Jaden was twenty-four when I first met him. It was his third admission to an inpatient psychiatric facility due to worsening psychosis. A typical story, in some sense, of medication non-compliance, escalating behavior at home and the family fearing for safety. A typically devastating story too, where just a few years earlier, this mild mannered and pleasant young man was going through high school, with all the promise and hope parents could have for their child. Smart, athletic and with big dreams, ready to take on college and beyond.

Shortly after college had started, Jaden began to struggle. Initially describing that things around him “just didn’t feel right.” He withdrew from his friends, and started missing classes. He then began interpreting ordinary events as carrying hidden meaning directed at him, and stopped leaving his apartment altogether. His roommate found him with a razorblade, cutting into his neck to find the hidden “thought transmitter” that had ruined his life. He did so just in time to call 911 and get him help.

Jaden did not like psychiatric medications. “Dr. Halassa, they just don’t work… I don’t really see the point.” His parents disagreed, and felt that there were some benefits. “You don’t seem to hear the voices anymore, that’s progress.”

“Mom, there is more to this than voices! I need to get my life back!”

And there it was. There is more to life than voices indeed. We do have medications that reduce psychotic symptoms, so from that perspective antipsychotic medications are helpful. However, in many people they do not restore functional capacity. The ability to reason, plan and interact with the world the way they did prior to their first psychotic break. Jaden’s parents agreed: “we would love to have our son back… will we?”

I cannot describe how gut-wrenching these conversations can be. It’s not that there are no examples of people overcoming psychotic illnesses and regaining function. There are, and I did write about that previously and discuss relevant literature. So there is certainly hope for people, but the issue is that we do not yet have a clear way of identifying the pathology in the broad entity we call schizophrenia, nor do we have effective medications that can target the broader syndrome beyond psychosis. To understand just how formidable this task is, I want to walk through what we currently think is going on, where the dominant theory falls short, and what a more complete picture might look like.

Schizophrenia beyond excess dopamine

For most of the past half century, dopamine sat at the center of the pathophysiological models for what we think schizophrenia is. The fact that giving people dopamine augmenting medications (like stimulants) precipitates psychosis and that traditional antipsychotic medications appear to work by blocking dopamine D2 receptors led naturally to the inference that excess dopamine was the problem. Plus, evidence indicated that people early in psychotic illness did have excess dopamine as measured through positron emission tomography (PET).

Shitij Kapur elaborated this connection via the aberrant salience hypothesis, which proposed that dysregulated dopamine signaling causes the brain to assign inappropriate significance to irrelevant stimuli. The world starts to feel charged with meaning that isn’t there. Random events feel connected. The mind builds explanatory narratives to account for these experiences, and because from the patient’s perspective the perceptions are real, those narratives have the internal logic of genuine beliefs.

As many readers of this newsletter know, I don’t particularly love the idea of molecular explanations of mental phenomena, so what are the circuit and computational correlates of this framework? Honestly, I’m not particularly clear on how exactly changes in striatal dopamine are supposed to relate to the experience of psychosis. There are many proposals that relate to the role of dopamine in assigning values to states, but no testable roadmap for how this translates to what people describe as hearing voices. Does this make the connection between dopamine and psychotic symptoms weaker? Possibly, but again, things that tend to increase striatal dopamine tend to increase psychosis (I’ve definitely seen people become psychotic on L-dopa who have never had a history of psychotic experiences, for example) and D2 blockers tend to make these symptoms better.

But the main issue is that schizophrenia is way more than voices, as Jaden pointed out. There are abnormal beliefs which we call delusions, many of which are not particularly ‘fixed’ by D2 blockade. Also, people experience loss of function due to changes in their ability to reason and think clearly. These are the cognitive deficits associated with the disorder which are also associated with ‘negative’ symptoms. Negative symptoms are hard to describe with natural language (you really have to see them to know what they are), but my best attempt at relaying them to you is the following: it’s the lack of initiation of action (and possibly thought), with what appears to be an intact ability to respond when input is provided.

Both cognitive and negative symptoms are linked to functional loss (loss of relationships, employment, education attainment) and both are thought to be a consequence of neural changes that are more causal to the illness than dopamine. Meaning, dopamine dysregulation and the positive symptoms appear to be secondary manifestations of something more fundamental that sits at the core of what we think the schizophrenia(s) are.

The Kraepelinian Core

Emil Kraepelin noticed the central problem more than a century ago, but lacked the tools we currently have to explain it. When he separated dementia praecox from manic-depressive illness, the distinction was primarily prognostic. Manic-depressive patients relapsed and remitted, while the dementia praecox patients deteriorated, often regardless of what happened to their psychosis. Kraepelin’s insight was that the trajectory was the illness, not the psychotic episodes that waxed and waned. He thought that an illness core may be eroding function steadily, in a way that was independent of psychosis altogether. That observation has held up across more than a century of follow-up data, and it is part of what makes schizophrenia so different from many other conditions we treat in psychiatry.

The evidence accumulated since Kraepelin points, imperfectly but consistently, toward the prefrontal cortex and the circuits that support how it does its work. The prefrontal cortex is responsible for the kinds of thinking that is most prized by us humans: maintaining goals across time, adapting to changing contexts, filtering out irrelevant information, and making decisions under uncertainty. Like all other cortical areas, the prefrontal cortex is built from both excitatory and inhibitory neurons; excitatory neurons are the main conduits of signals as they can drive one another (to sustain working memory or integrate signals across time for example) and project instructions to other areas, whereas inhibitory neurons control the activity patterns locally. There are many more excitatory neurons than inhibitory ones, but inhibitory function is critical for maintaining appropriate operating regime in cortical circuits, including operations like signal amplification and noise suppression. People have coined the term Excitatory/Inhibitory Balance (which was really developed for synaptic inputs on individual neurons, but gets used a lot more liberally to describe macroprocesses too) as a set point parameter for circuit function. In schizophrenia, converging evidence from postmortem studies, neuroimaging, and animal models points toward disruptions in inhibitory control that may give rise to loss of signal fidelity and accumulation of neural noise.

This formulation makes sense, broadly, when thinking about reasoning in schizophrenia and why it may be difficult for people to think clearly. Our lab has done a bit of work in this area (but we’ve also followed the literature very closely so it’s not just one experiment I’m highlighting). What we have observed is that patients with schizophrenia perform comparably to healthy individuals when task demands are straightforward, but some fail disproportionately as ambiguity increases. One view is that this is the consequence of a perturbed prefrontal network with particular failure modes that makes the brain susceptible to a variety of changes in the ability to reason, plan and initiate action. It also changes the ability to control downstream circuits including the dopaminergic system. Again, classical antipsychotic medications are able to address the dopaminergic abnormalities but do very little in addressing the prefrontal changes that may be at the core of these illnesses.

Schizophrenia may be hard to treat because we are not getting to its core.

Image illustrating prefrontal ‘top down control’ (prefrontal cortex: red), including that of dopaminergic function (ventral tegmental area: green). Modified from Ott and Nieder, 2019 TICS

Psychosis is not schizophrenia

If dopamine dysregulation is downstream of a more foundational core of schizophrenia, shouldn’t we be re-thinking the relationship between the two? Rather than being the defining feature of the illness, it becomes a clue: evidence that something has failed to regulate the dopaminergic system, resulting in surges that are associated with psychotic experiences. And one of the more interesting things about those surges is that they can apparently be triggered by multiple pathways.

Psychotic mania, which can happen in bipolar disorder, is a useful example. Molecular imaging studies find elevated dopamine receptor availability specifically in psychotic mania, with no significant difference between non-psychotic mania patients and healthy controls. A PET imaging study comparing bipolar disorder with psychosis directly to schizophrenia found that the dopaminergic signature in bipolar psychosis was comparable to what is seen in schizophrenia.

ADHD offers yet another route. In a subset of patients, particularly with high-dose amphetamines, psychosis can emerge. What’s intriguing about this is that there is a high co-occurrence of ADHD and bipolar disorder, which appears to be innate rather than treatment-driven. This suggests that some patients may carry a genetic liability spanning both conditions, placing them closer to the threshold to begin with. The diagnostic complexity these patients present clinically may reflect something important about their underlying biology that we have yet to understand.

What all of this suggests is that psychosis is better understood as a state than as the defining feature of any particular illness. In schizophrenia, the push toward that state may come from disrupted prefrontal function. In bipolar disorder and in stimulant-induced psychosis, it may come from other underlying mechanisms. This is a very important set of research questions that we need both public investment in and new talent to continue the quest of answering them.

But can excess dopamine drive schizophrenia?

This brings me to a question I find quite unsettling, and one the field has not fully worked through. If schizophrenia involves primary circuit disruption that produces dopamine dysregulation as a downstream consequence, what happens when you invert the sequence? When you start with dopamine and drive it hard enough, for long enough, we may be arriving at something that begins to resemble the underlying pathology we see in schizophrenia.

The animal literature on chronic amphetamine exposure offers some evidence that you can, at least partially. When dopamine is repeatedly and forcefully elevated, a secondary deficit in prefrontal function appears to emerge as a consequence. The cognitive deficits that follow, particularly in the kinds of flexible contextual thinking that depend on intact prefrontal function, begin to resemble what we observe in schizophrenia. In humans, chronic heavy methamphetamine use can produce persistent syndromes that extend well beyond the period of active drug use and include not just psychosis but negative symptoms and cognitive impairment. Strikingly, these cases are indistinguishable from schizophrenia that emerges the way it did in Jaden.

What it suggests is that dopamine dysregulation is not only a consequence of the primary pathology. When it is severe and sustained enough, it may begin to produce something like that pathology from the other direction. Whether that circularity has implications for how we think about illness progression in patients like Jaden, who experience repeated psychotic episodes with incomplete recovery between them, is very important. Because it helps inform whether psychotic episodes have a kindling effect, or not. Meaning, if people are slipping into psychosis (because of medication non-compliance for example), and when they are out of it, their frontal function is worse off than before, they need to know that so they are clear on the cost-benefits of taking medications, no matter how imperfect they currently are.

Map showing global prevalence of methamphetamine use. From this Lancet article.

For schizophrenia prevalence, check out this interactive map here.

Reasons for optimism

Jaden’s frustration was spot on. The voices were the part of his illness that our treatments could reach. They were not the part that determined whether he would get his life back.

Schizophrenia may be hard to tackle because psychosis, the feature that defines it in most people’s minds and that drives most of its treatment, is not the most devastating part of the illness. The illness may live in circuits that support how a person thinks, adapts, and engages with the world, circuits that were already showing signs of disruption before the first psychotic break and that tend to remain disrupted after psychotic surges are brought under control. The gap between what antipsychotics can do and what patients and families are actually asking for may be a consequence of having organized the entire treatment framework around a downstream signal while the upstream pathology has largely gone unaddressed.

But there is lots to be optimistic about.

The most significant development in schizophrenia pharmacology in decades is KarXT (Cobenfy), approved by the FDA in September 2024. Cobenfy combines xanomeline, a muscarinic cholinergic receptor agonist, with trospium chloride, which prevents xanomeline from acting outside the central nervous system and causing intolerable side effects. What makes it historically notable is simple: it is the first drug approved for schizophrenia in over seventy years that does not work by blocking dopamine D2 receptors. Cobenfy targets muscarinic M1 and M4 receptors, thought to act upstream of dopamine. M4 which is distributed subcortically, impacts dopamine release (and psychotic symptoms), while M1 which is enriched in the prefrontal cortex may get closer to the Kraepelinian Core of the illness.

This is consistent with both placebo controlled clinical trial data showing improvements on both negative and cognitive symptoms as well as preliminary real world reports. While we still need larger trials to confirm, the idea is that Cobenfy may be doing something different from traditional antipsychotics and that the patients most likely to benefit from that difference can be identified clinically. It is certainly going at least one step upstream of D2 blockers.

Other pharmaceutical companies are trying even further upstream, to target mechanisms of prefrontal excitatory/inhibitory balance more directly. One example I am familiar with is Ovid Therapeutics, which is developing a class of compounds that directly activate KCC2, the potassium-chloride cotransporter that controls chloride concentrations inside neurons. Chloride homeostasis is what determines whether GABA, the brain’s primary inhibitory neurotransmitter, can work effectively. Postmortem studies in schizophrenia have found reduced KCC2 in the prefrontal cortex, and animal models of schizophrenia-like pathology (in our hands) replicate the finding. Restoring KCC2 function would, in principle, address the inhibitory deficit directly rather than compensating for its downstream consequences. The first intravenous KCC2 activator, OV350, completed Phase 1 safety testing in late 2025 with a clean safety profile and electrophysiological signals consistent with central activity. An oral candidate, OV4071, is entering clinical trials in 2026, and will be applied to schizophrenia in 2027 if all goes well. I recently participated in an Ovid event and talked about our work here (you just need to enter some info to gain access to the full recording). Oliver (Ollie) Howes also spoke right before me and introduced the prefrontal E/I balance framework in relationship to his work.

None of this means we are close to having something to offer Jaden that addresses what he was asking for. The gap between what these drugs might eventually do and what he needed in that room remains large. But for the first time in a long while, the therapeutic programs being developed are starting to engage the biology that the evidence points toward, rather than cycling through variations on the same dopamine-blocking approach that has dominated since the 1950s. That is the right direction and for now, it’s the best we are all able to do.

If you found this piece useful, please consider subscribing and sharing it with someone who might find it interesting. Paid subscriptions help support the time it takes to produce this work.

Disclaimer: Jaden is a fictional character who draws upon a composite of patients I’ve interacted with over the years. I cannot overstate the sense of privilege I feel to have been a part of these folks’ and their families’ lives.

Second note: This essay is very long. While it can be read from top-to-bottom—and is written assuming you will—you will lose little by simply choosing specific sections you find interesting and reading only those.

1. Introduction

2. Some thoughts

   1. The business case for biosecurity requires another pandemic for it to work

   2. The screening architecture assumes a chokepoint that’s disappearing

   3. Targeting humans with bioweapons is (probably) genuinely difficult

   4. Agricultural bioterrorism is (probably) really easy

   5. The monitoring architecture is useful for detection, but not defense

   6. Machine learning may be very useful for rapid-response therapeutics

   7. Pathogen-agnostic defenses are extraordinary, but who pays for it?

3. Conclusion

Introduction

It is easy to scare yourself about biosecurity, and it is getting easier every day. Everyone has their moment when the fear first crept into their throat. Mine was when I read the article titled ‘AIs can provide expert-level virology assistance’, which found that LLMs—even ones as ancient as Gemini 1.5 Pro—are more than capable of happily providing the knowledge needed to debug BSL-4-sounding questions about wet-lab experiments.

As with any paranoia worth having, there are good objections to it.

Most recently, the non-profit Active Site published the largest randomized control trial of its kind—153 novices, 8 weeks, a BSL-2 lab in Cambridge—studying how much access to frontier LLMs (Opus 4, o3, Gemini 2.5, all with safety classifiers off) gave participants ‘uplift’ on performing a set of viral genetics workflow (including virus production), compared to only access to the internet. Their conclusions are the following: ‘We observed no significant difference in the primary endpoint of workflow completion (5.2% LLM vs. 6.6% Internet; P = 0.759), nor in the success rate of individual tasks’. with the caveat that the LLM has numerically higher success rates on 4 out of 5 tasks, just not high enough to reach significance level. YouTube, not the LLMs, was rated most helpful by both groups.

So, while frontier models are theoretically capable of providing virology assistance, it doesn’t immediately seem like they can bootstrap someone into wet-lab competence; the hands are still the hard part. There are counterpoints to this as well, like, ‘LLMs probably help non-novices a lot!’, and ‘the study is underpowered!’ and so on. I agree with some of these. The truth is almost certainly somewhere in the middle: LLMs can help a novice with wet-lab work, but they don’t help an infinite amount.

Yet, I still believe it is still hard to actually turn all this into something evil. And no, I do not think that gesturing towards ‘automated labs’ is a good counter argument. Doing things in the world of atoms is difficult. Especially here. Why? Didn’t I just write a month back about how cloud labs are the final end-state of lab automation plays, so can’t they be hacked into doing something ulterior? Man, maybe. But you should consider the fact that these cloud labs are, at the moment, barely functional enough to do the things their paying customers want them to do, let alone serve as unwitting accomplices in a bioterror plot. Yes, they will improve, but their improvement is on a very jagged frontier. Liquid handler automation is going splendid. Liter-scale creation, purification, and aerosolization of BSL-4 substances automation is not going so splendidly. Also, even in the case where automation suddenly rapidly accelerates, it is almost certainly economically not viable for these labs to care about servicing the likely small consumer market of ‘large-scale non-therapeutic virus creation’.

I’ll discuss this more deeply in the upcoming sections, but it feels that doing something as ambitious as bioweapon creation will likely be extremely annoying to do for the foreseeable future, and I am consistently on the side that only a well-funded actor would be capable of such a thing. And why wouldn’t those actors opt for much simpler acts of violence that would roughly accomplish the same thing?

This all said: I sympathize with the bioterrorism-phobia that is sweeping my simcluster. If you stare for long enough at the AI trendlines, and also observe the increasingly WW3-y vibes that the world is emanating, it is difficult to not feel at least some worry. Maybe a genuine bioterrorism incident is not too far away. And maybe it will be far, far worse than anything can imagine.

Or maybe not. Biosecurity is one of those topics that can either feel extraordinarily bleak in its prognosis, or like things are obviously going to be fine. As with many things in the world, I think both sides have a bit of a point, and I think holding them in tension is the only honest way to consider how the future may go. In this essay, I’ll share some of my own thoughts on the field at large, and the specific themes that arose in my discussions with people.

Some thoughts

The business case for biosecurity requires another pandemic for it to work

As with all problems that, if not solved, may lead to the depopulation of the planet, we can depend on venture capitalists to search for a market opportunity. A few companies have emerged in the last few months as the vanguard of this effort: Valthos ($30M Series A for being the Palantir of biosecurity), Red Queen Bio ($15M seed for designing therapeutics against bioterrorism threats), and Aclid ($4M seed for DNA synthesis screening infrastructure). There are others too, but we’ll stick with these ones for now for illustration purposes.

I have zero doubt that these companies, or something akin to them, are worth having around. What I cannot quite figure out is the business model. The usual answer for the ‘who pays for this?’ questions in these sorts of public-goods-situations are government agencies: BARDA, DoD, DHS, CDC and so on. This is not so bad of an idea.

Let’s take a look at the United States’ 2026 budget proposal for the biodefense-adjacent areas to get a sense of these agencies’ funding.

In the proposal, BARDA is being cut by $361 million, a roughly 36% decrease from its prior state. Project BioShield, the program that actually buys finished countermeasures, is on track to lose $100 million. The CDC budget is halved, coming at around a $5.4 billion loss. DTRA down $150 million. One article more deeply analyzing the many, many other various biodefense cuts being made had this to say about it:

With the Trump Administration’s priority of reducing federal spending, the funds requested for biodefense have been significantly reduced. Very few biodefense programs saw increases in their funding or even a continuation at their previous funding levels.

How about the Department of Defense (War)? Mixed picture: while the overall budget of the department was increased, the bio-adjacent programs within it saw a drop.

One notable example is the Defense Threat Reduction Agency (DTRA), a key government agency that prevents and mitigates deliberate biological threats to the US globally, for which the PBR requests $708 million. This is $150 million less than the $858 million requested in FY25. Similarly, the $1.61 billion FY26 request for the DoD Chemical and Biological Defense Program is $46 million less than requested for FY25.

In other words: the agencies that would theoretically buy tools from, say, Valthos, are the same agencies that the current administration is intending to either gut or barely increase the budget of.

There is good news: this budget did not come to pass.

Congress rejected nearly every one of the proposals: the CDC’s budget was not reduced, while BARDA, Project BioShield, and NIH’s budget actually slightly increased. There is one unfortunate budget stain—Kennedy pulling $500 million from a BARDA program developing mRNA vaccines against various respiratory viruses—but things overall turned out fine, though I cannot find specific numbers on how things fared on the DoD end. But it is a little worrying that the administration is not particularly sympathetic to biosecurity concerns. Why? Because if your primary customer is prone to wild swings of financial unpredictability, and it is only thanks to the grace of Congress that those sentiments are not actively reflected in their budget, it almost certainly hurts the capacity for these companies to plan for the future.

Earlier I mentioned that Valthos intends to be the Palantir for biosecurity. This is not a presumption on my end, they have basically said this. The CEO (Kathleen McMahon) is an alum of the company, and has stated that Valthos plans to apply “many of the same principles she learned at Palantir, about working with officials as well as commercial customers”. But an easy counterargument to this is that Palantir’s government business was built during the post-9/11 spending surge, when homeland security funding went from $16 billion to $69 billion. Biodefense is holding steady, for now, but not seeing the same dramatic jumps.

You could imagine that a pretty simple steelman for these objections is not dissimilar to the usual AI-wrapper-SaaS advice people give: build not for where the models are today, but where they are going. And if you trust the trend-lines, it is not inconceivable that there is a catalyzing event in our near future—a genuine, bona-fide bioterror incident—which will unlock massive government spending the way 9/11 created the entire homeland security industry overnight. In this setting, the companies that already have working products and government relationships when that moment arrives will be the Palantir of biosecurity. The ones that don’t will be too late.

The game then, is to survive until this catalyzing event occurs. If it does, Valthos may be able to gobble up all the government contracts it wants, Red Queen Bio may find the DoD suddenly desperate to fund therapeutics platforms that have a biosecurity veneer to them, and Aclid may discover that its few dozen synthesis company customers grow to have even stricter compliance requirements. If it doesn’t, it is tough to imagine that these companies don’t go either bankrupt or stay growth-capped. Because of this, you shouldn’t be surprised at all that these companies acquired the funding that they did! The game of venture capital is to play ‘big if true’ bets, continuously, forever, and few areas are as well-shaped to that as biosecurity.

Well, maybe. You could argue that the SARS-CoV-2 virus maybe couldn’t be the catalyzing incident for the government, since it is still unclear whether it was a lab-leak or not, but what about the 2001 anthrax attacks? How come that didn’t spur a massive amount of increased federal biodefense funding? In fact, it did. Total US biodefense funding jumped from roughly $700 million in 2001 to about $4 billion in 2002, peaking at nearly $8 billion in 2005. What was the money used on? A fairly large chunk was put into anthrax-specific [stuff]. As a case study, consider Emergent BioSolutions, the sole producer behind ‘BioThrax’, the only FDA-approved anthrax vaccine. They received: one $1.6B contract for a second-generation anthrax vaccine, one $1.25B five-year contract for delivering 44.75 million doses of an older vaccine candidate, followed by a $911 million CDC contract for another 29.4 million doses. A 2021 Congressional investigation found that, for the past decade, nearly half of the Strategic National Stockpile’s budget had gone to purchasing this anthrax vaccine, a product whose price had been raised 800% since 1998. And it is still ongoing, with a $21.5 million delivery order to the Department of War was issued as recently as January 2026.

This is, on some level, completely understandable. You prepare for the thing that just happened to you. But it should make us a little nervous about the “catalyzing event” thesis for biosecurity companies, because the empirical reality is that it may not unlock general biodefense spending so much as it locks in countermeasures that are overly anchored on the specific threat and threat vector of that particular incident.

So, perhaps it is worth exploring outside of this customer base. While governments are the biggest buyer, they surely are not the only one. After all, didn’t Kathleen’s comment mention commercial buyers too? There is another group on the table: DNA synthesis companies. A fairly high fraction of the current biosecurity framework rests on a pretty simple idea: that biological information must pass through synthesis companies to become biological reality. To actually make a [thing], you need physical DNA, and to get physical DNA, you order it from a commercial provider. Why not create a layer to screen the DNA being ordered, ensuring that whatever it is, it isn’t dangerous? This is, as previously mentioned, Aclid’s business model, alongside TwentyTwo, SecureDNA (a non-profit), and likely others.

How is that going?

The preventative architecture assumes a chokepoint that’s disappearing

There seem to be three problems with DNA-screening-as-biosecurity.

The first of which is that the screening only works if you’re ordering sequences long enough to screen. According to HHS guidelines, the current screening threshold is 50 nucleotides, but oligonucleotides—short DNA fragments often used in legitimate research—can be ordered, assembled, and stitched together into longer sequences. This is not theoretical. In 2018, Canadian researchers synthesized a functional horsepox virus from mail-ordered DNA fragments for about $100,000. Fairly, this is annoying to do, but a sufficiently dedicated adversary may be happy to do annoying things.

The second is that screening assumes you’re looking for known threats, which is to say, sequences with similarities to characterized pathogens. But if AI biological design tools might enable the creation of de novo pathogens, things that don’t have a match in any database because they’ve never existed before, then the screening becomes useless. And you needn’t even hop your way to truly de novo stuff, you could just redesign the existing bad pathogens in ways that make them invisible to screening tools. For example, Microsoft has a “paraphrasing” paper that was exactly this, redesigning known, toxic proteins in ways that evade sequence-based screening while preserving function. To counter this, you’d need to predict function from sequence alone, which is one of the hardest open problems in the field, especially because ‘function’ in biology is one of those super fuzzy, contextual words that can have a bunch of different meanings. It is certainly possible to do—see the podcast I did with Yunha Hwang, an MIT professor creating tools to automatically annotate the function of metagenomes—but it’s not easy.

The third problem is the biggest, and it is that benchtop DNA synthesizers are getting longer-range. In other words, you could neatly side-step all these screening checks by buying your own DNA-creation machine, and running synthesis in your bedroom. Right now, the best commercially available benchtop synthesizers tops out at about 120 base pairs per well, which, given that real viruses are on the order of dozens of kilobases, means we’re safe for now. But there is no functional reason that they cannot get any better. In fact, according to a fantastic Institute for Progress (IFP) report, it’s just around the corner. Enzymatic (as opposed to chemical) DNA synthesis is likely less than a decade off, comfortably pushing DNA synthesis capabilities to the kilobase realm. This all said: a few people I talked to mentioned that ‘long-range DNA synthesis has been a few years away for a decade-plus now’, so maybe we can discount this a little, but it’s worth paying attention to. Especially because, as we mentioned earlier, a DNA synthesizer needn’t be capable of full viral genome synthesis to be dangerous, since you can simply splice its outputs together.

This is all quite a pickle.

Yes, you could lock down the benchtop synthesizers, such that any attempt to use them would involve making an external call to some pathogen database to screen your request. But if the ML design tools get good enough, you can just do continuous zero-shot designs of something that doesn’t match anything in the database, and iterate from there. And even if the models don’t get good at that sort of in-vivo prediction behavior—which, despite what you may hear, is a genuine possibility for at least some time—you could simply split your order across multiple machines, synthesizing fragments that are each too short to trigger any screening individually, but that assemble into something very much on a select agent list once stitched together.

This last point is also called a split-order attack. The IFP report discusses this last point as well, and is refreshingly blunt about the prognosis.

Moreover, an offline device is vulnerable to the whole class of split-order attacks, whereby the adversary can combine the outputs of two or more devices that are small enough to evade screening in isolation, but together would be recognized. Without some centralized connectivity, such an attack is impossible to defend against.

Are we doomed?

Maybe. The optimistic angle is that the government can be awfully good at shutting things down when it wants to, and the track record in other domains is quite encouraging. When the Combat Methamphetamine Epidemic Act passed in 2005, putting pseudoephedrine behind the counter and requiring ID and purchase logs, domestic meth lab incidents dropped by over 65% within two years. Nuclear materials are an even stronger case: the NRC administers over 20,000 active licenses for radioactive materials in the US alone, coordinated across 40 states and backed by the international IAEA safeguards regime. This has almost certainly contributed to the fact that there has not been a single case of nuclear terrorism. When the government decides something absolutely cannot be allowed to proliferate, and builds the institutional machinery to back that up, it can, against all odds, work.

But the fundamental problem here is that preventing bioterrorism requires a level of governmental diligence that is closer to nuclear-level than meth-level, and right now it is far behind both. To be fair, there are clear structural differences between biology and nuclear/meth, the biggest one being that biology is much more dual-use. Benchtop synthesizers have far, far more legitimate uses than malevolent ones, and the upside of restricting them is a lot harder to argue for then, say, restricting access to pseudoephedrine.

Well, what should be done? The IFP proposal, to its credit, has some pretty clear demands: a mandatory Biosecurity Readiness Certification before any benchtop synthesizer can be legally sold, standardized customer screening for both devices and reagents, and a reagent track-and-trace system modeled on the Drug Supply Chain Security Act for pharmaceuticals. None of this is crazy, and rhymes with what has already been done for meth and nuclear material.

What is actually being done? Unfortunately for all of us, every federal document governing DNA synthesis security in the United States right now is (somewhat) voluntary, though there is a nuance here we’ll get to in a bit. The only binding rules are export controls, which have, circa 2026, already been violated. The IFP essay from earlier happily reports that Telesis disclosed in their SEC filings that their DNA assembly systems have accidentally ended up in embargoed countries through distributors.

Oops! Does uranium ever accidentally end up in embargoed countries?

Well, actually, yes. The IAEA has logged over 4,300 incidents of nuclear material outside regulatory control since 1993, 353 of which involved trafficking or malicious use, 13 that involved high enriched uranium, and 2 that involved plutonium. But importantly, the last time someone got their hands on kilogram quantities of weapons-usable material was 1994. The system leaks at the margins, but it doesn’t leak at the catastrophic level.

The security model that you’ll continuously hear repeated among biosecurity experts is the ‘swiss-cheese’ model, in which the purpose of the regulatory apparatus is to present enough overlapping layers of defense such that no actors, other than the absolute most determined, are willing to go through the trouble. The defenses against nuclear and meth are swiss-cheese-y, and the ideal solution for DNA screening will likely be similar. Possible to defeat, but difficult, annoying, and legally scary to attempt.

And at least one layer of cheese is present: I mentioned that screening is largely voluntary on the part of the synthesis companies, but there is an important caveat. It is required for federally funded entities to purchase synthetic nucleic acids only from providers or manufacturers that adhere to the US Framework for Nucleic Acid Synthesis Screening. In other words, if a DNA synthesis company wants to sell to the enormous market of federally funded researchers (most of the U.S. life sciences market), they effectively must implement screening.

Well…kinda. This particular screening requirement was the intended purpose of one piece of legislation that was passed in 2024, but the current administration issued an executive order in 2025 to replace it with something [better] within 90 days. These 90 days have come and gone, and there is yet for anything to pass to mandate it again. This said, the biggest DNA synthesis providers (Twist and the like) see the writing on the wall, and have already implemented the screening that they imagine will be required of them, but it is unlikely smaller DNA synthesis providers have. Circa February 2026, there is a bill going through the Senate to address this current regulatory gap.

But what about all the problems from before? Split-order screening, AI-assisted genome redesign, and DNA benchtop synthesizers? Legally mandated screening is surely useless given those. We need more layers of cheese to defend against these!

Many smart people have thought about these challenges, and there are ways to solve them if you can get widespread buy-in from the synthesis providers. You could create centralized repositories of DNA orders that are aggregated from multiple providers, you could assemble private saturation mutagenesis viral datasets to catch most attempted redesigns from bad actors, and you can install hardware locks on benchtop synthesizers to prevent them from being used without connection to the aforementioned centralized repository.

None of this is scientific fiction! There are groups actively working on all of them, and some are even wrapped up in the Feb 2026 bill I just mentioned. But we’ll see how realistic they are to implement in practice.

Targeting humans with bioweapons is (probably) genuinely difficult

There is something under-appreciated worth discussing: making and spreading bioweapons is not easy. I mentioned this at the start, but there is a lot more color to add.

If you talk to biosecurity folks for long enough, they will eventually mention Aum Shinrikyo. Aum is a Japanese doomsday cult that, in the 1990s, had everything a would-be bioterrorist could ask for: hundreds of millions of dollars, a graduate-trained virologist who had studied at Kyoto University running their bioweapons program (Seiichi Endo), dedicated lab facilities, and years of total freedom from law enforcement scrutiny. They believed the end of the world was upon them, and that their mission was to hurry the whole thing along. On their journey to do exactly this, they attempted ten biological attacks.

Every single one failed. Their most ambitious effort was the 1993 anthrax attack on Kameido, Tokyo, where cult members sprayed a liquid suspension of Bacillus anthracis spores, or anthrax, from a cooling tower on the roof of their headquarters onto the streets below. Nothing happened. It turned out they’d acquired a vaccine strain of anthrax, one that is, to quote the CDC’s postmortem, “generally regarded as nonpathogenic for immunocompetent people.” Even if they’d had the right strain, the spore concentration in their slurry was about 10⁴ per milliliter, versus the 10⁹ to 10¹⁰ considered optimal for a liquid bioweapon. They had a botulinum toxin program too, in which they attempted multiple attacks using vans fitted with sprayers. Once again, zero effect. The toxin was likely either degraded during processing, too dilute to have any effect, or produced from a non-toxigenic strain because they couldn’t maintain proper anaerobic fermentation conditions. It is unclear as of today.

An account of the many difficulties the group faced in actually creating usable bioweapons is well-described in this 2011 report, which has some real comedic gems:

Mice on which the yellow liquid [Botulinum Neurotoxin] was tested showed no toxic effects, and one cult member reportedly slipped into a fermenting tank and nearly drowned, but subsequently showed no signs of illness.

The same report notes that even Aum’s manner of spreading their pathogens may have interfered with their efficacy:

In the even more unlikely event that Aum had produced and successfully stored volumes of a virulent strain, it is possible that poor dissemination capabilities might have damaged the material or failed to aerosolize it so that sufficient quantities could be inhaled.

For example, the cult employed a homemade nozzle that reportedly spouted rather than sprayed and dispersed material during the day, exposing it to UV radiation and thermal updrafts that would have reduced concentrations at ground level.

The group did finally end up partially succeeding, but only after switching to chemical weapons: sarin nerve gas, which ended up killing 13 people and injuring thousands on the Tokyo subway in 1995.

But, you may protest, the 1990’s was a long time ago. We have nanopores now. We have Alphafold3 now. We have a (somewhat) mature field of synthetic biology.

All very true, but consider what actually went wrong for Aum. They used the wrong strains, their fermentation got contaminated, their concentrations were off by five orders of magnitude, their aerosolization likely didn’t work, a guy fell into a fermenter and was fine. These were problems of bioprocess engineering, strain selection, maintaining sterile culture conditions, building dissemination devices that produce the right particle size, and overall wet-lab competence. Some of these are pure information problems, yes, and some of them are fixed by using easier-to-produce viruses (rather than bacteria), yes. But others are iterative, hands-on, tacit protocol development work. Of those, none would be aided by the current generation of structural biology models, and only some would be aided by LLMs given the Active Site results I mentioned at the start of this essay.

There are other case studies to consider too. Canonically, there are three other historical bioweapons programs of note: the Soviet Union’s in the 1970s, Iraq’s program under Saddam in the 1960s, and the US’s own Cold-War-era investigation into bioweapons in the 1960s. Unlike Aum, all three had one thing in common: they were state programs, with thousands of employees, dedicated production facilities, and decades of institutional knowledge.

How did these groups fare?

Iraq’s program, despite Saddam’s enthusiasm, produced anthrax and botulinum toxin of such inconsistent quality that US intelligence assessments after the Gulf War concluded the weapons would have been largely ineffective in most deployment scenarios.

The US program—which weaponized anthrax, botulinum toxin, tularemia, brucellosis, and Q fever—had a slightly different takeaway, but one that’s still directionally aligned with what we’ve discussed. After nearly three decades of doing comically dangerous acts like releasing simulant organisms in the San Francisco Bay Area and the New York subway to study how pathogens would move through civilian infrastructure, the conclusion wasn’t exactly that bioweapons didn’t work, it was that they were strategically irrelevant. At this point, the US already had a nuclear arsenal that can glass a continent in an afternoon, and the marginal value of a weapon that is unpredictable, uncontrollable, and might blow back on your own population became effectively zero. Nixon shut the program down in 1969, and there were few complaints against the decision.

Next, the Soviet program, also known as ‘Biopreparat’. It was the largest biological weapons program in human history, employing over 60,000 people at its peak, and spent years trying to weaponize smallpox and plague. And it worked. Some insane lines from a Frontiers article about the program attached here, bolding by me:

Some Biopreparat and military facilities continuously produced agents and filled the delivery systems kept on standby. For example, the Soviets annually made about two metric tons of antibiotic-resistant pneumonic plague and 20 tons of liquid smallpox grown in eggs. Refrigerated bunkers stored the bulk smallpox, which had a 6 to 12-month shelf life, and also contained filling lines for munitions and spray tanks.

….The Corpus One building of The State Scientific Center of Applied Microbiology at Obolensk contains 42-story tall fermenters, separated into different biosafety containment zones, to make plague and other agents.

Building 221 at The Scientific Experimental and Production Base at Stepnogorsk housed 10 four-story-high, 20,000-liter fermenters and could make 300 metric tons of anthrax in 10 months. Other production lines at Kurgan, Penza, and Sverdlovsk could add hundreds more tons to the USSR’s prodigious capability to make biowarfare agents and fill munitions on short notice.

Fortunately for us, the Soviet economy collapsed before this stockpile could be used for anything world-ending.

I think there are a few takeaways here. One—from the US’s experience—is that bioweapons are fundamentally not worth it if the end goal is to wag a very large stick towards your enemy. Two—from Aum’s and Iraq’s experience—is that bioweapons are genuinely hard to create and disperse, even with significant resources and time. And three—from the Soviets experience—is that if you throw enough of a country’s industrial base at the problem, the engineering/scientific barriers can be overcome, but the scale of effort required is immense.

These are, alongside Aum, four, isolated cases from decades back. How much could we learn from such an isolated slice of history? Should we really let our mental models be informed by this?

Unfortunately, it is the best we’ve got. We do know there are other ongoing bioweapons programs today. In an April 2024 compliance report released by the the U.S. Department of State, they state that North Korea and Russia are definitely running a bioweapon program, and it is possible that Iran and China are also. Should this freak us out? Maybe. On one hand, we should take seriously the US opinion that bioweapons kind of suck, and that there are easier ways to kill many people. On the other hand, the strategic value of bioweapons is not just in killing many people, but also in plausible deniability. Either way, whether these programs perform as intended in a real-world deployment scenario is a very different question, and one that neither the compliance report nor this essay is not positioned to answer.

Agricultural bioterrorism is (probably) really easy

Unfortunately, most of what I said earlier referred to pathogens meant to target humans. The calculus changes dramatically when your targets are cows or a wheat field, or so-called ‘agroterrorism’. This isn’t great news, especially because if you spend any time reading the biosecurity discourse, you will notice that relatively few people discuss this topic, and, of the folks who mention it, the word ‘overlooked’ pops up a worrying amount.

Over the next few paragraphs, I’ll try to give some intuition as to why agroterrorism is uniquely challenging to combat.

First, the actual design of the pathogen.

Unlike most of the other, nastier viruses and bacteria that cause humans to bleed from every orifice, many incredibly dangerous agricultural pathogens do not require BSL-3/4 equipment to safely create. As a result, the barrier to entry in agroterrorism is incredibly low. While the Soviet Union bioweapons program had to regularly deal with unfortunate cases of accidental Marburg, smallpox, and anthrax leaks—even while having BSL-3-ready labs!—a bad actor here can freely muck around with designing whatever they want with little threat. And if you’re feeling especially thrifty, you don’t even need a novel gain-of-function chimera. You need foot-and-mouth disease, which already exists in nature, is endemic in parts of Africa and Asia, and is one of the most contagious diseases known to veterinary medicine.

In fact, we know this because a former Soviet Union bioweapons producer—Kenneth Alibek—told us. In a 2006 report, he extensively discussed his work, with one paper having a particularly good paraphrasing:

Alibek describes the Soviets as producing anti-livestock, anti-crop, and combined anti-livestock/anti-personnel pathogens. During the course of its existence, the Soviet’s anti-agricultural bioweapons program produced and weaponized the anti-crop pathogens Wheat Rust, Rye Blast, and Rice Blast; the anti-livestock pathogens African Swine Fever, Rinderpest, and foot-and-mouth disease…

…The Soviets used simple, rudimentary techniques to develop these effective antiagriculture pathogens. They developed anti-crop fungal pathogens through a simple ground cultivation technique, while anti-livestock pathogens were developed in live animals…

All of these techniques, as Alibek points out, could easily be utilized by unsophisticated terrorist organizations to develop bioweapons designed to cause mass casualties of agriculture.

Next, distribution.

If you want to cause a human pandemic, you need aerosolization, you need to calculate incubation times, you need sophisticated delivery mechanisms. Agricultural pathogens require none of this. As one paper puts it, deploying plant or animal pathogens could be as simple as “atomizing unprocessed pathogen near the target organisms or, in the case of animals, directly applying the pathogen to the nose and mouth of the organisms.”. Why is it so easy? Is there something special about agricultural pathogens? No, but there is something special about how modern agriculture is done, in that it involves thousands of nearly-genetically-identical plants and animals in astonishingly dense conditions. The environment does the work. All this, with virtually zero risk to the adversary, given that this would not be done in crowded cities with cameras on every corner, but on sprawling, isolated farms that have essentially zero security infrastructure.

Finally detection.

Unlike human disease surveillance, which benefits from the fact that sick people tend to show up at hospitals and demand attention, cows and wheat do not. As a result, agricultural disease relies on a very error prone set of steps for its detection to ever occur: one, the farmer noticing something is wrong with their animals, two, the farmer reporting it to the government, and three, the authorities being dispatched.

We’re going to spend the next few paragraphs discussing these three steps, because each step is a point of failure, and they fail constantly.

First, the farmer notices something is wrong. This is hard. You have to realize the scale that modern agriculture operates at.

A single large-scale poultry operation can house 50,000 turkeys or hundreds of thousands of laying hens in a single building. A feedlot might hold 100,000 head of cattle. The average dairy herd in states like California or Idaho now exceeds a thousand cows. And the trend is accelerating: U.S. livestock production has been consolidating into fewer, much larger operations for decades, with the economics of scale constantly toward ever-increasing density. As a matter of example: an outbreak of H5N1 among cattle populations in the United States began in December 2023, . How long was the lag between initial infection and actual detection? According to a Science paper from April 2025, the virus circulated entirely undetected for over 4 months. Clinical signs—reduced milk production, decreased feed intake, and changes in milk quality—were first noticed by veterinarians in late January 2024. Only on March 25, 2024 was the virus confirmed to exist after genetic sampling of the cows milk. By that point, the virus had already reached 26 dairy cattle premises across eight states and six poultry premises in three states.

Let’s say the farmer eventually realizes that something is wrong. Now they need to report it to the correct authorities. But why would they? There is something extraordinarily perverse about the reporting incentives at play here: farmers are actively disincentivized from flagging unusual disease, because a confirmed outbreak of a notifiable disease may wipe out their entire livelihood. Remember: these pathogens are often so virulent, so adaptive, that mass culling of their herd will be what is demanded of them. So, if you’re a rancher staring at a few sick animals, the economically rational move is to wait and see if they get better, not to call a vet and risk having your entire herd destroyed. Once again, there is empirical proof here: how Indonesian farmers handled avian bird flu in 2006. A paragraph from a zoonotic disease book is instructive:

Those smallholder poultry keepers questioned the severity of the avian influenza threat to their birds….Some continued to consume and sell diseased dead birds. Small to medium-sized contract poultry farmers feared that government officials might cull their birds before definitive laboratory confirmation of the disease, and they were skeptical of compensation schemes or believed compensation was too low. These poultry farmers reported the deaths of chickens to contractors, who in turn sought the services of private veterinarians to determine the causes of bird death, making effective disease surveillance difficult. Smallholder poultry farmers and keepers feared reporting incidents directly to the government. This fear was not limited to a concern about losing their own birds, but also to the social risk of angering nearby neighbors, whose birds would be subject to culling within a 2–5 km radius of an outbreak location.

You may ask: in the case of animals, why can’t we just vaccinate them? You can! But export regulations prevent most farmers from doing so, because standard vaccines make it impossible to distinguish a vaccinated animal from an infected one. Vaccines that include marker proteins allowing serological tests to tell vaccinated animals apart from infected ones do exist, or so-called DIVA vaccines, but adoption has been glacial.

Finally, let’s say, against their better judgement, the farmer reports it. What happens then?

How the U.S. government actually responds to agricultural threats is theoretically fairly straightforward. Human pathogens fall under HHS, via the CDC. Agricultural pathogens fall under the USDA, via its Animal and Plant Health Inspection Service (APHIS). There is a select agent list for each, plus an overlap category for things that threaten both. The jurisdictional lines are reasonably clear. The problem with the agency technically in charge, the USDA, is that it is also the agency whose mission includes promoting the very industry it would need to disrupt in a crisis.

To understand this better, we can look at a fascinating Vanity Fair investigation that interviewed over 55 people across USDA, CDC, HHS, and the White House, all of whom were involved in the same H5N1 cattle outbreak we just discussed. Since the virus was first confirmed in 2024, the two organizations were barely aligned: the White House was planning a public-health-directed response, while the USDA was prioritizing the needs of the dairy industry.

Within weeks of the diagnosis, APHIS employees began calling state veterinarians from personal cell phones to confide that they had been instructed not to discuss, not to engage, and to discontinue even routine conversations with health officials in the field unless talking points were pre-approved. The USDA sat on genetic sequencing data for weeks, sharing samples an average of 24 days after collection—compared to 8 days for the CDC—and without basic metadata like the date or state of collection, rendering the data effectively useless for real-time monitoring. The same farmer incentive problem from before reared its ugly head too: dairy farmers simply opted not to test, and some forced veterinarians off their property. At least five veterinarians who had been outspoken in responding to the outbreak were fired from their jobs. By the time a Federal Order requiring pre-movement testing was issued, the virus had already spread across multiple states. And the testing regime was widely regarded as obviously insufficient: just 30 animals per herd, with farmers reportedly prescreening in private labs to cherry-pick healthy animals.

Because why not? Who was going to stop them?

This was a naturally occurring virus, both in viral origin and how it was spread. Yet, the federal response still took months to coalesce into something real.

And as much as you may think the APHIS bungled this, it is difficult to imagine their future responses will look much better. As of mid-2025, APHIS lost roughly 1,377 staff under the administration’s workforce reduction push, about 16% of its employees. The USDA also accidentally fired several employees working on the H5N1 response, and had to scramble to rescind those termination letters within days. Yes, it may be the case that the organization is bloated beyond a reasonable doubt, and the cuts were deserved. But the cuts have not been accompanied by any visible effort to fix the structural problem here: the fact that the USDA is simultaneously the regulator of and the lobbyist for the industry it oversees.

But there is an important question to ask. What is the ultimate impact of all this? What actually happens if a successful agroterrorism attack occurs? Because if it’s insignificant, just a rounding error, then none of this should be a concern.

It is not a rounding error. The 2001 foot-and-mouth (FMD) outbreak in the UK resulted in over 6 million animals culled, cost the public sector £3+ billion and the private sector £5+ billion, was severe enough to delay that year’s general election by a month, and lead to the dissolution of the Ministry of Agriculture entirely. Simulation models for the United States are even uglier. A study modeling FMD outbreak originating in a single California dairy farm found that median national agricultural losses ranged from $2.3 billion to $69.0 billion depending on detection delay, with every additional hour of delay at the 21-day mark costing roughly $565 million and another 2,000 animals to be slaughtered. What about a deliberate, state-actor attack? Another simulation model estimated the economic impact of a FMD agroterrorism scenario—vast, widespread dispersal of the pathogen—put possible losses between $37 billion and $228 billion across three scenarios, from a contained state-level outbreak to a large multi-state attack.

But there is at least some argument that, under some mental models, it actually is a rounding error. The United States’ agricultural GDP is roughly $1.4 trillion, while the overall GDP is $29 trillion. Even the worst-case FMD simulation represents about a 16% hit to agriculture, and a 1% hit to the broader US economy. This is not nothing, it may completely devastate the nation, but it also is not civilization-ending.

Yet, while agroterrorism perhaps isn’t a standard x-risk scenario, when evaluated against the “is this a serious national security threat“ standard, the answer feels like it is an obvious yes. This raises a rather important question. If everything I’ve said is true—and I’m pretty sure it is—why hasn’t there been a significant agroterrorism event…ever? I have no idea, and it too was a point of confusion among most those I talked to. The best argument I’ve heard is that, if the ultimate goal of bioterrorism is to either terrify a nation or outright end the world, neither the aesthetics nor net-effect of agroterrorism is well suited for either.

However, one person I talked to did say there has, in fact, been one case of minor agroterrorism they are aware of: in late 2019, drones controlled by gangs dropped [items] infected with African swine fever into commercial pig farms in China. Why were the gangs trying to spread swine fever? So that the farmer would be forced to sell their potentially infected meat cheaply to the gangs, who would then sell it on as healthy stock. This feels like a rather roundabout way to make money, but it happened. Moreover, it may be the case that stuff like this occurs far more than anyone realizes, since the whole racket was only discovered because Chinese farmers resorted to radio jammers to prevent the drones from flying near the farms, which ran afoul of the regional aviation authority.

The monitoring architecture is useful for detection, but not defense

The United States has two main systems for detecting biological threats in the environment: one that watches the air, and one that watches the sewage.

Let’s start with the air. BioWatch is a federal program to detect the release of pathogens into the air as part of a terrorist attack on major American cities, created in 2001 in response to the anthrax attacks. Here is how it works:

As currently deployed, BioWatch collectors draw air through filters that field technicians collect daily and transport to laboratories, where professional technicians analyze the material collected on the filter for evidence of biological threats [via PCR]. The entire collection and analysis process takes up to 36 hours to detect the presence of a potential pathogen of interest.

A positive result triggers what is known as a BioWatch Actionable Result (BAR), an indication that genetic material consistent with a target pathogen was present on a BioWatch filter. Upon declaration of a BAR, local, state, and federal officials then assess relevant information and determine the course of action to pursue.

Very cool, isn’t it? Here’s what one of the air filter boxes look like:

The problem with the system, and this is a big one, is that it has literally never once been useful. Never. Not once. Every single time a BAR has been announced, the subsequent investigation has concluded that it was either a false positive or an environmental anomaly indistinguishable from something dangerous. A Department of Homeland Security page has this helpful note about it:

Out of these more than 7 million tests, BioWatch has reported 149 instances in which naturally-occurring biological pathogens were detected from environmental sources. Many of the pathogens the BioWatch system is designed to detect occur naturally in the environment, such as the bacteria that causes anthrax, which has been used as a weapon, but is also found in nature. For example, near the nation’s Southwest border there have been a number of instances where a bacterium that is endemic in the environment has been identified. Thankfully, none of the instances were actual attacks.

It also has these lines that I thought were quite funny:

The detection of commonly occurring environmental agents is not a “false positive.” Much like a home smoke detector goes off for both burnt toast and a major fire, the smoke detector is meant to notify you of a potential fire before it’s too late. BioWatch works very much the same way.

A smoke detector that has gone off 149 times over two decades and never once for an actual fire is almost certainly not a functioning smoke detector. And this particular smoke detector cost hundreds of millions to set up, and tens of millions a year to maintain! To be clear: there is no technological reason that these can’t be made better, and there are startups, such as Pilgrim Labs, that are working on improving similar air-detection systems. If curious, I found the Pilgrim’s founders interview here to be worth watching.

On the sewage side, the whole endeavor is actually going fairly well. But before we go on: monitoring the air is obvious, but why monitor sewage? Because nearly every pathogen that infects a human being eventually ends up in the toilet. Because of this, looking through sewage is perhaps the most honest epidemiological data source available, because people cannot choose not to participate.

And we’re doing very well in monitoring this sludge, or doing so-called ‘wastewater screening’. A lot of people in biosecurity complain that ‘the federal government learned nothing from COVID’, and they are mostly right, with one huge counterexample: the national wastewater surveillance infrastructure, which was largely built in response to the pandemic. The National Wastewater Surveillance System (NWSS), launched by the CDC in September 2020, established that you could detect community-level viral trends days before clinical cases appeared, using nothing more than the genetic material people flush down the toilet, without requiring any of them to consent to testing, show up at a clinic, or even know they’re sick.

But the problem with the NWSS, as it is currently deployed, is that it is a targeted system, relying on qPCR to identify specific, known threats. And among the 500-600 sites where NWSS monitoring stations are deployed, they measure three things: COVID-19, Influenza A, and RSV.

80% of them also measure three more things: Measles, H5N1, and Monkeypox.

There’s an awful lot missing, isn’t there? What about all the other types of Influenza? Norovirus? And the scarier ones too, Nipah, Ebola, Tularemia, all of them are entirely absent.

The answer is, in principle, to switch away from qPCR and do metagenomic sequencing: instead of looking for specific pathogens, you sequence everything in the sample and computationally figure out what’s there. I’ve written about metagenomics in the context of microbiomes, so you can look there for a deeper explanation on how it works.

Why isn’t anyone doing this?

In fact, there is someone doing this, and this leads us to what I’d consider one of the crown jewels of what the U.S. nonprofit-biosecurity-complex has managed to accomplish: SecureBio Detection, previously known as Nucleic Acid Observatory (NAO), which has been building a pilot metagenomics-based wastewater screening network in the US since 2021. Circa November 2025, they maintain 31 sampling sites across the US, in 19 cities, sequencing about 60 billion read pairs weekly. And they’ve already stumbled across a few interesting things, such as detecting measles in wastewater from Kauaʻi County, Hawaii and West Nile Virus in Missouri—the latter of which ended up having real, confirmed cases to go alongside it! There is an ongoing effort to have something similar at the federal level—the so-called ‘Biothreat Radar’—but it doesn’t seem to actually exist yet. But SecureBio Detection continues!

This is quite promising. This is a bonafide, national-scale attempt to detect both known and unknown biological threads, and it works! They are also doing some interesting ML work in being able to automatically detect, via a metagenomic language model, whether unknown metagenomes are simply uncharacterized, innocuous microbes (i.e. nearly all microbes) or human-targeting pathogens worth worrying about.

But, despite how good wastewater screening is, it is worth remembering that detection is not defense. This may seem like a semantic point, of course detection isn’t defense, but certainly it should allow you to defend faster or better.

But does it really?

If you’re detecting something known—a COVID variant, a resurgent influenza strain—then yes, detection may accelerate response, because you already know what to make against it. But if you’re detecting something novel, then what exactly happens next? Designing vaccines that elicit neutralizing antibodies is difficult in the best of circumstances, clinical trials take time, and, in the meantime, the underlying pathogen will continue to mutate, potentially diverging from whatever you’re designing against it. This is surprisingly under-discussed, but it is worth marinating in the fact that, yes, BioNTech’s and Moderna’s capacity to generate a COVID-19 vaccine so quickly was indeed an extraordinary feat of logistics and science, but the usage of the spike protein segment as an immunogen in the vaccine was informed by two decades of prior coronaviruses research. In the case of a brand new, chimeric virus that has no immediate cousin, a few weeks of advance notice is just a longer window in which to watch the curve steepen.

Finally, in both cases, either a natural or engineered pathogen, there exists one last problem: coordination. There is no pre-negotiated decision tree for what happens after something scary is detected, no threshold that, once crossed, triggers automatic funding for therapeutic stockpiling or accelerated clinical development. There probably should be one! But there isn’t today and, as far as I can tell, there aren’t plans for one to exist. Ultimately, the value of early warning is bounded by the speed of the response it enables, and that speed seems extremely limited today.

Machine learning may be very useful for rapid-response therapeutics

This section is me going off-script from the experts I talked to. The pipeline I will describe below does not exist in any meaningful capacity, but there are inklings of it found across the therapeutics-for-biosecurity plays out there, so it feels like the mental framework is informative regardless. As in, the logical steps mentioned here may massively diverge from what will realistically occur, but the types of models, timelines, and decision calculus used likely will not.

The Coalition for Epidemic Preparedness Innovations, or CEPI, has an initiative that identifies exactly what you’d want your government to be capable of in the case of a major pandemic: the 100 Days Mission. As in, from the day of realizing, ‘we probably should mount a response to this weird sequence we found’, therapeutic options should be ready to go within three months for population-scale deployment. It took 326 days to get the first COVID-19 vaccine authorized, and that was widely regarded as the fastest vaccine development in human history. How could 100 days be possible?

Luckily for us, they’ve defended the position at length in a paper. Long story short: this is not an unreasonable timeline if you’re in a coronavirus-y situation, where your adversary is something that millions of hours of research has already gone into characterizing. Why? Because the second you can identify the ideal immunogen—or, what you should be sticking in your vaccine to elicit the antibody repertoire that neutralizes the virus—you’re done with the major technical design challenge. Like I mentioned earlier, the spike protein was the obvious immunogen for SARS-CoV-2, informed by two decades of prior coronavirus research going back to SARS-1 and MERS. Thus, the fun little party story of BioNTech and Moderna having a vaccine candidate within days of receiving the SARS-CoV-2 sequence.

So, mRNA basically hands us our vaccine. Now we just need to deal with the two other bottlenecks: manufacturing scale-up and clinical trials. I think it’s interesting to discuss how things may be sped up here—and the arguments for how you’d speed them up are within the realm of possibility—but it does lead us off-topic, so I’ll place those in a very long footnote.1

But remember, what I’ve described so far is the rosy scenario, where we are dealing with something we already mostly understand. What about things that are wholly new? This includes not only de novo pathogens, but also mostly natural ones that have immune-escaped the established immunogens through either evolutionary or otherwise methods. For these cases, the same CEPI paper admits that things are harder, and that a 200 or 300 day turnaround time should be the goal.

But is that possible? Remember, now the vaccine design problem becomes quite difficult. Which viral protein subunit do you use as the immunogen? Which conformation elicits neutralizing versus non-neutralizing antibodies? Which epitopes are conserved enough that you’re not designing a vaccine that will be obsolete by the time it’s manufactured? These are not easy questions to answer! And if you get them wrong, you waste months manufacturing the wrong thing. The same CEPI paper from earlier optimistically states that immunogen/antigen design for these novel pathogens would take just a few months if we really worked hard at it.

But it feels like getting to this speed of development would almost certainly require immense technological leaps. One of my favorite podcast episodes was my interview with a founder of a vaccine development startup: Soham Sankaran of PopVax. In it, I ask a lot of questions about why immunogen design for vaccines is so hard, and I will paraphrase his answers in the footnotes.2 To keep it short: it’s really, really hard.

Now, the question of the evening: can machine-learning help us with this?

Probably not. At least not in a significant way anytime soon. ML seems useful in the margins for, say, figuring out how to scaffold specific immunogens of interest such that they are ‘correctly’ presented to the immune system, but we are far off from a model being able to reliably respond to a query like ‘here is the structure of the virus I am scared of, please design an immunogen that I can encode into an mRNA vaccine that will elicit broadly neutralizing antibodies’.

At least, that’s the consensus from everyone I talked to. But if we’re willing to stretch our brains a little, I think one can imagine a scenario in which ML, as it exists today, may end up being extraordinarily useful for how we respond to pandemics. And it comes down to the fact that mRNA is such a stupidly, insanely versatile platform. You don’t need to encode an immunogen in the mRNA. Instead, you could simply encode the antibodies that you’d want the immunogen to elicit.

What, you may scream, surely you can’t do that. But you can! As far back as 2021, Moderna injected adult humans with an mRNA vaccine that had, as its payload, monoclonal antibodies against the Chikungunya virus. And it worked quite well! Moderna has since shelved this particular asset, but for reasons that seem more portfolio-optimization-y than the drug not having enough efficacy. Luckily, there is ongoing work outside Moderna in exploring mRNA-encoded nanobodies, which have the advantage of being far smaller than typical antibodies, so less stressful for our weak, mammalian cells to pump out. And upon looking it up, I have discovered that I am not the first one to find this absurdly relevant to biosecurity efforts! One 2026 review paper echoes my sentiment, and expands on it: ‘mRNA-encoded antibody approaches have been explored in preclinical models of Zika virus, Ebola virus, and rabies, where a single intramuscular dose provided prophylactic and therapeutic benefits in animal models’.

Insane, right? Now, you may immediately spot problems with this. For instance: antibodies don’t last very long in our blood stream, on the order of 2-3 weeks. How useful could this possibly be in a pandemic, where circulating pathogenic material may linger around for months? But fixing this is fully within the realm of possibility. Engineering the Fc region, or the bottom section of the ‘Y’ shape of an antibody, can reliably and dramatically expand its therapeutic window. In fact, we needn’t even theorize on this, because the same 2021 Moderna paper also included these Fc mutations: 2 alterations (M428L and N434S), leading to a 69 day half life. And there is no reason to believe that this cannot be pushed even further, given that at least one anti-viral antibody has been shown to have a half-life on the order of 5-6 months.

The next question: where will we get useful antibodies from?

Modern ML methods for designing antibodies against arbitrary targets are not perfect, but they really are quite good. In 2025, the Baker lab published what is, to my knowledge, the most significant result in computational antibody design to date: a fine-tuned version of RFdiffusion that can generate de novo antibodies—VHHs, scFvs, and full antibodies—targeting user-specified epitopes. Most relevant for us, when the model was given a particular target and epitope—C. difficile toxin B and a specific epitope that had never had an antibody designed against it—the model generated moderate-affinity antibodies, with cryo-EM confirming its binding. Now, as I mentioned in the footnotes, binding to a virus is not the same thing as neutralization of a virus, and we usually only care about the latter. I agree that this is a bottleneck that ML cannot easily solve, but it also does not feel like a huge bottleneck, especially if these models work well. Consider the fact that binding is necessary, but not sufficient for neutralization, and if you just screen a bunch of binders, all generated for free, surely you can vastly speed up the process of identifying a neutralizing antibody.

Of course, in the case of a pandemic going on long enough, you could bypass all this by simply fishing out neutralizing antibodies from infected patients, or at least use those as a parent for further ML-driven optimization.

Our final problem is that pathogens usually mutate, which means that even if we turn every human into a factory of identical antibodies against a particular sequence, those same antibodies may soon become useless due to immune escape. This is why the natural immune response—as offered by either an immunogen or antigens from the pathogen itself—can be so efficacious, as the polyclonal antibody repertoire elicited by natural infection or vaccination targets dozens of epitopes simultaneously, making it extraordinarily difficult for the virus to escape all of them at once. This too is not theory: every single monoclonal antibody therapy authorized against SARS-CoV-2 was eventually rendered obsolete by Omicron and its descendants.

Are we doomed?

Let’s not give up, and instead take a closer look at what two issues we need to solve to overcome this obstacle. First, we need to choose not just any neutralizing antibodies for our vaccine, but ones that target sites where escape is costly to the virus, or functionally constrained epitopes where mutations would compromise receptor binding or some other essential function. Second, we need to deploy cocktails of antibodies targeting non-overlapping epitopes, such that the probability of simultaneous escape across all of them becomes vanishingly small.

I propose to you that there are viable ML-based solutions to both of these.

For identification of immune-escape-y-epitopes, we can look to EVEscape, a protein model from the Debora Mark’s lab at Harvard. The model combines evolutionary sequence information with structural and biophysical data to predict, for a given viral protein, which mutations are most likely to emerge and evade existing immunity. Flip the interpretation and you get the inverse: sites where EVEscape predicts low escape potential are precisely the sites where you want your antibodies to bind, because the virus cannot easily mutate away from them without crippling itself. This is not a solved problem, but models like these are surely directionally useful, and certainly better than guessing.

For cocktail design, consider EscapeMap. EscapeMap integrates deep mutational scanning (DMS) data from SARS-CoV-2 across dozens of neutralizing monoclonal antibodies with a generative sequence model to identify something very useful: negatively correlated escape routes. Two antibodies have negatively correlated escape if the mutations that evade one tend to make the virus more sensitive to the other. Cocktails built from such pairs are inherently resistant to simultaneous escape, because the virus cannot run from both at once. As published, EscapeMap is SARS-CoV-2-specific; the underlying DMS data took years to generate, and you wouldn’t have it on day one of a new pandemic. But the framework (should) generalizes to any pandemic and a DMS-esque dataset will emerge if it goes on for long enough, allowing you to eventually design broadly-neutralizing cocktails of antibodies. If we’re being especially galaxy-brained, given a sufficiently good protein model, perhaps you don’t need any DMS data at all! After designing your de novo antibodies, you could run in-silico DMS to predict how every possible mutation on the target surface would affect binding to each candidate, cross-reference those with EVEscape-style fitness predictions to filter for mutations the virus can actually tolerate, and look for the same negative correlations. I realize this isn’t fully possible today, that the impact of single-amino-acid substitutions are still badly grasped by these models, and a whole host of other failure modes. But the models will only get better.

When all of this is put together, this pipeline should allow us to do something extraordinary within weeks of a novel pathogen being sequenced:

1. Discover neutralizing antibodies against them, either via ML or patient serum.

2. Create a cocktail of antibodies with negatively correlated escape routes via in-silico screening or a DMS dataset.

3. Fc-engineer them for a long half-life.

4. Encode the whole thing into mRNA.

5. Manufacture it.

If we do this early enough, and distribute the vaccines fast enough, we could potentially kill the spread of even the most virulent pathogens. Of course, manufacturing is historically the next major bottleneck, but if our wastewater screening and ensuing rapid-responses are quick enough, we may need to manufacture orders of magnitude fewer doses.

I realize that there are many catches here, and that what I’ve presented is grossly optimistic. All of this is a multi-layered, mostly-computational solution, and every one of these layers are error prone. All antibody generation methods have plenty of failure modes, EveScape is not consistently useful across viruses (though further lines of research claim to have improved on it), EscapeMap is hyper-focused on SARS-CoV-2 and it may very be that the framework actually cannot easily transfer to new pathogens, and antibody-encoded-into-mRNA is—for however clever it may sound—still in its early days of efficacy-and-adverse-effect studying.

But each one of these are improving, and I think the trend-lines are promising. I am much more optimistic on the value of ML here than in perhaps any other layer of the biosecurity defense workflow, and time will tell how much that optimism is warranted.

Pathogen-agnostic defenses are extraordinary. But who pays for it?

Finally, the last section. This one will be short.

Everything discussed so far shares a common architectural assumption: that you know, or can figure out, what you’re looking for. This is hard! And it is made all the more difficult by the fact that the coordinated effort needed to respond to these discoveries is not something that we’re historically very good at. But there is a one category of biosecurity defense that sidesteps this problem entirely, since they work against all pathogens. And once they are deployed, they largely work for extended periods (months to years!) by themselves, with no logistical effort needed from anybody.

What are they? Far-UVC and glycol vapors.

I’m going to be honest: the more I looked into this subject, the more I found that every conceivable thing that could be written about it has been, and where it hasn’t, it’d require conversations with a lot more people and significantly lengthen this already long essay. So I’ll defer to other people here. For far-UVC I’d recommend visiting faruvc.org for an introduction, and, if you’re sufficiently convinced, aerolamp.net to pick one up for yourself. Glycol vapors have a lot less easy reading material, but there is one article published a year back by Blueprint Biosecurity—a nonprofit who also funds far-UVC work—and various related articles written on Jeff Kaufman’s blog, who works in biosecurity.

To keep it short: If we could tile the interior of enough buildings with these solutions, you could, in theory, render the entire human indoor environment continuously hostile to airborne pathogens; far-UVC through physical degradation of their DNA, and glycol vapors through (probably) desiccation. This would affect all airborne pathogens. Named ones, unnamed ones, engineered ones, ones that have never existed before and will never exist again except in the brief window between their release and their death to one of these two. And it would do all this with no harm to you. Of course, these technologies still have room to improve, but their problems are mostly ones of logistics, optimization, and scalability.

So why don’t we see these far-UVC lamps and glycol vapor fumers in every building in the world? Why aren’t we sterilizing our air the same way we sterilize our water?

You could quibble with the details here, about how far-UVC is still very expensive, the evidence base for glycol vapors is still being figured out, and the like. But it’s tough for me to consider the question of ‘why isn’t this being massively funded’ without concluding that the problem is that there is no for-profit entity that really benefits from it. The benefits of clean air are diffuse, accruing to everyone who breathes in a building, none of whom are the institution writing the check. Hospitals are the one exception, but they are a sliver of all interior environments that humans reside in, and obviously will not offer the scale necessary to put a dent into pandemics. This means that these technologies can only be deployed and studied by a very small group of hobbyists, early adopters, and academic labs.

Okay, but isn’t this the point of governments? This is a clear public good! This is territory that is hard to get perfect visibility into, but my instinct is that the evidence base for governmental-buy-in is simply difficult to produce.

A recent Works In Progress article over far-UVC had this to say:

Measuring infection control is challenging and seldom undertaken, particularly in public spaces. Epidemiological data is expensive and difficult to gather, and there is currently no way to measure the amount of viable, infectious pathogens in the air in real time. Office attendance can be tracked, but controlling for how users mix outside the office space is immensely difficult, and measuring the real-world effect of small-scale deployments in public areas is almost impossible. Studies aiming to cause deliberate disease transmission in controlled environments have failed to work in practice because they have been too small to generate enough infections.

While this is a bitter pill, there is a sweet one that it offers us: the implementation problems with pathogen-agnostic defenses are extremely ‘money-shaped’ in a way that few other biosecurity solutions are. All the subject needs is proof, in the form of randomized control trials, in aggregating individual use experiments, in subsidizing institutions to try it out—more money to push over to the ‘this obviously works’ finish-line. So, if there are any biosecurity-curious philanthropists reading this: I highly encourage you to explore far-UVC or glycol vapors.

Especially because unlike almost every other type of biosecurity solution we’ve discussed so far, these solutions will yield public benefits even in the absence of bioterrorism. In fact, the same Works In Progress article over far-UVC never even mentions biosecurity, and is focused more on public health, ending with this line: ‘Tuberculosis and coronaviruses [may] join typhoid and cholera as tragedies of the past, and seasonal flu and common colds would become rare rather than routine if clean air were as universal and expected as clean water.’.

It’s a great pitch, and I am very excited to see more deployment of these technologies in the coming years. It just feels like one of the more obvious areas to push forwards on in this field.

Conclusion

So, what should you be scared of?

I can’t speak for you, but I can say what I’m scared of. I am scared of a well-funded terrorist organization constructing their own lab, out of which they create natural pathogens—potentially with a few AI-assisted mutations to allow them to immune-escape existing defenses—using either split-order attacks or ordering from DNA synthesis companies who don’t screen. I am scared of these groups spreading it in well-populated cities or farmland. I am scared that it will either kill several million people and/or cause billions in economic damage, and though its spread will be noticed by wastewater screening, it will be months until the necessary resources are allocated to defend against it. And I am scared that all of this will happen within the next few years.

What am I not scared of? I am not scared of state-actors, because most states have too much to lose by violating the Biological Weapons Convention and, if they are willing to let loose anyway, I believe they would opt for either easier-to-use-and-control chemical or nuclear weapons instead. I am not scared of people creating extremely engineered pathogens that have capabilities far beyond existing ones—because the existing ones are already quite good and difficult enough to work with—especially because even if the AI tools get good enough to make it worth it, I believe the same AI tools will be just as useful in countermeasure design. And yes, I realize ‘attack requires one success while defense requires comprehensive coverage’, but I also believe the swiss-cheese security model will prevail. Finally, I am not scared of individual actors, because the economics of bioweapons production likely do not work in their favor. Yes, they can rent upstream services—virus production, purification—but the downstream weaponization work requires custom protocols that CROs have no economic incentive to develop. Moreover, given that the weaponization will almost definitely be a bespoke, hands-on R&D project and not one that is easily automated, it feels unlikely that nobody at the CRO will raise an eyebrow.

That’s my threat model at least. I realize it has holes. For example, it may be the case that state actors are worth worrying about, entirely because the appeal of bioweapons is that you deploy them with plausible deniability. Hard to do that with a nuke! You may also accuse me of not paying close enough attention to the trendlines, and that maybe I am correct about the 2026 threats, but not the 2030 ones, so perhaps a disgruntled salaryman will really be able to someday easily design mega-Ebola to depopulate the planet. Maybe!

Ultimately, you can get infinitely paranoid about biosecurity if you really want to, or you can assume Nothing Ever Happens, and I think where I have landed is a comfy middle ground. I am grateful that there exist people who work in biosecurity who are infinitely paranoid, and through writing this essay, I have become far more sympathetic to their viewpoint.

To end this off: in all my conversations, everyone generally agreed that an honest-to-god, bioterrorist attack is unlikely. It is a low probability event. But low probability events with civilizational consequences are still worth preparing for. The heartening thing here is the bottleneck to preparation here is almost entirely institutional, economic, and coordinative, not scientific. The disheartening thing is that fixing these ultimately requires political will, and sans a catalyzing event to unlock it, that political will does not currently exist. Of course, one could argue that perhaps we will never need it, that the Pathogen that people in this space are breathlessly building defenses against will never arrive, that it is all paranoia, tech-rotted minds coming up with entirely hallucinated demons. But that argument feels far less convincing now than before I started writing this essay, and, if I did my job right, I hope it will feel less convincing to you, too.

New-ish substacker Linch Zhang at The Linchpin has a fun post that ranks the “ways of knowing,” which he summarizes in this tier list:

The first thing I noticed is that it’s strange to rank things like “reading,” “science,” and “non-expert intuitions” in the same list, but having thought about it, it kinda makes sense. We might consider the elements in this chart as answers to the question, “If I want to know something, what techniques can I use?”, conflated with answers for “How does human civilization produce new knowledge?” This conflation makes things a little confusing, but the list is interesting nonetheless.

The second thing I noticed, because this is a topic I know well, is that cultural evolution is ranked very low, on par with folk wisdom and divine revelation.

Broadly speaking, cultural evolution is the application of evolutionary concepts from biology, like mutation and natural selection, to cultural traits. These traits cover any type of information that is socially transmitted: ideas, customs, styles, technologies, non-innate behaviors, etc. “Evolution” means the way this information changes over time, driven by mechanisms like innovation (analogous to mutation) and cultural selection (analogous to natural selection, i.e., how we choose to keep certain cultural traits and not others, and accumulate them).

Cultural evolution can be said to produce new knowledge, but Linch finds it unimpressive. In his list, “literacy is ranked highly . . . because it delivers extraordinary returns on humanity’s investment compared to, say, cultural evolution’s millennia of trial and error through humanity’s history and pre-history.” He adds that “cultural evolution (F tier) is vastly overrated.”

I’m really curious about who, exactly, is doing all this overrating, since I hold the opinion that cultural evolution is vastly underrated. In fact, it’s often not rated at all: most people aren’t even aware of it. It’s surprising to see it on the chart, to be honest. But once we assume the writer knows what cultural evolution is, it’s even more surprising to see it ranked in the “Fail” tier.

Here’s how Linch justifies it:

Let’s be specific about cultural evolution, since Henrich’s The Secret of Our Success has made it trendy. It’s genuinely fascinating that Fijians learned to process manioc to remove cyanide without understanding chemistry. It’s clever that some societies use divination to randomize hunting locations. But compare manioc processing to penicillin discovery, randomized hunting to GPS satellites, traditional boat-building to the Apollo program.

Cultural evolution is real and occasionally produces useful knowledge. But it’s slow, unreliable, and limited to problems your ancestors faced repeatedly over generations. When COVID hit, folk wisdom offered better funeral rites; science delivered mRNA vaccines in under a year.

The epistemic methods that gave us antibiotics, electricity, and the internet simply dwarf accumulated folk wisdom’s contributions. A cultural evolution supporter might argue that cultural evolution discovered precursors to what I think of as our best tools: literacy, mathematics, and the scientific method. I don’t dispute this, but cultural evolution’s heyday is long gone. Humanity has largely superseded cultural evolution’s slowness and fickleness with faster, more reliable epistemic methods.

There are several confusions here — starting with the fact that the Fiji example in The Secret of Our Success isn’t about manioc, but about taboos on eating shark while pregnant. The manioc example is from Amazonia! But okay, minor mistake, it doesn’t matter to the argument. There are bigger issues, and looking at them will help us get a much clearer picture of what cultural evolution is and why it matters.

The #1 problem in Linch’s post is his implicit definition of cultural evolution. It’s not exactly wrong, but it’s too strict. He suggests that cultural evolution takes “millennia of trial and error,” and that it is “slow, unreliable, and limited to problems your ancestors faced repeatedly over generations.” None of this is true, except the trial-and-error part. Cultural evolution does operate by trial and error, but that can happen far faster than generations or millennia, and it’s not particularly unreliable, since we can control the selection mechanism in many cases.

Here are some toy examples at various time scales, with different ways to perform the trial-and-error step and the selection step:

• Centuries and millennia: Chimpanzees try to eat termites by digging in termite nests with their hands for 100,000 years. Then one chimp randomly pokes a nest with a stick and discovers that she can get termites more efficiently this way (trial and error). Later, another chimp sees the first chimp do it and be rewarded with many juicy termites, and imitates the behavior (selection). The knowledge of using sticks slowly spreads, but it takes hundreds of years for the entire population to adopt it.

• Decades or generations: The manioc story. Various tribes of the Tukanoan people in the Colombian Amazon eat manioc, a.k.a. cassava, as their staple food. Manioc causes cyanide poisoning in the long term, but you don’t notice it with either sight or taste, and without modern science, it’s basically impossible to link manioc to disease. One tribe randomly has this habit of waiting two days before eating the manioc fiber, which detoxifies the manioc, though they don’t know it; another just eats it immediately (trial and error). Over a few decades, the members of the first tribe are overall healthier, allowing them to reproduce and survive better, while the tribe that doesn’t detoxify eventually goes extinct (selection).

• Years: Wolfgang Amadeus Mozart is messing around at his keyboard (trial and error), and invents a new kind of musical flourish in one of his symphonies. He publishes the symphony. Other up-and-coming composers find it cool, and within a few years, they also use the technique in their works, changing the course of classical music history (selection).1

• Months and weeks: Dozens of SaaS startups get founded in the AI space, taking advantage of recent LLM advances. They experiment with various projects (trial and error), and within a few weeks, some of them start getting traction since they fill a user need (selection).

• Hours and minutes: Thousands of people post jokes and hot takes on Twitter, most of which are bad and boring (trial and error); but every once in a while, one of them becomes viral because it’s particularly funny, interesting, or egregious, which encourages people to share it further (selection).

Sometimes, of course, the selection step is a totally uncontrollable, even unnoticeable phenomenon, as in the manioc story. But many other times it occurs through rational human behavior, such as users buying SaaS or choosing to retweet something they like (or dislike). The trial-and-error step can also be controlled to some extent: Mozart or startup founders can deliberately choose to experiment, or even induce creativity using specific techniques or substances.

What happened in Linch’s post, I suspect, is that he overfit on the examples in Henrich’s book. Henrich is an anthropologist, and most of the evidence he describes comes from studies of hunter-gatherer societies, humans vs. great apes, and child psychology. This makes sense for a field that tries to encompass so much (many things are cultural!) and yet is still fairly young (a couple of decades). You’ll make more progress if you study the mechanics of human culture at the boundary between humans and other animals, without the confounding effect of the layers and layers of culture that make up the life of an adult in a complex modern civilization. But the idea generalizes far beyond anthropology.2

A related error in Linch’s post is to consider cultural evolution as distinct from the other ways of knowing in the list.

It would be both boring and useless to claim that all of the items are examples of cultural evolution. To be sure, you certainly could make this argument: In a broad sense, cultural evolution covers almost every kind of knowledge process used by humans, with the exception of the knowledge encoded in our genes. But some of the items in Linch’s list are more closely associated with it than others. And those few items, in fact, cover most of it. I don’t know what’s left if you take them out from under the cultural evolution umbrella.

First, folk wisdom. Also rated as F-tier. Cited twice in Linch’s three paragraphs on cultural evolution (“The epistemic methods that gave us antibiotics, electricity, and the internet simply dwarf accumulated folk wisdom’s contributions”). It’s unclear to me what the distinction is here. Maybe the idea of accumulation? But folk wisdom without accumulation doesn’t really make sense; isn’t the point of folk wisdom that it has been transmitted across generations? Otherwise it’s just random life advice by an elderly person. It seems that Linch conflated the two concepts and should have just merged them into one, called “folk wisdom” or maybe “culturally evolved folk wisdom.”

Second, mimicry. B-tier: one of the better ways of knowing, according to the list. But if you read The Secret of Our Success, you’ll realize that this actually is the main mechanism of cultural evolution. A successful hunter gets imitated by younger aspiring hunters. Toddlers learn about the world by imitating their parents and older siblings. A core argument from Henrich is that humans are particularly good at transmitting cultural information in this way, unlike chimps and other animal species.

Third, expert intuition, classified as C-tier, and non-expert intuitions, D-tier. We humans are good at mimicking and learning, but not from just anyone. Henrich describes various quirks of human psychology that help us identify experts, or at least the most successful individuals in a particular task. This is why we’re sensitive to prestige: it’s an (imperfect) signal of expertise. So Linch’s two items on “intuition” are clearly part of cultural evolution: figuring out who to pay attention to is a key mechanism.

Fourth, literacy/reading. This is Linch’s top way of knowing, at the S+ tier. Literacy certainly isn’t needed for cultural evolution to occur, but boy, does it help with the accumulation part.

Fifth, social media, ranked as F--. For what it’s worth, I don’t fully agree that this is the absolute worst way of knowing, but regardless, it’s a clear example of trial and error and selection, as seen in my earlier Twitter example.

The other ways of knowing can also be massaged into fitting within the cultural evolution frame. Science, for example, definitely also works through trial and error, followed by selection of the best knowledge through criticism. Indeed, there are epistemologists like Karl Popper who argue that all knowledge production is evolutionary. But even if we don’t go all in, notice how the items I picked cover the full spectrum from F-- to S+. This makes it pretty hard to decide in which tier to actually put cultural evolution, if we wanted to update the list…

… which brings us to this question: should cultural evolution even be in the list at all?

In some sense, sure, it’s a way of knowing. But unlike the other items in the tier list, it’s not really something you choose to do. Most of the time, it just happens: knowledge gets created and selected, and has thereby evolved. And to the extent that you can choose to do it, by deliberately innovating and selecting, that’s an indistinguishable process from most other ways of knowing.

If I wanted to make a better list, I would split it in two, removing the conflation I mentioned earlier. One tier list for the question, “If I want to know something, what techniques can I use?” and another for “How does human civilization produce new knowledge?” Literacy would appear in the former, but not in the latter, since it’s not possible to produce knowledge new to humanity just from reading (at least until we meet literate aliens). Cultural evolution wouldn’t be in the former, but would show up in the latter, probably ranked somewhere in the middle.

For that matter, Darwinian biological evolution should be in the latter list too, ranked lower than cultural evolution. Mutation + natural selection does produce new knowledge, encoded in DNA. It’s a very lousy way to learn, it takes millions of years and it generally involves killing off every organism that makes a mistake, but it works.

I made this very quickly and didn’t put that much thought into each individual item, but here’s what my two-part list looks like:

I learned about cultural evolution when I happened to be a research intern in Joe Henrich’s lab at Harvard for half a year, back in 2016, which is also right after The Secret of Our Success was published. Ever since, I’ve been puzzling over why the idea is so little-known and often misunderstood.

Part of the answer might be that it’s surprisingly counter-intuitive, like biological evolution by natural selection. The mechanism is simple, but it’s difficult to grasp that complex living things or complex cultural traits evolve out of mutation and selection, without any sort of goal-oriented (teleological) processes.

In addition, cultural evolution is more complex than biological evolution, because it operates along more mechanisms. In biology, the information is stored in a standardized way, in genes. Genetics certainly is complex, but it’s tractable. Also, in a sexual species, genes are passed to a new organism almost exclusively when two organisms reproduce. (There are exceptions, like horizontal gene transfer, but they’re rare in complex organisms like animals and plants. We can also do this artificially with genetic engineering.) By contrast, in cultural evolution, the equivalent of a gene, the meme, is very poorly defined. This has prevented memetics from establishing itself as a formal field. And memes can spread in a multitude of methods, including sexual reproduction (parents passing their behaviors and values to their children), but also education, media, the influence of friends, etc. Not to mention that there are different levels of memes that interact in various ways: whether a group adopts the cultural practice of eating shellfish may be influenced by the “larger” cultural practice of being Jewish.

But in another sense, cultural evolution might still be underrated simply because it hasn’t been written about that much. It’s a young field, and it might still be difficult to tease the ideas apart from the few mainstream books, papers, and blog posts that have been published on the topic. Which is why I’ve been writing about it often, though rarely as the main idea of a post.

I’ve been meaning to write about it more formally several times, and for some mysterious reason I find it very difficult. So I’m happy I got to leverage the motivation of proving someone wrong. Hopefully Linch appreciates that this, too, is cultural evolution in action. A substacker writes about cultural evolution in a mistaken way (trial and error); another critiques the problems in it (selection); we collectively arrive at better knowledge. It’s not the worst way to make progress!

This is my 3rd Roots of Progress Blog Building Intensive post! Thanks to Linch Zhang for providing the fuel, to Venki for drawing my attention to Linch’s post, to BBI fellows Steven Adler, Hiya Jain, and Andrew Burleson for feedback, and to Mike Riggs for editing.

My favorite color has changed throughout my life, cycling through the entire spectrum of visible light and beyond. I don’t remember when blue was the chosen one, exactly; maybe when I was 13 or so. After that, yellow, purple, orange, green, and pink occupied the top spot for various periods. Blue never made a comeback. I saw it as a banal, common color. After all, the sky is made of it, and the sky is everywhere.

Then I realized when compiling the tech tree that blue is the most fascinating color, because it is the hardest of the common colors to create artificially.1 You can’t just take a piece of the sky and put it into a painting. And blue pigments are fairly rare in minerals, plants, and animals. So blue had to be invented, time and time again, from 4000 BC to the 21st century. It is the most technological color, and I’m willing to claim that this is why it is usually, in science fiction and elsewhere, used to represent the future.

The story of blue starts with indigo. It is an organic dye made from plants in the Indigofera genus, which grow throughout the tropical and subtropical regions of the world. The first known traces of indigo dye come from the New World, in ancient Peru, 6,000 years ago, using Indigofera suffruticosa, or anil.2 In the Old World, it was known from Africa to East Asia, but became particularly associated with India (hence indi-go), where Indigofera tinctoria was domesticated. Indigo soon became a luxury, traded from India to Greco-Roman and then medieval Europe, where the same blue dye could only be made from a less productive plant, woad or Isatis tinctoria. Eventually the “blue gold” became an important colonial crop in the Caribbean and was part of the story of slavery, next to sugar, tobacco, and cotton.

Before indigo was a thing in the Old World (that started circa 2400 BC), the Egyptians had already become obsessed with the color blue. Besides the sky, it was available in the form of semiprecious stones like turquoise and lapis lazuli, cobalt oxide (more on that later), as well as the mineral azurite, which they mined in Sinai and the Eastern Desert.

Azurite would later enjoy a fruitful career as the main blue pigment in European painting, but to the Egyptians it was costly, and besides it isn’t the most stable blue color: it degrades and fades when in contact with air. And so they created the first synthetic pigment in history: Egyptian blue. The oldest evidence of it is in a bowl dated to 3250 BC. Egyptian blue is a calcium copper silicate with formula CaCuSi₄O₁₀ or CaOCuO(SiO₂)₄. Its method of manufacturing, in a rare example of lost technology, was forgotten towards the end of antiquity, but has been plausibly reconstructed. It likely involved heating together quartz sand (silica) and some source of copper (either copper ores or scraps from the bronze industry), together with an alkali (like natron) and a calcium oxide (unintentionally added as impurities in the other materials).

In another cradle of civilization, a very similar story unfolded from about 800 BC. So similar, in fact, that it has been speculated that knowledge of Egyptian blue spread along the early silk road, all the way to China, where Han blue (together with Han purple) makes an appearance during the Zhou dynasty. Han blue has almost the same chemical formula as Egyptian blue, but replaces calcium with barium: BaCuSi₄O₁₀. It may also have been an independent invention, perhaps the work of Taoist alchemists and glassmakers. Its use declined after the Han dynasty, and few examples survive.

Much later, China would become famous for another application of blue: the “blue and white” porcelain style. The blue here comes from cobalt oxide, which had colored Egyptian faience since at least 1500 BC, though nobody at the time knew what cobalt was. You could make cobalt blue in the form of glass and then grind it into a pigment called smalt. Despite porcelain originating in China, it seems that the use of smalt for the blue and white style began in Iraq. It spread from the Middle East to China, and then from China to the rest of the world including Europe, where it would eventually allow Swedish chemist Georg Brandt to identify cobalt as an element in 1735, the first time a new metal was discovered since antiquity.3

Meanwhile, in the New World, the local indigo dye was being combined with a clay called palygorskite to create what became known as Maya blue, which was the main blue pigment in Mesoamerican art from about 800, and was still used as late as the 19th century, though it, too, was forgotten about for a while.

But none of the pigments mentioned so far, not azurite, not cobalt blue, not Egyptian blue, could rival with the purest and deepest of blues, the one that came from grinding the rare lapis lazuli stone into a powder. Lapis lazuli had been extracted from mines in what is now Afghanistan since ancient times, but was used for paint starting around the 5th to 7th centuries, for use in Zoroastrian and Buddhist religious art. This pigment became known to medieval and Renaissance Europeans as ultramarine, meaning “beyond the sea,” since it had to be imported at great cost from central Asia (which, to the Venetian merchants who mostly controlled this trade, was beyond the Mediterranean sea, I suppose). Nobody has written about this more eloquently than Scott Alexander:

Here is the process for getting ultramarine. First, go to Afghanistan. Keep in mind, you start in England or France or wherever. Afghanistan is four thousand miles away. Your path takes you through tall mountains, burning deserts, and several dozen Muslim countries that are still pissed about the whole Crusades thing. Still alive? Climb 7,000 feet through the mountains of Kuran Wa Munjan until you reach the mines of Sar-i-Sang. There, in a freezing desert, the wretched of the earth work themselves to an early grave breaking apart the rocks of Badakhshan to mine a few hundred kilograms per year of blue stone - the only lapis lazuli production in the known world.

Buy the stone and retrace your path through the burning deserts and vengeful Muslims until you’re back in England or France or wherever. Still alive? That was the easy part. Now you need to go through a chemical extraction process that makes the Philosopher’s Stone look like freshman chem lab. “The lengthy process of pulverization, sifting, and washing to produce ultramarine makes the natural pigment … roughly ten times more expensive than the stone it came from.”

Finally you have ultramarine! How much? I can’t find good numbers, but Claude estimates that the ultramarine production of all of medieval Europe was around the order of 30 kg per year - not enough to paint a medium-sized wall. Ultramarine had to be saved for ultra-high-value applications.

In practice, the medievals converged on a single use case - painting the Virgin Mary’s coat.

By the beginning of the 18th century, Egyptian blue had been long forgotten, and painters in Europe relied on indigo, smalt, azurite, and when they could get their hands on it, ultramarine. But this was the modern, enlightened era of European science. Great things were to come.

It began with a chance discovery. In Berlin around 1706, a paintmaker, Johann Jacob Diesbach, was trying to prepare red dye from cochineal.4 The details of the story are not totally ascertained, but it seems that his intended mix of cochineal insects, ferric sulfate, and potash had been tainted by another substance, perhaps bone oil from the alchemist Johann Konrad Dippel. The result was a deep blue pigment, soon to be known as Prussian blue, since Berlin was the capital of Prussia. It immediately found its niche in the art market: a deep blue, like ultramarine, but which unlike ultramarine didn’t cost more than literal gold. Within a couple of years, painters were already depicting the Virgin Mary’s robes with Prussian blue.

Thus Prussian blue became the first modern synthetic pigment. It spread far and wide, even to isolationist Japan. Large quantities of Prussian blue began entering the country around 1829, through the single trading post the Japanese allowed with Westerners, at Dejima, and very soon after, revolutionized the woodblock printing art of ukiyo-e. As early as 1831, one of the most famous works in art history was created with abundant Prussian blue.

Prussian blue is also the blue of blueprints, created with the cyanotype process, one of the first ways to make many copies of a document. The blueprint was invented in 1842 by John Herschel and became the standard for engineering drawings; it was also used abundantly to duplicate photographs. Though it has become obsolete (replaced by whiteprint and then xerography, the currently dominant photocopying technique), it survives as the word to describe any technical, detailed plan.

Prussian blue was only the first of a series of synthetic blue pigments that span the history of industrial civilization. In 1789, cerulean appears, the creation of Albrecht Höpfner in Switzerland. It is another compound of cobalt, but combined with tin: a cobalt stannate (Co₂SnO₄). It would become available to artists in paint form only in the middle of the 19th century.

Around the same period, in 1799 or 1802 (sources differ), the French chemist Louis Jacques Thénard reinvented cobalt blue. It was a commission from another chemist, Jean-Antoine Chaptal, who happened to be a minister in the government of the First French Republic. Thénard investigated the pigments at the Sèvres porcelain factory, but used a different method than the originators of smalt pigments in Egypt, Iraq, or China, using aluminium (formula: CoAl₂O₄). By the middle of the 19th century, the leader in the production of cobalt aluminate was Blaafarveværket, a large industrial enterprise in Norway.

In this golden age of blue pigment synthesis, would it be possible to create even synthetic ultramarine? Goethe, already, had noticed the blue deposits on lime kilns when visiting Sicily in 1787. The locals used it for decoration as if it were lapis lazuli. The same phenomenon was observed in limeworks in France in the 1810s, and in 1824, the Société d’encouragement pour l’industrie nationale — a government organization dedicated to further French industry in response to the industrial revolution in Britain, and led by the aforementioned Jean-Antoine Chaptal — announced a prize of 6,000 francs to anyone who could make ultramarine for much cheaper than the price of lapis lazuli. In 1826, Jean-Baptiste Guimet succeeded in Lyon. He won the prize and established a thriving business, though he kept his methods secret and, as a result, forever has to share credit with Christian Gmelin in Tübingen, who did publish the process. It involves heating up clay, sodium hydroxide, and coal together.5

Artists and decorators now had their main blue pigments. Soon, industry and science would extend the use of blue to other domains. In 1897, it became practical to prepare artificial indigo in industrial quantities, eventually replacing all use of the plant. Today 80,000 tonnes are produced per year, mostly for the purpose of dying textiles and especially the denim of jeans.

Around the turn of the 20th century, artificial food colorings became widespread, derived primarily from coal tar. One of them would become known as brilliant blue FCF or Blue No. 1. Together with Blue No. 2, which is made from indigo, it is one of the two main blue colorings, and has a strong association with the familiar-yet-mysterious blue raspberry flavor.

The 1920s saw the introduction of another synthetic pigment, phthalo blue (also known as copper phthalocyanine), perhaps in a way harking back to the copper-derived compounds of ancient Egypt. It has grown to be most widely produced blue.

Though the discovery of new pigments is a rare occurrence, it still happens. In 2009, a serendipitous discovery at the Oregon State University led to YInMn Blue, so named because it contains yttrium, indium, and manganese, and pronounced “yinmin.” It is a near-perfect blue that furthermore avoids the toxicity and environmental problems of the pre-existing pigments.

I have this half-baked theory that science fiction is associated with blue because of blue LEDs. Consider this chart, which I wrote about in an old post:

There are a number of competing hypotheses for why science fiction movie directors and video game designers overwhelmingly choose blue as the color of fictional user interfaces. They include:

• Accidental reasons from filmmaking considerations (from Mark Coleran):

  • Using simple interfaces with primary RGB colors on black looks better in film than ordinary liquid-crystal screens, so most UIs in video media is either red, green, or blue

  • Red is associated with weapons, and green with vintage electronics (which commonly used green-phosphor monochrome monitors), leaving blue as the generic and/or futuristic choice

  • Blue is easier to color-correct: a lot of filmed material tends to look blue before color correction, but you don’t need this when the image is supposed to be blue

• Cultural associations:

  • Blue fits well with science fiction thanks to associations like coldness, knowledge, otherworldliness, and creative transcendence (found in some academic paper in Korean thanks to Elicit)

  • Something something near-far Robin Hanson something something6

  • Blue is rare in nature except the for sea and sky, so “there’s something fundamentally mystical, unnatural, and inhuman about it” (from Noessel, cited in ‘Future Screens are Mostly Blue’)

• Most science fiction creators simply copy the tropes of existing science fiction, so they choose blue because it already is the “science fiction color.” (And picking something else is likely to be interpreted as an intentional deviation for a specific purpose.)

I’d guess that the actual, immediate reason for most blue in science fiction is the last one. Unless creators make a conscious artistic choice to deviate, they tend to copy what’s typical and expected in their chosen genre. Yet those norms and expectations have to come from somewhere. It’s possible that the items in the first part of the list, about decisions having to do with the techniques of filmmaking rather than with the artistic meaning of blue, are the actual cause for some early shows that snowballed into the ubiquity of blue interfaces today, but that seems a bit like post-hoc justification to me unless we can find evidence of those decisions being made.

So it’s probably cultural associations, but most of them also just kick the can further. Overall I suppose I somewhat agree with Noessel: it may well come down to the difficulty in finding blue in nature. Or finding it in technology, considering the history of blue pigments that we just went over.

And not just pigments. There is another realm in which blue has proven incredibly difficult to produce: light. In fact we found the solution so recently that it is why, I speculate, the future is still strongly associated with blue.

The story of how we produce light is a fun one, spanning all of our technological history and involving dozens of solutions, from prehistoric oil lamps to candles to coal gas to cold-cathode tubes. But most of those solutions have produced light somewhere between white and the reddish yellow of flames or black-body radiation. If you wanted blue light, you could make a bulb out of blue glass (with cobalt!) and put an incandescent filament in it. This worked okay, but blue light bulbs did tend to be less satisfying than the other colors, or at least that’s what I remember from Christmas lights when I was a kid.7

There were other strategies for blue lights: construct a tube like the familiar red-glowing neon ones, and put mercury vapor in it. Low-brightness phosphors for RGB screens. Greenish-blue vacuum fluorescent displays. So, blue light was not an unsolved problem. But it wasn’t as conclusively solved as light in the other spectral colors.

In the 1960s, light-emitting diodes started becoming practical. LEDs are the most efficient way of creating light by far, but the properties of the materials they are made of — semiconductors that emit light when traversed with an electrical current — make it much easier to generate radiation in the less energetic part of the electromagnetic spectrum. Thus the first practical LED emitted infrared light, by Texas Instruments in 1962. Later that same year, the first visible-light LED was made at General Electric, in red. Displays made of red LEDs soon became widespread in electronic devices (replacing Nixie tubes, also reddish) after some further advances by Hewlett-Packard circa 1968.

Humanity then gradually conquered the rest of the visible light spectrum, with orange, yellow, and green LEDs being developed in the 1970s. But blue remained elusive. A practical, bright blue LED would not be made, despite much research being poured into it by electronics companies around the world, until a breakthrough by Shuji Nakamura in Japan in 1993.

This completed the color spectrum and enabled us to create white LEDs, which are now quickly replacing nearly every lighting technology since they cost so little and are so customizable. We can say we have essentially “solved” lighting. Blue LEDs also enabled the first blue lasers in the mid-1990s.

This is a very recent development! For a very long time, blue would have been the color that only “future tech” could create. Then, for a brief period, it would have been the color of cutting-edge tech. Now, 30 years later, the tech exists and is widespread, but we still have the memory of that time. And furthermore no other color can take its place as the inaccessible one; we’ve conquered the entire spectrum.8

Does the futuristic quality of blue really come from LEDs? Maybe, maybe not. I’m not sure there’s a direct causal link.

But given the full history of blue pigments, I wouldn’t be surprised to find some truth in this speculative scenario: that there were just enough innovations in blue, a steady trickle of serendipitous discoveries and long-term research efforts to produce better versions of it, to keep it in the mind of humans as the color of the artificial and the cutting-edge. If you wore indigo-dyed clothes in ancient India, you were one step more removed from nature than the person who wore plain cotton. If you used blue pottery in Egypt or Iraq or China, you were clearly cooler than the people who used plain terracotta. If you hired an engineer or architect at the peak of the Industrial Revolution in the late 19th century, they’d be way more efficient at their job if they duplicated their drawings with Prussian blue instead of copying them by hand. And today, if you want to make your city a herald of the high-tech future, you decorate everything with programmable blue LEDs. No other color would do.

This post was written as part of the Roots of Progress Institute’s Blog-Building Intensive, and I thank the fellows who provided feedback: Allison Lehman, Kelly Vedi.

Spring is finally coming to Canada.

It never shows up all at once, just a flicker at first. A slant of sunlight across the kitchen table, river water rustling beneath the winter ice. It’s like this every year. And yet the long winter always makes me doubt spring will come again.

In Greek myth, Demeter is the goddess of the harvest. When her daughter, Persephone, is taken to the underworld by Hades, Demeter’s grief halts the growth of all things. A deal is eventually struck: Because she ate a few pomegranate seeds in the underworld, Persephone must spend part of each year below but will be allowed to return to Demeter every spring. Even knowing this, Demeter mourns her absence as if the return were never guaranteed.

By the end of winter, I am Demeter in abeyance, disbelieving the cycle. As if the cold will stay forever; as if the underworld has won.

But today, I stand corrected. One warm day is all it takes. The soil softens. Possibility stirs.

Science can feel like that, too.

We work inside inherited frameworks, mistaking the scaffolding of our understanding for reality itself. And then, sometimes, a shift. A sliver of evidence cracks through the surface and what once felt certain begins to thaw.

In neuroscience and immunology, just such a crack appeared in 2007. Suddenly, some conditions doctors had long diagnosed as psychosis, personality changes, cognitive decline, or just psychiatric illness, appeared instead to be effects of the immune system attacking the brain.

What had scientists discovered? A specific kind of autoimmune attack. Patients were making antibodies that targeted the NMDA receptor—a critical junction for passing information between neurons (synaptic communication). The disorder was eventually dubbed anti-NMDA receptor encephalitis, and later, when other antibodies with other targets were discovered able to cause similar conditions, autoimmune encephalitis.

The discovery marked more than a new diagnosis: namely, it challenged a foundational belief shared by both neuroscience and immunology. For most of modern science, the brain was considered “immune privileged”—a kind of no-fly zone for immune cells. The idea was that these two systems, though essential, operated on opposite sides of a biological firewall (the blood-brain barrier).

But that firewall turns out more porous than we thought. In the past several decades, neuroscience and immunology have come to a growing recognition that the immune system doesn’t just defend the body from infections. It also shapes, regulates, and sometimes disrupts what we think of as mental life.

The Pre-History of Autoimmune Encephalitis

The identification of autoimmune encephalitis was the culmination of more than 40 years of mysterious clinical observations: patients with seizures, memory loss, and mood changes who were later found to have cancer. The illness first showed up in the medical literature in 1960, when clinicians described three adults with what they called “subacute encephalitis of later adult life mainly affecting the limbic areas.”

If we’re being honest, the whole diagnosis is basically medical hand-waving. “Subacute” just means “not super-fast, but not slow either,” “encephalitis” means “something’s inflamed in the brain,” and “limbic areas” refers to a fuzzy collection of brain bits involved in emotion and memory. In other words, something weird is happening in the emotional memory parts of the brain, kind of suddenly, and we have no idea why. Not exactly diagnostic precision, but it was a start.

Many patients with this condition deteriorated rapidly, and a significant number were only identified post-mortem, when pathologists would find limbic inflammation, neuronal loss, and immune cell infiltrates—often alongside “occult” (literally, hidden) tumours like small-cell lung cancer or testicular germ-cell tumours.

Eventually, this cluster of symptoms came to be known as paraneoplastic limbic encephalitis, a term that, again, sounds more precise than it is. It basically meant: there’s inflammation in the emotional-memory circuits of the brain, and oh look, the patient also had cancer.

The prevailing theory at the time was that certain tumours were triggering an immune response that accidentally spilled over into the nervous system. But the antibodies identified in these patients weren’t targeting the outside of cells, where they could interfere with function. Instead, they recognized intracellular proteins, buried deep inside neurons. And antibodies, as any immunologist will tell you, can’t easily get inside living cells.

That made the antibodies unlikely culprits. Before 2007, they were thought to be a consequence of the disease—evidence that the immune system had already done damage. According to this model, it was a broader inflammatory process, most likely involving cytotoxic T cells, that was attacking neurons first. When those neurons died, they spilled their internal contents into the surrounding tissue, including proteins the immune system had never seen before. Then antibodies were generated against those proteins.

In other words, the antibodies were downstream, not upstream. These antibodies weren’t early warning signs or causes; they were immunological debris of damage done.

The figure below illustrates this hypothesized model: a tumour triggers a general immune activation, leading to neuronal cell death, and only then do antibodies against intracellular proteins appear—not as instigators, but as clues left at the scene.

Treatment efforts were mostly ineffective. By the time doctors figured out what was going on, neurons were often already lost. And because symptoms frequently came before a cancer diagnosis, the connection wasn’t always obvious. Patients were often misdiagnosed with psychiatric illness, admitted to the wrong ward, and left to decline. You can imagine how many cases were missed.

So, for decades, neurologists were stuck with fragments: a mysterious syndrome that attacked personality and memory, an occasional tumour, a few tantalizing antibodies, and a vague sense that the immune system was somehow involved. But nothing that tied it all together.

Then, in 2007, a change in scientific technique changed the narrative.

Previously, researchers used fixed tissue slices or biochemical assays which were enriched for proteins inside of cells (intracellular proteins) to study the disease. But in 2007, researchers using rat brain sections and live cultured neurons—techniques that preserved the delicate structure of the cell surface—observed antibodies binding in real time to the outside of neurons, where they could actually interfere with synaptic function. So, these newly identified antibodies weren’t targeting proteins buried inside cells. They were on the surface, targeting a receptor that’s essential for how neurons communicate: the NMDA receptor, a key player in learning and memory.

And that changed everything. If you remember one of the basic rules for identifying causes in science—what epidemiologists call temporality—then you’ll see why this matters: a true cause has to come before the effect. An antibody that alters the properties of a neuron’s signalling channels is very likely to be the cause, not the consequence, of cognitive impairment.

These antibodies weren’t just markers anymore: they were agents of causality, the kind of smoking gun every translational scientist dreams of finding. They didn’t kill neurons outright; they disrupted synaptic function, impairing signalling.

For the first time, medicine had uncovered a psychiatric syndrome caused by a treatable autoimmune process. And here’s the extraordinary part: the damage was reversible. When doctors treated the immune response and, where applicable, removed the tumour, patients came back to their former selves.

Once the basics of autoimmune encephalitis were understood, interesting variations were identified. Many new cases—especially in children and young adults—had no tumours at all. The initial cancer link had cracked open the mystery, but the real story was broader: an autoimmune attack, often in people with personal or family histories of autoimmunity, sometimes following infection, that targeted the brain and produced dramatic changes in behaviour and cognition.

The discovery of autoimmune encephalitis landed like an epistemic shockwave, forcing the field to reckon with just how much it had overlooked.

But it was also something else, something rare and worth celebrating: a clear win for translational neuroscience. A field that’s more Mets than Yankees when it comes to its win–loss record—plenty of grit, a lot of heartbreak, and far more swings and misses than walk-off victories. But here, finally, was an extraordinary win. A baffling clinical syndrome with a patient population hiding in plain sight. Eventually, a molecular mechanism uncovered and a treatment strategy that worked. And best of all, patients who, if diagnosed early enough, recovered.

If you followed my previous series on Alzheimer’s disease, you’ll know how rare this kind of clarity is in neurology.

You’d like to think these revolutions in medical comprehension would spread instantly to specialists worldwide, but that rarely happens. In the case of autoimmune encephalitis, it took more than recognition and publication in academic papers—it took a compelling story. In 2012, journalist Susannah Cahalan published Brain on Fire, a first-person account of her descent into apparent psychosis that turned out to be autoimmune encephalitis. Her memoir exposed a hidden epidemic; at the time, only about 10 percent of cases were being correctly diagnosed. But it also exposed something else: the power of how we think about biological systems to inhibit our ability to understand them.

The Double-Edged Sword of Metaphor

For most of the 20th century, we described the immune system as a military force, defending the borders of the body against invasion. The brain was considered “immune privileged”—too precious for warfare, sealed off by evolution from immune influence. This metaphor wasn’t just poetic; it shaped what we looked for, what we measured, and what we could see.

But reality—as ever—resists neat conceptual packaging.

We now know the immune system isn’t just a warrior—it’s a curator, a gardener, a maintainer of order, and a mediator of meaning throughout the body, including the brain.

In this series, we’ll explore how the discovery of autoimmune encephalitis has pushed us to rethink not just the relationship between the brain and immune system, but how biological systems interact, how diseases emerge, and how our frameworks for understanding can both illuminate and obscure the truth. Throughout these discussions, we’ll return to the role of language and metaphor in science—how the words and models we choose shape what questions we ask, what experiments we design, how we interpret those results and ultimately what we can discover.

The military metaphors that dominated immunology for a century weren’t wrong, exactly. They were incomplete. They were useful except when they weren’t. They helped us understand one aspect of immune function while obscuring others. This pattern—of metaphors both revealing and concealing—runs through the history of science. We need models, categories, poetic association and even myths to make sense of complexity, but we also need the flexibility to step outside them, to remember that naming is not the same as knowing.

500 Million Years of Your Friendly and Fearsome Immune System

By the time you finish reading this sentence, your immune system will have surveilled, evaluated, and responded to millions of signals within your body. Even when you’re sitting listlessly on the sofa, mad at yourself for doing “nothing,” your immune system is busy monitoring trillions of microbes living in your gut, responding to countless particles in the air you breathe, and maintaining constant surveillance of every tissue in your body. This is both amazing and a matter of survival.

But the very sophistication that keeps us alive can turn against us, and for a growing number of people, this remarkable system becomes a source of danger. Autoimmune disorders now affect nearly 17 million people in North America, or roughly 1-in-20. Different autoimmune conditions tend to emerge at different times in life: Type 1 diabetes in childhood, lupus and multiple sclerosis in early adulthood, and rheumatoid arthritis in middle age. And now we’re discovering that conditions once thought entirely psychiatric or neurological—certain cases of psychosis, memory loss, personality changes, seizures, and dementia—can be caused by the immune system attacking the brain.

What other conditions might we still fail to recognize as immune-related? Are autoimmune disorders in fact separate diseases, or do they cluster in different manifestations of broader patterns we’re only beginning to recognize? These questions are fundamental to a better understanding of why autoimmunity occurs, and more importantly, how to prevent and treat it. To get closer to that understanding, we need to look at how our immune system evolved.

Take It Back to When Life Got Complicated

The story begins with an ancient accident that gave vertebrates their most powerful defense system. It’s a story that helps explain both why our immune system is remarkably effective and why it sometimes turns against us. More crucially, it’s a story that matters to anyone who might develop an autoimmune condition—which, given current trends, could be any of us (except me, because I already have a few).

Somewhere in the ancient oceans of the Cambrian era, before dinosaurs, before mammals, when trilobites were finding their legs and vertebrates were developing backbones, something extraordinary happened inside a primitive vertebrate. A piece of rogue DNA, a transposable element (think: molecular hitchhiker with a crowbar), inserted itself into the genome near genes involved in recognizing pathogens.

Transposable elements (transposons) are selfish bits of DNA—genomic drifters that copy and paste themselves across genomes and species. (Richard Dawkins popularized the idea of “selfish genes” in the 1970s; Francis Crick and Leslie Orgel later extended it to DNA itself.) These sequences don’t prioritize serving the organism that hosts them. They serve themselves, hitching rides through evolutionary time by being exceptionally good at getting duplicated.

Where do they come from? Most likely from viruses or ancient bacteria, early genomic parasites that learned how to hack into host DNA and stay there. We don’t know exactly who left the calling card that became our essential immune gene; we only know what they left behind: a pair of enzymes capable of cutting and rearranging DNA with extraordinary precision.

What of consequence do they do? Usually, nothing. They insert, and that’s it. They’re eventually silenced, deleted, or eroded by time. But this one stuck and it rewired the immune system. This was the accident that would make complex vertebrate life possible.

This single insertion became the cornerstone of adaptive immunity—the sophisticated defense system that allows vertebrates to survive childhood, heal from injuries, fight off infections, and live long enough to develop complex social bonds and sophisticated brains. Without it, there would be no long-lived vertebrates, no social mammals, no human culture. A chance genetic event became the gateway to biological complexity.

That bit of DNA became what we now call RAG—short for Recombination Activating Gene. What it activated was nothing short of the immunological big bang.

Why the Cambrian? Multicellular life was inventing itself in real time: specialized tissues, nervous systems, body plans with appendages, eyes, guts. With this explosion of biological novelty came new ecological pressures: predators, pathogens, and symbiotic microbes. Life was no longer a solo act; it became a crowded stage and a simple, hardcoded immune system couldn’t keep up.

To survive in this newly competitive world, organisms needed a flexible immune strategy, one that could recognize unfamiliar threats, adapt to evolving pathogens, and remember past encounters. That was the niche waiting for, even desperate for, adaptive immunity. And the RAG insertion was the catalyst.

How a DNA Hitchhiker Changed Everything

We’ve already described the origin moment: a mobile genetic element—a transposon—inserted itself into the genome of a primitive vertebrate. Most of these genetic stowaways aren’t useful. They barge in, disrupt important genes, and get kicked out or silenced. But this one was different. It brought tools.

It carried genes that encoded enzymes with an unusual skill: they could cut and paste DNA with remarkable precision. Think: molecular scissors with a built-in ruler. This might sound more dangerous than useful, and it could have been, but this time the host survived and evolution took notice.

Instead of evolution wiping out the intruder, organisms with the transposon survived better. Over millions of years, the host genome began to shape it and mutations softened its sharp edges. Regulatory sequences evolved to control where and when the scissors did their cutting. The parasite was slowly domesticated, tamed, and repurposed.

Eventually, it rewired the architecture of immune recognition in vertebrates forever.

Specifically, that insertion enabled a process that lets immune cells shuffle gene segments to create a vast library of unique receptors—each one capable of recognizing a different molecular shape, like a virus or infected cell. These receptors are the key to immune detection, and in the case of B cells, they form the blueprint for antibodies.

Two-Minute Immunology Primer

Let’s take a moment to dust off your pandemic-era immunology knowledge.

Remember antibodies? Those little Y-shaped proteins everyone suddenly had strong opinions about in 2020? They’re not just floating around aimlessly—they’re made by specialized immune cells called B cells, and their job is to recognize, bind to, and help neutralize foreign invaders like viruses and bacteria.

B cells have cousins called T cells, which also play a starring role in your immune defense—responding to infected cells, coordinating responses, and helping make sure the right cells are activated at the right time.

Together, B and T cells belong to a group called lymphocytes, and they have a superpower that makes them unlike any other cells in your body: each one carries a unique molecular sensor—a receptor on its surface that can recognize a specific pathogen or infected cell. In the case of B cells, that sensor does double duty: it helps the cell recognize a threat and, once activated, serves as the blueprint for the antibodies the cell will secrete. These antibodies flood the bloodstream and tissues, each one acting like a lock that’s shaped to catch a specific molecular key—usually a fragment of a virus, bacterium, or other invader.

Their job? To tag invaders for destruction, block their function, or deliver them into the jaws of other immune cells. In short: tag, neutralize, and bag.

Here’s the kicker: your B and T cells can make trillions of different-shaped receptors. Orders of magnitude more than the number of genes you have.

How do they pull this off? That’s where RAG comes in.

Inside every developing B or T cell, long before it ever sees a virus or a vaccine, the RAG proteins go to work. They act like molecular editors, cutting and rearranging small chunks of DNA—modular gene segments called V(D)Js:

• V for variable

• D for diversity

• J for joining

These segments are like genetic Lego bricks. RAG picks one of each, snips them out and glues them together in all sorts of ways, and—voilà—a brand-new, one-of-a-kind receptor blueprint is born. This process is called V(D)J recombination, and it’s the reason your immune system can recognize almost anything, including pathogens it’s never seen before.

This is creativity by design, but also by accident. The DNA gets stitched back together by a kind of cellular duct tape (your DNA repair machinery), which introduces even more variation. Every cell ends up with a slightly different combination.

Most of these receptors won’t bind to anything useful. Some bind to intruders that can harm you. And some might bind to you yourself—to self-molecules. Which brings us to the next critical step in immune development: selection.

Developing B and T cells go through a training program—boot camp in the bone marrow or thymus—where the ones that react strongly to your own tissues are typically destroyed. This process is called central tolerance, and it’s the immune system’s first safeguard against self-attack.

And to think—it all traces back to a molecular accident in the Cambrian Era. A rogue piece of DNA jumped into a genome and, instead of wrecking the place, rewrote the rules of immune recognition. Five hundred million years later, your B cells are still running drills based on that ancient insertion event—shuffling your immune genome and adapting to molecular targets.

So far, so good.

But now comes the question that hangs over all of this: If the immune system can generate trillions of different receptors, how does it avoid turning on its own tissues, to autoimmunity?

The short answer is: it doesn’t. Not always.

The Classical View: How Autoimmunity Slips Through the Guardrails

The very mechanisms that give rise to incredible receptor diversity also open the door to autoimmunity. Some self-reactive cells slip through training. Others mutate later in life. Some stay quiet for years until an environmental trigger wakes them up. Occasionally, the immune system gets confused—mistaking part of your body for an invading pathogen.

This is how researchers believe autoimmune diseases like lupus, rheumatoid arthritis, and autoimmune encephalitis arise.

But it’s not the only possibility. And as we’ll explore later, this entire explanation rests on a framework—a powerful framework that is often predictive, but not a complete one.

Let’s take a moment to introduce the classical model of how autoimmunity arises. This is the theory that underpins nearly every clinical trial, therapeutic strategy, and standard of care today.

It’s a model I know personally. I’ve lived inside its logic. And thanks to therapies based on it, I’m—after adding coffee—a functional human being.

According to the classical model, your immune system is built to distinguish self from non-self. Most self-reactive cells are eliminated during development (this is central tolerance). For those that escape, there’s a backup system—peripheral tolerance—which includes regulatory T cells, immune checkpoints, and protective barriers around sensitive organs like the brain.

Autoimmunity, in this model, arises when:

• A self-reactive B or T cell slips through the central and peripheral checkpoints;

• It encounters an environmental trigger—like an infection or tissue damage—that activates it; and

• It begins to expand and form a clone that mistakenly attacks the body.

This classical model is also known as clonal selection theory, and for over 50 years, it has dominated our understanding of autoimmune disease. It’s largely internally consistent and helps explain why many patients respond to therapies like B cell depletion, T cell suppression, or cytokine blockade.

Like many people with an autoimmune diagnosis, I’ve benefited profoundly from immunosuppressive treatments that emerged from this framework. They’ve pulled me back from the edge more than once.

And yet… the closer you look, the more cracks begin to show in the clonal selection theory of autoimmunity. We have some puzzles that we’ll encounter throughout this series that prompt the following questions:

• Why do some healthy people have autoantibodies without any symptoms?

• Why do some autoimmune diseases wax and wane?

• Why are some diseases so targeted (just one organ) while others affect the whole body?

• And why do treatments based on this model sometimes just fail?

These are signs that something deeper is going on—that this classical model, while useful, is incomplete. And that’s where we’re headed in this series. I cannot promise complete answers, but we can journey through the literature together.

Importantly, the classical model didn’t stop us from discovering autoimmune encephalitis, which is the great victory we will be dissecting in the next post. It didn’t stop us from saving lives.

But it has stopped us from connecting what we’ve learned into a broader, unifying theory—one that might not just treat autoimmune disease but transform how we harness the immune system to fight a much wider range of disorders.

In upcoming posts, we’ll examine the cracks in the clonal selection framework, explore ideas like molecular mimicry, degeneracy, pleiotropy, randomness, and the idea that self-recognition isn’t a static boundary, but a dynamic negotiation between immune cells, tissues, and environment.

Stay tuned—and when the day feels heavy, consider this: you are the living legacy of a story that began 500 million years ago with a wandering transposon and a lucky break. It might just add a little bounce to your step.

As we all know, the march of technological progress is best summarized by this meme from Linkedin:

Inventors constantly come up with exciting new inventions, each of them with the potential to change everything forever. But only a fraction of these ever establish themselves as a persistent part of civilization, and the rest vanish from collective consciousness. Before shutting down forever, though, the alternate branches of the tech tree leave some faint traces behind: over-optimistic sci-fi stories, outdated educational cartoons, and, sometimes, some obscure accessories that briefly made it to mass production before being quietly discontinued.

The classical example of an abandoned timeline is the Glorious Atomic Future, as described in the 1957 Disney cartoon Our Friend the Atom. A scientist with a suspiciously German accent explains all the wonderful things nuclear power will bring to our lives:

Sadly, the glorious atomic future somewhat failed to materialize, and, by the early 1960s, the project to rip a second Panama canal by detonating a necklace of nuclear bombs was canceled, because we are ruled by bureaucrats who hate fun and efficiency.

While the Our-Friend-the-Atom timeline remains out of reach from most hobbyists, not all alternate timelines are permanently closed to exploration. There are other timelines that you can explore from the comfort of your home, just by buying a few second-hand items off eBay.

I recently spent a few months in one of these abandoned timelines: the one where the microwave oven replaced the stove.

First, I had to get myself a copy of the world’s saddest book.

Microwave Cooking, for One

Marie T. Smith’s Microwave Cooking for One is an old forgotten book of microwave recipes from the 1980s. In the mid-2010s, it garnered the momentary attention of the Internet as “the world’s saddest cookbook”:

To the modern eye, it seems obvious that microwave cooking can only be about reheating ready-made frozen food. It’s about staring blankly at the buzzing white box, waiting for the four dreadful beeps that give you permission to eat. It’s about consuming lukewarm processed slop on a rickety formica table, with only the crackling of a flickering neon light piercing through the silence.

But this is completely misinterpreting Microwave Cooking for One’s vision. Two important pieces of context are missing. First – the book was published in 1985. Compare to the adoption S-curve of the microwave oven:

When MCfO was published, microwave cooking was still a new entrant to the world of household electronics. Market researchers were speculating about how the food and packaging industries would adapt their products to the new era and how deep the transformation would go. Many saw the microwave revolution as a material necessity: women were massively entering the workforce, and soon nobody would have much time to spend behind a stove. In 1985, the microwave future looked inevitable.

Second – Marie T. Smith is a microwave maximalist. She spent ten years putting every comestible object in the microwave to see what happens. Look at the items on the book cover – some are obviously impossible to prepare with a microwave, right? Well, that’s where you’re wrong. Marie T. Smith figured out a way to prepare absolutely everything. If you are a disciple of her philosophy, you shouldn’t even own a stove. Smith herself hasn’t owned one since the early 1970s. As she explains in the cookbook’s introduction, Smith believed the microwave would ultimately replace stove-top cooking, the same way stove-top cooking had replaced campfire-top cooking.

So, my goal is twofold: first, I want to know if there’s any merit to all of these forgotten microwaving techniques. Something that can make plasma out of grapes, set your house on fire and bring frozen hamsters back to life cannot be fundamentally bad. But also, I want to get a glimpse of what the world looks like in the uchronia where Marie T. Smith won and Big Teflon lost. Why did we drift apart from this timeline?

Out of the frying pan, into the magnetron

Before we start experimenting, it’s helpful to have a coarse intuition of how microwave ovens work. Microwaves use a device called a magnetron to emit radiation with wavelengths around 5-10 cm, and send it to bounce around the closed chamber where you put your food. The idea that electromagnetic radiation can heat stuff up isn’t particularly strange (we’ve all been exposed to the sun), but microwaves do it in an odd spooky way. Microwaves’ frequency is too low to be absorbed directly by food molecules. Instead, it is just low enough that, in effect, the electric field around the molecules regularly changes direction. If the molecules have a dipole moment (as water does), they start wiggling around, and the friction generates plenty of heat.

As far as I can tell, this kind of light-matter interaction doesn’t occur to a noticeable degree anywhere on Earth, except in our microwave ovens. This is going to be important later: the microwave is weird, and it often behaves contrary to our day-to-day intuitions. (For example, it’s surprisingly hard to melt ice cubes in the microwave. This is because the water molecules are locked in a lattice, so they can’t spin as much as they would in a liquid.) Thus, to tame the microwave, the first thing we’ll need is an open mind.

With that in mind, let’s open the grimoire of Microwave Cooking for One and see what kind of blood magic we can conjure from it.

The book cover, with its smiling middle-aged woman and its abundance of provisions, makes it look like it’s going to be nice and wholesome.

It’s not going to be nice and wholesome.

Microwave cooking is not about intuition. It’s about discipline. The timing and the wattage matter, but so do the exact shape and size of the vessels. Smith gives us a list of specific hardware with exceedingly modern names like the Cook’n’Pour® Saucepan or the CorningWare™ Menu-ette® so we can get reproducible results. If you were used to counting carrots in carrot units, that has to stop – carrots are measured in ounces, with a scale, and for volume you use a metal measuring cup. Glass ones are simply too inaccurate for where we are going.

The actual recipe section starts with the recipe for a bowl of cereal, which I am 70% sure is a joke:

Whenever a cooking time is specified, Smith includes “(____)” as a placeholder, so you can write in your own value, optimized for your particular setup. If your hot cereal is anything short of delicious, you are invited to do your own step of gradient descent.

A lot of recipes in the book involve stacking various objects under, above, and around the food. For vegetables, Smith generally recommends slicing them thinly, putting them between a cardboard plate and towel paper, then microwaving the ensemble. This works great. I tried it with onion and carrots, and it does make nice crispy vegetables, similar to what you get when you steam the vegetables in a rice cooker (also a great technique). I’d still say the rice cooker gives better results, but for situations where you absolutely need your carrots done in under two minutes, the microwave method is hard to beat.

But cardboard contraptions, on their own, can only take us this far. They do little to overcome the true frontier for microwave-only cooking: the Maillard Reaction. Around 150°C, amino acids and sugars combine to form dark-colored tasty compounds, also known as browning. For a good browning, you must rapidly reach temperatures well above the boiling point of water. This is particularly difficult to do in a microwave – which is why people tend to use the microwave specifically for things that don’t require the Maillard reaction.

But this is because people are weak. True radicals, like Marie T. Smith and myself, are able to obtain a perfectly fine Maillard reaction in their microwave ovens. All you need is the right cookware. Are you ready to use the full extent of microwave capabilities?

Tradwife futurism

In 1938, chemists from DuPont were trying to create a revolutionary refrigerant, when they accidentally synthesized a new compound they called teflon. It took until the early 1950s for the wife of a random engineer to suggest that teflon could be used to coat frying pans, and it worked. This led to the development of the teflon-coated frying pan.

In parallel, in 1953, chemists from Corning were trying to create photosensitive glass that could be etched using UV light, when they accidentally synthesized a new compound they called pyroceram. Pyroceram is almost unbreakable, extremely resistant to heat shocks, and remarkably non-sticky. Most importantly, the bottom can be coated with tin oxide, which enables it to absorb microwave radiation and become arbitrarily hot. This led to the development of the microwave browning skillet.

In the stove-top timeline where we live, the teflon-coated pan has become ubiquitous. But in the alternate microwave timeline, nobody has heard of teflon pans, and everybody owns a pyroceram browning skillet instead.

I know most of you are meta-contrarian edgelords, but nothing today will smash your Overton window harder than the 1986 cooking TV show Good Days, where Marie T. Smith is seen microwaving a complete cheeseburger on live TV using such a skillet.

I acquired mine second-hand from eBay and it quickly became one of my favorite objects. I could only describe its aesthetics as tradwife futurism. The overall design and cute colonial house drawings give it clear 1980s grandma vibes, but the three standoffs and metal-coated bottom give it a strange futuristic quality. It truly feels like an object from another timeline.

The key trick is to put the empty skillet alone in the microwave and let it accumulate as much heat as you desire1 before adding the food. Then, supposedly, you can get any degree of searing you like by following the right sequence of bleeps and bloops.

According to Marie Smith, this is superior to traditional stove-top cooking in many ways – it’s faster, consumes less energy, and requires less effort to clean the dishes. Let’s try a few basic recipes to see how well it works.

You’ll microwave steak and pasta, and you’ll be happy

Let’s start with something maximally outrageous: the microwaved steak with onions. I’d typically use olive oil, but the first step in Smith’s recipe is to rub the steak in butter, making this recipe a heresy for at least three groups of people.

The onions are cooked with the veggie cooking method again, and the steak is done with a masterful use of the browning skillet.

I split the meat in two halves, so I could directly compare the orthodox and heretical methods.2 The results were very promising. It takes a little bit of practice to get things exactly right, but not much more than the traditional method. The Pyroceram pan was about as easy to clean as the Teflon one. I didn’t measure the energy cost, but the microwave would probably win on that front. So far, the alternate timeline holds up quite well.

As a second eval, I tried sunny-side up eggs. On the face of it, it’s the simplest possible recipe, but it’s surprisingly hard to master. The problem is that different parts of the egg have different optimal cooking temperatures. Adam Ragusea has a video showcasing half a dozen techniques, none of which feature a microwave.

What does Marie Smith have to say about this? She employs a multi-step method. Like with the steak, we start by preheating the browning skillet. Then, we quickly coat it with butter, which should instantly start to boil. This is when we add the egg, sprinkle it lightly with water, and put it back in the oven for 45 (___) seconds. (Why the water sprinkling? Smith doesn’t explain. Maybe it’s meant to ensure the egg receives heat from all directions?)

Here again, I was pleased with the result – I’d go as far as saying it works better than the pan. With that success, I went on to try the next step of difficulty: poached eggs.

Poached eggs are my secret internal benchmark. Never in my life have I managed to make proper poached eggs, despite trying every weird trick and lifehack I came across. Will MCfO break my streak of bad luck?

Like for veggies, the egg is poached in the middle of an assemblage of multiple imbricated containers filled with specific amounts of water and pre-heated in a multi-step procedure. We are also told that the egg yolk must be punctured with a fork before cooking. (What happens if you don’t? The book doesn’t say, and I would rather not know.)

The recipe calls for 1 minute and 10 seconds of cooking at full power. Around the 1 minute and 5 seconds mark, my egg violently exploded, sending the various vessels to bounce around the walls of the oven. And listen, as I said, I came to this book with an open mind, but I expect a cookbook to give you at least enough information to avoid a literal explosion. So I wrote “LESS” in the “(____)” and never tried this recipe again.

The rest of the book is mostly made of variations of these basic methods. Some recipes sound like they would plausibly work, but were not interesting enough for me to try (for example, the pasta recipes primarily involve boiling water in the microwave and cooking pasta in it).

All in all, I think I believe most of the claims Smith makes about the microwave. Would it be possible to survive in a bunker with just a laptop, a microwave and a Cook’n’Pour SaucePan®? I think so. It probably saves energy, it definitely saves time washing the dishes, and getting a perfect browning is entirely within reach. There were failures, and many recipes would require a few rounds of practice before getting everything right, but the same is true for stove-top cooking.

On the other hand, there’s a reason the book is called Microwave Cooking for One and not Microwave Cooking for a Large, Loving Family. It’s not just because it is targeted at lonely losers. It’s because microwave cooking becomes exponentially more complicated as you increase the number of guests. I am not saying that the microwave technology in itself cannot be scaled up – if you really want to, it can:

But these industrial giant microwaves are processing a steady stream of regular, standard-sized pieces of food. Home cooking is different. Each potato comes in a different size and shape. So, while baking one potato according to MCfO’s guidance is easy and works wonderfully, things quickly get out of hand when you try baking multiple potatoes at the same time. Here is the sad truth: baking potatoes in the microwave is an NP-hard problem. For a general-purpose home-cooking technology, that’s a serious setback.

The weird thing is, the microwave maximalists of the 1980s got the sociology mostly right. People are preparing meals for themselves for longer and longer stretches of their lives. Women are indeed spending less time in the kitchen. The future where people cook For One – the one that was supposed to make the microwave timeline inevitable, arrived exactly as planned. And yet, the microwave stayed a lowly reheating device. Something else must be going on. Maybe the real forking path happened at the level of vibes?

Microvibes

To start with the obvious, the microwave has always been spooky, scary tech. Microwave heating was discovered by accident in 1945 by an engineer while he was developing new radar technologies for the US military. These are the worst possible circumstances to discover some new cooking tech – microwave manufacturers had to persuade normal civilians, who just watched Hiroshima on live TV, to irradiate their food with invisible electromagnetic waves coming from an object called “the magnetron”. Add that to the generally weird and counterintuitive behavior of food in the microwave, and it’s not surprising that people treated the device with suspicion.

Second, microwave cooking fell victim to the same curse that threatens every new easy-to-use technology: it became low-status tech. In Inadequate Equilibria, Eliezer makes a similar point about velcro: the earliest adopters of velcro were toddlers and the elderly – the people who had the most trouble tying their shoes. So Velcro became unforgivably unfashionable. I think a similar process happened with microwaves. While microwave ovens can cook pretty much any meal to any degree of sophistication, the place where they truly excel is reheating shitty canned meals, and soon the two became inseparable in the collective mind, preventing microwaves from reaching their full potential for more elaborate cuisine.

Third, compared to frying things in a pan, microwave cooking is just fundamentally less fun. I actually enjoy seeing my food transform into something visibly delicious before my eyes. But microwave cooking, even when done perfectly right, gives you none of that. You can still hear the noises, but not knowing what produced them makes them significantly more ominous. Some advanced recipes in MCoF call for 8 minutes at full power, and 8 minutes feel like a lot of time when you are helplessly listening to the monstrous anger of the oil, the stuttering onions’ rapid rattle, and the shrill, demented choirs of wailing pork ribs.

With all that said, I do think Microwave Cooking for One is an admirable cookbook. The recipes are probably not the finest cuisine, but they’ll expand your cooking possibilities more than any other recipe book.3 What I find uniquely cool about Marie T. Smith is that she started with no credentials or qualifications: she was a random housewife who simply fell in love with a new piece of technology, spent a decade pushing it to its limits, and published her findings as a cookbook. Just a woman and a magnetron. You can just explore your own branch of the tech tree!

Let’s not oversell it – if your reference class is “tech visionaries”, maybe that’s taking it a bit too far. If your reference class is “Middle-aged Americans from the eighties who claim they can expand your horizons using waves”, then Marie T. Smith is easily top percentile.

Is our way of representing the laws of nature actually the truth, or are physical laws better understood as one possible set of approximate compressions we happened to stumble upon? I would argue for the latter. I would also argue that we may be at a unique moment in history. A point in time where it is becoming possible, for the first time, to systematically explore many alternative cohesive systems for formulating natural laws. But let’s start with the first point.

Consider this: the human brain is a strange substrate on which to conduct science.

1. In raw capacity to perform operations, our brain is comparable to a supercomputer.

2. Our working memory is quite mediocre. We can hold references to highly complex concepts in mind (like World War 2), but we can never consciously attend to more than a tiny fraction of the complete whole at once.

3. Our capacity for arithmetic is comically poor relative to our overall brain capacity. At one point, a few hundred years ago, the ability to perform calculations that a birthday card could now handle was used as a proxy for human intelligence. How good are you at long division?

Clearly the first trait is beneficial for conducting science. Clearly the second and third are not. For the most part, scientific progress only accelerated once we could externalize working memory and computation. Writing was one early mechanism for this. What can be achieved through writing emulates what could be achieved without it, if our memory and compute were far better. Modern computers and machine learning represent our latest steps in this externalization.

I would argue that the way we describe and handle natural laws is inevitably shaped by both our cognitive strengths and weaknesses. The purpose of natural laws is to enable prediction. We find ways to isolate and compress certain aspects of reality into forms that are tractable for us to reason about. And most often, we are specifically interested in performing concrete predictive computations.

If we view physics as a search for optimal data compression, we must ask: optimal for whom? Clearly the same underlying truth can be expressed in different ways. In the US, fuel economy is tracked in miles per gallon. In Europe, we use litres per 100 km. Neither is inherently worse, but each guides intuition in subtly different directions. The same is true of Newton versus Einstein.

Any good compression has a scope — a domain of validity — but also a certain ease of computability. Every measurement contains uncertainty; every prediction carries acceptable error. Rather than viewing physical laws as discoveries of metaphysical truth, I propose we treat them as solutions to an optimization problem.

Given experimental data, find a set of representations that:

• Minimises prediction error E

• Minimises description length L

• Minimises computational cost C

• Maximises scope S

We can, and will, have many overlapping representations. Some might minimize computational cost; others might maximize scope. With infinite compute and time, we could keep track of individual particles, and emergent properties like temperature would just introduce errors. This, however, would be utterly infeasible in reality.

Physics then becomes a search over a Pareto surface defined by simplicity, accuracy, and generality, with many valid and overlapping representations.

• Newton sits on one region of that surface: tiny E, moderate C, large S.

• Einstein occupies another: minimal E, large C, very large S.

• Quantum mechanics yet another: minimal E, very large C, very large S.

But why would these be unique optima? They are optima that happened to be accessible to humans, under human constraints. They also followed from prior definitions and compressions, dependent on the haphazard path our scientific history took. This is why I suspect that an unconstrained optimizer might settle on entirely different primitives. Newton and Einstein did not find truth in a vacuum. They located regions on the Pareto surface that were reachable given what had come before, shaped by our sensory priors, our evolutionarily crafted intuitions, and our computational limitations.

The Problem of Time

There is a deep suspicion in physics that one of our most fundamental variables may be something of an illusion: time.

As humans, we experience time as a sequence of distinct, fleeting moments. This is consistent with our weak working memory. We may be forced to process events sequentially simply because we lack the buffer to handle them globally. We stitch snapshots together and call it flow. This is likely also tied to the timescales at which human intervention needs to happen.

Now consider a tree. If a tree possessed intelligence, it would almost certainly treat time differently from us. Our concept of individual snapshots would not be the most evolutionarily advantageous architecture for a rooted organism responding to slow, sweeping arcs of environmental change. The tree might perceive time as large, mean gradients held in a vast working memory, spanning years. What tangible statistically significant changes emerging over months and years ought the tree to optimize for? This is obviously speculative.

What is not speculative is the kind of artificial intelligences we have begun creating — Large Language Models in particular. A LLM possesses a context window vastly larger than human working memory, and no concept of time that has been shaped by physical survival. A LLM might not represent a novel — or a physical process — as a sequence of words to be read one by one. Instead, it can hold the entire narrative arc as a single, simultaneous, richly structured representation.

To a mind with a massive context window, the concept of time as a linear progression of snapshots might seem terribly inefficient, a lossy compression. Rather than t as a variable that ticks forward, such a mind might compress reality into arcs or events-as-whole-objects, perceiving the full trajectory of a ball as a single static shape rather than a moving point. For most humans, a parabola plotted on a graph is an abstraction. Not so for a mind that naturally thinks in whole trajectories.

Would the laws we have formulated not be influenced by our particular vantage on time? I find it unlikely that they would be immune to it. Our intuition for time is also deeply non-relativistic. We evolved at slow speeds in a weak gravitational field. A mind evolved for survival in a regime where special relativity is decisive, perhaps navigating near a black hole or moving at a significant fraction of the speed of light, would have a radically different intuitive prior for time. Concepts like simultaneity, as we currently conceive of them, would be almost meaningless.

We can speculate on what those alternative architectures might look like:

• The Deep-Time Mind: a mind working at longer timescales with a vastly larger working memory, which views events as four-dimensional structures in spacetime rather than a sequence of nows. Just as we could view an apple as a thousand microscopic slices, but have no reason to, such a mind might have no reason to slice time at all.

• The Relativistic Mind: a mind fine-tuned for survival in high-velocity environments, for whom space and time are intuitively stretchy, codependent substrates, a lived experience rather than a complex mathematical derivation.

None of these minds would likely settle on the same priors for t that we have.

The question then becomes almost inevitable: if we fed a sufficiently advanced machine learning algorithm a large amount of physical data and tasked it with finding patterns and compressions — redundancies in reality — what would it find? How many overlapping, yet distinct and valid, compressions would we find, if we ran and re-ran the experiment?

The Contingency of Our Representations

There are many examples where we already know that different representations overlap. Newton’s laws and relativity cover much of the same ground. Within a certain domain and accuracy threshold, they say the same thing — but one is far more computationally efficient. There is a reason we have not shelved Newton since discovering Einstein, even though the scope overlap is 100 %.

Lagrangian and Newtonian mechanics are largely equivalent in scope, yet in practice one can be computationally superior in certain contexts while the other dominates elsewhere.

There is also one more human constraint worth noting here: high-fidelity long-term memory and retrieval. If this were orders of magnitude better in humans, perhaps we would not only seek to generalise our laws, but also to fragment them into a myriad of highly specialised compressions — limited in scope, but computationally optimal within their domain of validity. Allow me to direct your attention to Quake III Arena.

The Quake Constant Problem

There was a time when compute was a severe bottleneck in computer games. Any kind of motion through a 3D representation of space depends heavily on square roots and inverse square roots. We know how to compute these quite exactly, but what if there is not enough compute available to do so in a timely manner, as it was when Quake III was being developed? The necessary solution was to find a computationally optimal compression that worked well enough for the game’s required scope and accuracy:

The fast inverse square root constant in Quake III (0x5f3759df).

A human physicist might call it numerology. Ugly. Ad hoc. Not a law of nature. But within its narrow band of validity, it is exactly that. Under strict compute constraints, it may even be locally optimal.

How many such local optima exist in physics? Ugly, hyper-local compressions that look meaningless to us but are computationally superior within bounded regimes? Human taste filters these out (they seem devoid of deeper meaning) but perhaps that reflects only our inability to simultaneously manage generalisations and local optimisations in a structured way. An optimizer without aesthetic prejudice would not discard these valid and useful compressions.

And how many Newtonian, Lagrangian, or Einsteinian mechanics might there be — slightly different in scope, accuracy, and efficiency — that a system starting without priors could discover? How many broad scope unifications, necessary in the most complex, integrated cases? Finding different optimal compressions on a Pareto surface might give us some insight.

True Mathematics and Representational Contingency

This ultimately links to a deeper suspicion I have about mathematics: I doubt integers are fundamental.

They feel fundamental because we are organisms that individuate. I am one, you are two. We deal with singular objects in a world of discrete things. But according to one of our best theories, quantum mechanics, this discreteness may be illusory. Position, velocity, time are all indefinable in the way we commonly think about them. Reality is smeared out, without sharp borders. At the fundamental level, discreteness dissolves into wavefunctions; identity dissolves into process.

In this regard, intervals seem to be more fundamental than counts. And for many processes, tracking logarithms is more natural than tracking counts. This is true for our perception of sound and light, to name a couple of examples. Might there exist cognitive architectures for which logarithms are primitive and addition is derived? No five fingers acting as the starting point, so to speak. The foundation of our mathematics is initially rooted in whole numbers and Euclidean geometry.

I highly doubt this is the only valid starting point. And perhaps there are even useful compressions that cannot be expressed or proven within our system? Gödel tells us that no sufficiently rich formal system is complete. Maybe any such system tends to emerge through specific modes of interaction with reality? The “unreasonable effectiveness of mathematics” in describing physical reality has long puzzled scientists and philosophers. But perhaps this is not such a mystery. Human mathematics may have evolved precisely to compress certain physical observations.

Different axioms and starting points might be equivalent. There might be axioms seemingly unrelated to Euclidean geometry that, by necessity, lead to Euclid’s axioms being true, and vice versa. Yet no system can probably contain the full truth. Or, as Plato might have put it: perhaps each mathematical system is but one shadow cast by a deeper mathematical reality, if such a thing even exists.

Regardless: For two beings occupying the same underlying reality, what is deducible within their respective mathematical systems would likely overlap substantially, but not completely. Not if their modes of interaction with reality differ enough. Einsteinian spacetime, for instance, necessitates moving beyond Euclidean geometry. In real spacetime, parallel lines can cross.

Once again, I find myself wondering: what would a machine learning system find, if it searched the space of possible compressions? To name one tantalising example: calculus discretises the continuous through limits and infinitesimals, a move that feels natural to us. But what might calculus look like for a mind that experiences continuity as primitive? A mind finding whole numbers to be alien and impossibly exact edge-cases.

Instead of rediscovering F=ma, the AI might express mechanics through continuous deformation fields in phase space. In this alien physics, our concepts of ‘force’ and ‘mass’ might never explicitly appear. Yet, the predictions would match our experiments perfectly. It would be a physics of flow rather than objects—empirically equivalent, but conceptually unrecognizable.

The Actual Proposal

Back to the main point. What might we find through letting Machine-Learning find different compressions? The architectural details matter only insofar as they make possible a concrete experiment allowing us to:

1. Train an optimizer on experimental data.

2. Reward compression along multiple axes.

3. Reward finding many distinct points on the Pareto surface.

4. Reward different balances of Scope, Accuracy, Description Length, and Computational Cost.

5. Allow it to invent its own variables.

6. Allow it to request data needed to differentiate between competing generalisations.

7. Let it search.

Then observe:

• Do independent runs converge?

• Do they rediscover our laws?

• Do they produce alternative but equivalent compressions?

• Do they invent primitives that feel alien but are predictively powerful?

• Do they find compressions more computationally efficient for specific domains?

• Do they discover different unifications with different scopes?

• Do they produce generalisations diverging from our current laws in ways that can be empirically tested?

If convergence occurs, our laws may be structurally privileged. If divergence occurs, physics may be representation-contingent. Either result is interesting.

A further refinement: data fed to such a system could come in three forms. “Simulated” data, generated from our current best understanding with added uncertainty; “experimentally measured” data from real observations; and “specifically requested” data specifically generated to probe the limits of a given compression’s domain of validity, helping distinguish between competing generalisations, asked for by the model itself.

Finally, a thought that has long hovered at the edges of my mind: what if String Theory had been discovered before Relativity? Would that even have been possible? I raise this only as an example of how contingent our path through physics may have been. This is not a value judgement about which theory is more fundamentally true. It is simply a thought I find tantalising.

There are so many interesting possibilities to explore here, and the only way to obtain real answers is through experimentation. I have a number of concrete ideas for how such an exploration might be started, but those details are for another time.

On paper, titanium should be the world’s #1 structural metal. There isn’t another metal with such great overall properties– incredible toughness and specific strength, light weight, exceptional corrosion resistance, and performance even at extreme temperatures. Titanium is also abundant, the 9th most common element in the earth’s crust—cheap enough as ore to use in paint pigment and plastic filler.1 And making titanium metal should take only half as much energy as making aluminum.2 It should be cheap and ubiquitous.

But in reality, titanium is a rounding error in the world metals market – produced 5000x less than iron and almost 200x less than aluminum. Half a century after the US and Soviet governments conjured a commercial titanium industry from scratch to feed their cold war machines, we’ve seen virtually no progress in making the metal cheap or abundant. Learning curves that manifested quickly for aluminum and stainless steel simply didn’t appear for Ti—its “idiot index”, the cost of Ti parts as a multiple of the cost of ore and embodied energy, remains >10x that of steel. At $25-$50/kg, titanium is just too expensive to be widely used.

The total lack of progress keeps titanium in a reverse-Goldilocks zone where it loses to steel and aluminum on cost, and to composites on weight3, and so is used only where its lightness and toughness are absolutely essential. Outside of aerospace, defense, corrosion-resistant process equipment, artificial joints and premium sporting goods, titanium is practically irrelevant.

‍Why is titanium so expensive?

Here’s how we get Ti today:

• We mine TiO2 as ilmenite ore, which is widespread and fairly concentrated at 40-65% TiO2. (~$1/kg Ti)

• The ilmenite is purified to synthetic rutile (90-95% TiO2) (~$2/kg Ti) then carbo-chlorinated to TiCl4 which can be distilled for very high purities. ($3-4/kg Ti)

• Next comes the main act: the Kroll process, reducing TiCl4 with magnesium metal TiCl4+2Mg -> 2MgCl2 + Ti. This is a touchy operation in which TiCl4 is meticulously metered into molten magnesium – too fast and the heat liberated in the reaction would vaporize the Mg metal. Since we have to remove the Mg and MgCl as gases, and the process happens well under Ti’s melting temperature, the metal we get initially is in porous and hard-to-use “sponge” form --full of bubbles like lava rock (~$6-8/kg Ti).

• The porous sponge is crushed, ground, mixed with alloying metals (useful Ti most often shows up as Ti90Al6V4), pressed, welded into 10-15 ton “consumable electrodes”, and homogenized in a vacuum arc remelter (VAR). (~$15/kg Ti)

• The ingots are then whittled down to saleable parts with sundry processes and techniques, which themselves are painstaking for many of the same reasons that Ti is so awesome as a structural metal – it’s hard enough to destroy tooling, forms a thick oxide film in O2, melts only at a staggering 1670ºC etc. … (~$25-60/kg Ti)

The cost of titanium roughly doubles with each of these steps—and there are lots of steps!4

Why are there so many steps here? Part of it is that titanium is both electropositive and fairly refractory—you need the uncomfortable combination of very high temperatures and a very reducing environment to get it into a useful form. But the bigger reason is that titanium is reactive—ferociously so. Molten Ti dissolves nearly every other metal. It readily forms oxides, nitrides, hydrides, and carbides. Titanium powders will ignite in carbon dioxide, and even in helium with as little as 9% oxygen. Sometimes this reactivity can be annoying, requiring an inert environment, or causing embrittlement or loss of material. Other times it can be deadly.5

(But wait – isn’t titanium meant to be corrosion resistant? Shouldn’t it behave more like a staid transition metal than an electron-shooter like sodium? Like aluminum, titanium forms a passivating oxide on its surface that imparts its famous corrosion resistance, until about 600ºC. At that point, the titanium below the oxide skin begins to draw in oxygen at a meaningful rate, embrittling the part and exposing it to attack from whatever’s around.)

Titanium’s reactivity is especially punishing when you try to process it into useful shapes, which requires high temperatures where oxygen and anything else around gets in and ruins it. Consider for example just the ingot breakdown step. In order to make the ingots more pliable, producers heat them up to high temperatures and forge them in open air. As the surface area of the ingot increases, so does oxygen pickup, forming a brittle hide you have to shave off. That’s up to 50% of the metal you made, discarded useless on the mill room floor—just to forge an ingot! Frictions like this add up at every step on the value chain. Since Ti is only transformed incrementally at each step in this process, it all adds up to significant capital, time, labor and material loss tied into each step.

It’s not a matter of optimization or learning rates

As bad as things seem, there’s good reason to believe that the way we make titanium today is pretty close to its cost floor.

The TiCl4 precursor to Kroll is a high-volume commodity chemical (>7 MMT/yr) that costs ~$4/kg Ti. Making TiCl4 is a mature process that’s substantially optimized already. Kroll also requires magnesium metal which is recycled on-site for ~$1.2/kg Ti6 — only a small multiple of the energy costs involved. Throw in a 15% profit margin, and you have arrived at the VAT excluded price for the cheapest Ti sponge quoted on the Shanghai Metal Market: ~$6/kg. So Kroll costs hardly more than its material inputs. 60 years of Ti production have whittled most of the inefficiencies away. The real cost and complexity comes later, in trying to work Ti sponge into useful parts.

Clearly, what titanium needs is reimagining of the whole process that reduces the number of steps in the value-add chain. Since cost roughly doubles with each step, eliminating one or two steps could easily save more than improving each one by 10%.

This fact is widely recognized—and has been for half a century. As Brian Potter writes in his excellent history of the Ti industry, even as the Kroll process was first scaling in the 1950s, newer and better ways of making titanium were widely expected to materialize. William Kroll himself expected that an electrolytic process similar to Hall-Heroult for aluminum would come to replace his method.

But the sad truth is that there have been many serious and well-funded attempts to reinvent Ti production over the last 50 years, and all have failed to meaningfully displace the Kroll-centric approach we use now. Let’s look through what has been tried, and see what we can learn:

The cruel lessons from 60 years of failing to reinvent titanium

People have tried all sorts of methods: electrolysis, plasma, alternative reductants, additive manufacturing. None ever produced full-density Ti cheaper than Kroll:

1960s – Dow-Howmet/TIMET/RMI electrolysis – Electrolysis of TiCl₄ to porous Ti.

1994 – Hydrogen plasma reduction – Thermal plasma for reduction of TiCl₄ to Ti powder.

2000 – Armstrong process – Na reduction of TiCl₄ to Ti powder.

2000 – Farthing-Fray-Chen (FFC) Cambridge & Ono-Suzuki – Electrolysis of TiO₂ to produce porous Ti.

2003 – Cardarelli-Rio Tinto electrolysis – Direct electrolysis of molten TiO₂ to molten Ti.

2003 – Ginatta electrolysis – Electrolysis of TiCl₄ in fluoride/chloride salt to make molten Ti.

2004 – MER Corp electrolysis – Electrolysis of Ti oxycarbide to porous Ti.

2012 – ADMA process – Hybrid H₂/Mg reduction of TiCl₄ to TiH₂ powder.

2013 – Hydrogen Assisted Magnesiothermic Reduction (HAMR) – Hybrid H₂/Mg reduction of TiO₂ to Ti powder.

Here’s our longer review of all the attempts to reinvent Ti, and the lessons we’ve drawn from their failure.

We take away four main lessons:

• Start with TiCl4. Trying to start from TiO2 makes things much harder, and can’t meaningfully reduce costs.

• Any new process will have to make dense Ti directly if it’s going to be useful. Kroll already makes porous Ti efficiently. The problem is that porous Ti simply isn’t a valuable product.

• Beware additive techniques. These are the most promising of the alternative approaches, but signs indicate that they won’t reach the scales we’d need to revolutionize Ti.

• Beware electrolysis. It’s the most frequently attempted and least successful of the alternative approaches. Ti has several cruel chemical features that make it especially challenging to manipulate in solution.

1) Trying to start with TiO2 makes things much harder

Most of the attempts reinvent Ti since the early 2000s have tried to substitute TiO2 for Kroll’s TiCl4 starting material. That’s a way to save a few bucks on feedstock costs-- and the only possible route to getting Ti as cheap as Al or Mg. But unfortunately, today none of those projects produce titanium at commercially relevant tonnages. We think that much of the problem has to do with the hassle of purifying TiO2.

The vast majority of titanium used on earth is sold not as Ti metal, but as white TiO2 pigment. TiO2’s whiteness depends strongly on its purity. So the TiO2 pigment industry has come up with an excellent way of making it as pure as possible: carbo-chlorinating synthetic rutile, fractionally distilling the various volatile metal chlorides until just the TiCl4 remains, and then burning the purified TiCl4 in oxygen to get back to pure TiO2.

As complicated as this sounds, the chloride distillation route is just about as good a way as is thermodynamically possible for purifying a metal as chemically “sticky” as Ti. Since Ti metal has to meet strict purity standards (~0.1wt% for most elements), trying to make metal directly from TiO2 saddles you with a serious purity problem right at the start.

TiCl4 also serves a valuable function in keeping oxygen out of your reaction mixture. Oxygen is fiendishly soluble in Ti metal—so much so that its presence makes metallothermic reduction almost impossible.

Just say no to TiO2. TiO2 that is pure enough to be a feedstock for Ti metal is more expensive than TiCl4 anyway, and makes the reduction of Ti vastly harder. We want to delete steps, but not at the cost of pure titanium.

2) Porous Ti is not a valuable product

Most of the inefficiencies in the Kroll process have been whittled away over the last 60 years. At this point, Ti sponge sells at a price that is very nearly the sum of the cost of TiCl4, a high-volume commodity chemical, and the cost of producing magnesium with a well tamed electrolytic process. What exactly would you improve on?

Imagine you did invent a magical new process for making porous Ti sponge. Say it’s half the capital cost of Kroll and a quarter of the energy cost—just above the thermodynamic minimum for making Ti. You buy the cheapest pure feedstock that gets you the purity you need, TiCl4 at ~$4/kg. If you want a 10% profit margin, that leaves you only ~$2.3/kg Ti to finance, operate, and maintain your process if you’re going to compete with ~$7/kg Kroll sponge. Kick in $1.35/kg Ti for capex (half of Kroll’s capex of $30/kg Ti per year amortized at 8% for 20 years)7, $0.5/kg Ti for electricity costs, and ~$0.5/kg Ti for labor and maintenance—and even though your process is magical, it’ll still be borderline unfinanceable if all you can sell is Kroll-equivalent sponge!

Titanium sponge just isn’t where the value is. Saving a few tens of cents on $7/kg Ti that will later be turned into mill products that cost as much as $50/kg Ti won’t move the needle.

3) Additive manufacturing is unlikely to give titanium the volume it needs

So if sponge isn’t where the value is, maybe we need an all-new way to turn sponge Ti into finished parts?

A lot of people have seen it this way in recent years, and so there’s been a lot of excitement about additive manufacturing for Ti parts. It makes sense in some ways. After all, subtractive manufacturing sucks for titanium – its hardness and low thermal conductivity chew through tooling8, and its reactivity costs you material in high temp processing. Both Ti and additive manufacturing are mostly used at the top of the market where expense hardly matters. And a lot of the newfangled ways you might make Ti give you spongy or powdered metal that needs additive post processing to get to full density anyway.

But while additive manufacturing with Ti powder does some specific things very well, one thing it won’t do is make Ti cheaper than it is today. Additive manufacturing tools are generally less capital intensive but have lower throughput, while conventional methods get economies of vertical scale. This has played out in the market for a few years now and cost analyses show that there’s a breakeven manufacturing volume above which conventional Ti production is more cost-effective. The exact value varies but this paper suggests that conventional Ti casting beats additive manufacturing at scales above a few hundred parts (valve bodies made with hot isostatic pressing vs. conventional forming), and this analysis of a landing gear reports that high pressure die casting is more cost effective than selective laser sintering above ~40 parts produced.

There are also lingering concerns that the main approaches to additive Ti leave micro-pores in the final metal product that concentrate stress and make weaker overall parts, especially under cyclic loading. Fear of microporosity pushes additive manufacturing to use high-cost ($70-200/kg) ultra-spherical powders in order to minimize pore volumes9. But concerns over part strength have persisted especially among the usual sectors for early adoption—defense, aerospace, and medical equipment– where failures can be catastrophic and the cost for Ti is borne easily.

Additive manufacturing does shine for low production volumes of highly complicated parts that conventional manufacturing struggles to make at all (see SpaceX’s SuperDraco). But an abundant Ti future needs a new method that can make Ti at high volumes and low cost and full strength, and that method probably won’t be additive manufacturing.

4) Electrolysis isn’t out of the question but it comes with severe challenges

Electrolysis seems like a logical approach to making titanium. After all, the development of an electrolytic process for aluminum took it from a precious metal used only in jewelry to a ubiquitous structural material in just a few decades. Could it unlock similar cost-savings for titanium?

Probably not, we think:

Ti has too many valence states in solution. Titanium, unlike aluminum, adopts multiple valence states which engage in a phenomenon called redox cycling in an electrolyzer. Ti2+ can flow across the cell in one direction, drop off an electron, and flow backwards as Ti3+. So current flows through the cell for no reason, generating no titanium metal and producing nothing but joule heat! Electrolytic magnesium produced for the Kroll process at ~30% voltage efficiency and >90% current efficiency is just a more efficient way of packing electrons into a titanium feedstock.

Ti melts at too high a temperature, but lowering the temperature makes things worse. Part of the magic of the Hall-Heroult process for aluminum is cryolite, which lowers the melting temperature of Al2O3. Since Al metal also has a relatively low melting temperature (660ºC), a Hall-Heroult cell can operate at reasonable temperatures and still pour molten aluminum into shapes that can be sold directly to customers.

If only that were possible for a Ti electrolyzer, it could delete the costly post-processing steps in making Ti parts, and make finding compatible electrolyte and electrolyzer materials that much easier. But Ti’s 1670ºC melting point makes that impossible.

You could make building your electrolyzer much easier by running it below titanium’s melting point. Some people have tried this. But at meaningful current densities, the product will then be dendritic titanium – tree-like metal crystals embedded in frozen electrolyte. What would it take to make this titanium useful? First, you must clean the titanium of the salt and then melt it down to make homogenous titanium stock – the same steps that make Kroll’s sponge titanium expensive. A non-starter.

Overall, electrolysis is the most-frequently attempted and least successful of the alternative approaches to Ti. People will keep trying—but we will be surprised if anyone gets it to work.

‍Tempering our expectations

People who love titanium tend to compare it to steel and aluminum since Ti’s base properties are so much better. They’ll also point out that the ore is cheap and abundant and that the metal costs vastly more than it should, a staggering 60x over its theoretical cost floor (the cost of ore plus energy). That kind of argument makes it easy to convince yourself that Ti should be a main structural material if not the main structural material of the future—if only we could find a better way to make it.

But it’s worth offering a moderating opinion, both on how cheap titanium can really get, and on how useful it is.

Why you can’t say that titanium beats aluminum or steel categorically.

Consider the bicycles used in the Tour de France, ultra-premium machines made of the best possible materials. From 1904 to 1994 the Tour de France was won on steel bikes. From 1995 to 1998 it was won on aluminum bikes. And then from 1999 until now it’s been won exclusively on carbon fiber bikes. Why was Ti passed over completely?

Strength-to-weight ratio isn’t everything. Larger-diameter tubes in Al bike frames are stiffer than Ti and can even be stronger—not based on material properties, but based on geometry. In a bending tube in a bike frame, the critical buckling load scales with the cube of the outer diameter, but only linearly with stiffness. If you want a stiff and strong frame, or one that’s easy to shape into aerodynamic forms, aluminum with its lower density and better formability can end up as the winning choice.

This type of dynamic repeats all across industry. Properties like density end up being just as important as specific strength, and lesser metals that are easier to form into the right geometry end up competing with and even beating titanium.

Why titanium will always be pricier than Al or steel.

Hopefully we’ve landed this point by now: Ti is extremely hard to work with. The reasons why are written in its atomic structure. It’s semi-refractory and a sponge for nearly everything around it. The purity of the feedstock and reducing agent, inertness of the reactor volumes, and metal handling at elevated temperatures are crucial aspects that will increase capital and labor costs and decrease material utilization, no matter how clever your process is. Unlike titanium, aluminum and steel smelters operate in atmospheric conditions, the molten metals can be poured in open air to form shapes that brokers can store on shelves, and they are both compatible with common high temperature materials like alumina, silica, and graphite.10 None of that works for titanium.

Where does this all leave us?

Back when it was young and promising, ARPA-E ran a program called METALS which sought to reinvent many of the light metals for the modern era. They estimated substitutional elasticities for different structural metals to come up with some cost targets, how cheap these metals would have to get to really change the world. For Ti it was $10/kg at the ingot level. That’s the cost where they thought Ti could start to meaningfully displace steel and aluminum parts.

Is that possible? Despite all the harsh lessons over the last few pages, we think so.

All this reflection on past failed attempts to reinvent titanium production makes us think that the straightest line to cheaper titanium is to delete steps wherever possible and to not shy away from energy intensity. Paying extra energy for purification and reduction could get you to fewer total steps and direct castability. Fewer steps is key since titanium is a refractory metal that can soak in nearly everything around it, leaving it worse off. Every step wastes material and makes it harder to meet purity standards.

At this intermediate stage in our brainstorm on the topic, what looks best to us is a single-step metallothermic reduction of titanium chloride that produces molten titanium directly. That’s like a more energy-intense version of Kroll. With 150% of Kroll’s capex and opex you could make ingots that hit the ARPA-E target. Better yet, using more energy in the reduction step could leave you with molten Ti so you could skip the ingot step entirely, getting you customer-ready bars and sheets for less than ingot prices. If somebody could do that, we think they could beat Kroll Ti by wide margins.

It’d be hugely energy intensive—but that could be OK. People are already making metals off-grid with dirt cheap electricity. We’ve said before why we shouldn’t be shy of energy intensity in materials production— falling energy costs have always been a main driver for lower material costs, and energy-intensity has always gone hand in hand with building the world out of much cooler materials.11

High energy intensity, direct castability, deleting steps. If you’re also thinking about titanium along similar lines, please get in touch.

Introduction

Neurotech is complicated. This is because you need to understand at least five fields at once to actually grasp what is/isn’t possible: electrical engineering, mechanical engineering, biology, neuroscience, and computer science. And, if you’re really trying to cover all the bases: surgery, ultrasound and optical physics as well. And I’ve met relatively few people in my life who can operate at the intersection of three fields, much less eight! As a result, I’ve stayed away from the entire subject, hoping that I’d eventually learn what’s going on via osmosis.

This has not worked. Each time a new neurotech startup comes out, I’d optimistically chat about them with some friend in the field and they inevitably wave it off for some bizarre reason that I would never, ever understand. But the more questions I asked, the more confused I would get. And so, at a certain point, I’d just start politely nodding to their ‘Does that make sense?’ questions.

I have, for months, been wanting to write an article to codify the exact mental steps these people go through when evaluating these companies. After talking to many experts, I have decided that this is a mostly impossible task, but that there are at least a few, small, legible fractions of their decision-making framework that are amenable to being written out. This essay is the end result.

My hope is that this helps set up the mental scaffolding necessary to triage which approaches are tractable, and which ones are more speculative. Obviously, take all of my writing with a grain of salt; anything that touches the brain is going to be complicated, and while I will try to offer as much nuance as possible, I cannot promise I will offer as much as an Actual Expert can. Grab coffee with your local neurotech founder!

Questions

1. How relevant are the state measurements to the application?

At least some forms of neurotech, like brain-computer-interfaces, perform some notion of ‘brain state reading’ as part of their normal functionality.

Well, what exactly is ‘brain state’?

Unfortunately for us, ‘brain state’ lies in the same definitional scope as ‘cell state’. As in, there isn’t really a great ground truth for the concept. But there are things that we hope are related to it! For cells, those are counts of mRNA, proteins found, chromatin landscape of the genome, and so on. For brains, there are four main possibilities to get at a notion of state:

1. Measure the spiking activity of singular neurons (very invasive)

2. Measure the activity of local field potentials (can be slightly less invasive)

3. Measure hemodynamics (blood flow or oxygenation) changes (can be non-invasive, though higher-res invasive)

4. Measure electromagnetic fields outside the skull (usually non-invasive)

There is an ordering here; at the top, we have measurements that are closest to the actual electrical signaling that (probably) defines moment-to-moment neural computation. As we move down the list, each method becomes progressively more indirect, integrating over larger populations of neurons, longer time windows, and/or more layers of intermediary physiology.

This is perhaps overcomplicating things, but there’s one also, slightly more exotic approach not mentioned here (and that I won’t mention again), called biohybrid devices. In these systems, neurons grown ex-vivo are engrafted to a brain, and those neurons are measured directly, so it’s sort of an aggregate measure like LFP, but also it’s technically able to measure single spikes.

But keep in mind: none of these actually work at understanding the full totality of every single neuron firing in a brain, which is a largely physically intractable thing to perform. Which is fine and fair! Understanding totalities is a tall bar to meet. But it does mean that whenever we stumble across a new company, we should ask the question: how relevant is their method of understanding brain state to the [therapeutic area] they actually care about? Superficial cortical hemodynamics won’t reveal hippocampal spiking, 2-channel EEG won’t decode finger trajectories, and so on.

With this context, let’s consider Kernel, a neurotech company founded by the infamous Bryan Johnson in the mid-2010’s. Their primary product is called Kernel Flow, a headset that does time-domain functional near-infrared spectroscopy (TD-fNIRS) to measure brain state, which tracks blood oxygenation by measuring how light scatters through the skull. In other words, this is a hemodynamics measurement device.

It is non-invasive, portable, and looks like a bike helmet (which is an improvement compared to many other neurotech headsets!).

One common thing you’ll find on most neurotech websites is a ‘spec sheet’ of their device. For most places, you’ll need to formally request it, but Kernel helpfully provides easy access to it here.

In it, they note that the device has an imaging rate of 3.76Hz, which means it’s taking a full hemodynamic measurement about every 266 milliseconds across the surface of the brain. This is fast in absolute terms, but slow on the level of (at least some) cognitive processes, which often unfold on the order of tens of milliseconds. For example, the neural signatures involved in recognizing a face or initiating a movement can happen in less than 100 milliseconds. And to be clear, this is not something that can be altered by increasing the sampling rate; the slowness is inherent to hemodynamic measurements in general.

This means that by the time Flow finishes one hemodynamic snapshot, many of the neural events we care about have started and finished.

The spec sheet also notes that the device comes with 4 EEG electrodes, which have a far higher sampling rate of 1kHZ, or 1,000 measurements per second. At first glance, this seems like it might compensate for the sluggish hemodynamic signal by offering access to fast electrical activity. But in practice, 4 channels are entirely insufficient for learning really anything about the brain. Keep in mind that clinical-grade usually operates at the 32-channel-and-above level!

I found one paper that investigated the localization errors of EEG’s—as in, can you correctly place where in the brain a spike is occurring—across a range of channels: 256, 128, 64, 32, and 16. Not even 4! Yet, even at the 16-channel level, spatial localization was incredibly bad; one example of its failure case being that it mis-localized a temporal-lobe spike to the frontal lobe. Past that, noise like muscle and eye movement artifacts often dominates the EEG signal at the lowest channel counts.

And, again, this was on 16 channels! One can only imagine how much worse 4 channels is.

Of course, 4-channels of EEG data clearly offer something. In the context of the device, they may serve as a coarse sanity check or a minimal signal for synchronizing with the slower hemodynamic measurements. Which maybe is enough to be useful?

But we may be getting ahead of ourselves by getting lost in these details. It is entirely irrelevant to consider the absolute value of any given measurement decision being made here, because, again, what actually matters is the relevancy of those measurements to whatever the intended use case is. Clearly the devices measurements are, at least, trustworthy. But what is it meant to be used for?

Well…it’s vague. Kernel’s public messaging has shifted over the years—from “neuroenhancement” and “mental fitness” to, most recently, “brain biomarkers.”. I am not especially well positioned to answer whether this final resting spot is relevant to what Kernel is measuring, but it feels like it is? At least if you look at their publications, which do show that the device is capable of capturing global brain state changes when under the influence of psychoactive substances, e.g. ketamine. So, even if hemodynamics doesn’t meet the lofty goal of being able to detect face recognition, that’s fine! Static-on-the-order-of-minutes biomarkers are fully within their measuring purview.

Does that make Kernel useful? I don’t know the answer to that, but we’ll come back to the subject in a second.

2. What are the costs and burdens for the user?

In short: a device must earn its place in a patient’s life.

The historical arc of neurotech companies lay mainly in serving desperate people that have literally no other options: ALS, severe spine damage, locked-in syndrome, and the like. The giants of the field—Synchron, Blackrock Neurotech, and Neuralink—have all positioned themselves around these, and so their maximally invasive nature is perfectly fine with their patients. Now, fairly, Synchron apparently doesn’t have the greatest reputation and Blackrock is somewhat old-fashioned, so Neuralink could be considered the only giant, but all three did pop up a lot during my research!

Blackrock Neurotech are the creators of the Utah Array, which remains the gold standard for invasive, in-vivo neural recording. Neuralink, the newest and most-hyped, have iterated on the approach, developing ultra-thin probes that can be inserted into the brain to directly record signals. Synchron has the least invasive approach, with its primary device being an endovascular implant called the Stentrode, allowing neural signals to be read less invasively than a Utah Array or Neuralink (from a blood vessel in the brain rather than in the parenchyma), though at a severe cost of signal quality.

You could find faults with these hyper-invasive neurotech companies on the basis of ‘how realistically large is the patient population?’, but you can’t deny that amongst the patient population that does exist, they’d certainly benefit!

So…if you do spot a neurotech company that is targeting a less-than-desperate patient population, you should ask yourself: why would anyone sign up for this? Why would an insurance company pay for it? And most importantly, why would the FDA ever approve something with such a lopsided risk-reward ratio? This is also why you see a lot of neurotech companies pivot toward “wellness” applications when their original clinical thesis doesn’t pan out. Wellness doesn’t require FDA approval or insurance reimbursement! But it also doesn’t require the device to actually work.

But even if a neurotech company is targeting a less-than-desperate patient population and aren’t trying to push them towards surgery, it’s still worth thinking about the burdens they pose!

Neurotech devices can be onerous in more boring ways too, so much so that they can completely kill any desire for any non-desperate person to use it. One example is a device we’ve talked about: the Kernel Flow. Someone who I chatted with for this essay mentioned that they had tried it, and had this to say about it:

“[the headset] weighs like 4.5lbs. That is so. fucking. uncomfortable.”.

Now, it may be the case that the information that the device tells you is of such importance that it is worth putting up with the discomfort. Is the Kernel Flow worth it? I don’t know, I haven’t tried it! But in case you ever do personally try one of these wellness-focused devices, it is worth pondering how big of a chore it’d be to deal with.

3. How much is the approach ‘fighting physics’?

Speaking of ‘building things for less desperate patients’, two big neurotech names that often come up are Nudge and Forest Neurotech (the founder of whom I talked to for this article, who has since moved to Merge Labs).

Both of these startups are focusing on brain stimulation for mental health, though Forest’s ambitions also include TBI and spinal cord injuries. Depression, anxiety, and PTSD can be quite awful, but only the most severely affected patients (single-digit percentages of the total patient population) would likely be willing to receive a brain implant. And both of these companies are fully aware of that, which is why neither of them do brain implants.

But, even if you aren’t directly placing wires into the brain, there is still some room to play with how invasive you actually are. I think it’d be a useful exercise to discuss both Nudge and Forest’s approaches—the former non-invasive, the latter invasive (albeit slightly less invasive than a Neuralink, which goes directly into the brain parenchyma)— because they illustrate an interesting dichotomy I’ve found amongst neurotech startups: the degree to which they are attempting to ‘fight’ physics.

At the more invasive end, there’s Forest Neurotech. Forest was founded in October 2023 by two Caltech scientists—Sumner Norman and Tyson Aflalo—alongside Will Biederman from Verily. They’re structured as a nonprofit Focused Research Organization and backed by $50 million from Eric and Wendy Schmidt, Ken Griffin, ARI, James Fickel, and the Susan & Riley Bechtel Foundation. Their approach relies on ultrasound, built on Butterfly Network’s ultrasound-on-chip technology, that sits inside the skull but outside the brain’s dura mater; also called an ‘epidural implant’. Still invasive, but again, not touching the brain!

At the less invasive end, there’s Nudge, who just raised $100M back in July 2025 and has Fred Ehrsam, the co-founder of Coinbase, as part of the founding team. They also have an ultrasound device, but theirs is entirely non-invasive, and comes with a nice blog post to describe exactly what it is: …a high channel count, ultrasound phased array, packed into a helmet structure that can be used in an MRI machine.

So, yes, both of these are essentially focused ultrasound devices meant for neural stimulation, though I should add the nuance that Forest’s device is also capable of imaging. But, despite the surface similarities, one distinct split between the two is that, really, Nudge is attempting to fight physics a lot more than Forest.

Why? Because they must deal with the skull.

Nudge’s device works by sending out multiple ultrasound waves from an array of transducers that are timed so precisely that they constructively interfere at a single millimeter-scale point deep in the brain, stimulating a specific neuron population, usually millions of them. It is not dissimilar to the basic principle as noise-cancelling headphones, but in reverse: instead of waves cancelling each other out, they add up. The hope is that all the peaks of the waves arrive at the same spot at the same moment—constructive interference—and you get a region of high acoustic pressure that can change brain activity. As a sidepoint: you’d think this works by stimulating neurons! But apparently it can work both via stimulation or inhibition, depending on how the ultrasound is set up.

How is the Nudge approach fighting physics?

First, there’s absorption. The skull soaks up a substantial chunk of the emitted ultrasound energy and converts it into heat. One study found that the skull causes 4.7 to 7 times more attenuation than the scalp or brain tissue combined.

Second, aberration. Because the skull varies in thickness, density, and internal structure across its surface, different parts of your ultrasound wavefront travel at different speeds, so, by the time the waves reach the brain, they’re no longer in phase. If the whole point of focused ultrasound is getting all your waves to constructively interfere at a single point, the skull messes that up, and the intended focal spot gets smeared, shifted, or might not form properly at all.

And, finally, the skull varies enormously between individuals. The “skull density ratio”—a metric that captures how much trabecular (spongy) bone versus cortical (dense) bone you have—differs from person to person, and it dramatically affects how well ultrasound gets through.

Now, to be clear, Nudge is aware of all of these things, and the way they’ve structured their device is attempting to fight all these problems. For example, Nudge talks a fair bit about how their device is MRI-compatible. This is great! If you want to correct for aberrations (and for everyone’s brain being a different shape), you need to know what you’re correcting for, which means you need a detailed 3D model of that specific patient’s skull, which means you need an MRI (or better CT). You image the skull, you build a patient-specific acoustic model, you compute the corrections needed to counteract the distortions, and then you program those corrections into your transducer array. Problem solved!

Well, maybe. Fighting physics is a difficult problem, and we’ll see what they come up with. While there is already a focused ultrasound, FDA-approved device that has been used in thousands of surgeries similar to Nudge’s that can target the brain with millimeter-scale accuracy (albeit for ablating brain tissue, not stimulating it, but the physics are the same!), it is an open question whether Nudge can dramatically improve on the precision and convenience needed to make it useful for mental health applications.

On the other hand, Forest, by bypassing the skull, is almost certainly assured to hit the brain regions they most want, potentially reaching accuracies at the micron scale. Remember that these differences cube, i.e. the number of neurons in a 150 micron wide voxel vs. a 1.5 millimeter wide voxel is (1500^3)/(150^3) =1,000 times more neurons. So it’s safe to say that the Forest device is, theoretically, 2-3 orders of magnitude more precise in the volumes it interacts with than Nudge is. Now, Forest still isn’t exactly an easy bet, given that they now have to power something near an organ that really, really doesn’t like to get hot, figure out implant biocompatibility, and a bunch of other problems that come alongside invasive neurotech devices. But they at least do not have to fight the skull, and are thus assured a high degree of precision.

There is, of course, a reward for Nudge’s trouble. Nudge, if they succeed, also gets access to a much larger potential patient population, since no surgery is needed. This is opposed to Forest, who must limit themselves to a smaller, more desperate demographic.

As with anything in biology, there is an immense amount of nuance I am missing in this explanation. People actually in the neurotech field are likely at least a little annoyed with the above explanation, because it does leave out something important in this Nudge versus Forest, non-invasive versus invasive, physics-fighting versus physics-embracing debate: how much does it all matter anyway?

4. Do they know whether their advantages translates to clinical benefit?

The brain computer interface field is in a strange epistemic position where devices are being built to modulate brain regions whose exact anatomical boundaries aren’t agreed on (and may even diverge between individuals!), using mechanisms that aren’t fully understood, for conditions whose neural circuits are still being studied.

Because of this, despite all the problems I’ve listed out with going through the skull, Nudge will almost certainly have some successful clinical readouts. Why? It has nothing to do with the team at Nudge being particularly clever, but rather, because there is already existing proof that non-invasive ultrasound setups somehow work for some clinically relevant objectives.

Nudge is fun to refer to because they have a lot of online attention on them, but there are other players in the ultrasound simulation space too, ones who are more public with their clinical results. SPIRE Therapeutics is one such company and they, or at least people associated with the company (Thomas S Riis), have papers demonstrating tremor alleviation (n=3), chronic pain reduction (n=20), and, most relevant to this whole discussion, and depressive symptom improvement (n=22 + randomized + double-blind!), all using their noninvasive ultrasound device.

How is this possible? How do these successful results square with the skull problems from earlier?

Clearly, something is getting through the skull, and it seems to be having some clinically significant effect. Because of this, it could very well be possible that the relative broadness of Nudge’s and SPIRE’s (and others like them) stimulation is, in fact, perfectly fine, and being incredibly precise is simply not worth the effort. This all said, it is hard to give Forest a fair trial here, since they are basically the only ones going the invasive route for ultrasound, and their clinical trials (which use noninvasive devices) have just started circa early 2025. Maybe their results will be spectacular, and I’d recommend watching Sumner’s (the prior Forest CEO) appearance on Ashlee Vance’s podcast to learn more about early results there.

But really, this debate between invasive and non-invasive really belongs in the previous section, because the point I am trying to make here is a bit more broad than these two companies. What I’m really gesturing at is that being really good at [X popular neurotech metric] doesn’t alone equal something better! This is as true for precision as it is for everything else.

Staying on the example of precision, consider the absolute dumbest possible way you could approach brain stimulation: simply wash the entire brain with electricity and hope for the best.

This is, more or less, what electroconvulsive therapy (ECT) does. Electrodes are placed on your scalp, a generalized seizure is induced, and you repeat this a few times a week. You are, in the most literal sense, overwhelming the entire brain with synchronized electrical activity. And yet despite the insane lack of specificity, ECT remains the single most effective treatment we have for severe, treatment-resistant depression. Response rates hover around 50-70% in patients for whom nothing else has worked, with some rather insane outcomes, one review paper stating: “For the primary outcome of all-cause mortality, ECT was associated with a 30% reduction in overall mortality.” For some presentations, like depression with psychotic features, catatonia, or acute suicidality, it is essentially first-line.

This should be deeply humbling for anyone looking into the neuromodulation space. There are companies raising hundreds of millions of dollars to hit specific brain targets with millimeter, even micron precision, and meanwhile, the most effective neurostimulation-for-depression approach we’ve ever discovered involves no targeting whatsoever. Now, of course, there are genuine downsides to the ECT approach (cognitive side effects, the need for anesthesia, the inconvenience of repeated hospital visits, obviously doesn’t work for every neuropsychiatric disorder) that make it worth pursuing alternatives! But it does suggest that the relationship between targeting precision and clinical outcome is much more complex than you’d otherwise assume.

Consider the opposite failure mode. Early deep brain stimulation—the most spatiotemporally precise neurostimulation method currently available—trials for depression are instructive here. Researchers identified what they believed was “the depression circuit,” implanted electrodes in that exact area, delivered stimulation, and then watched as several major trials burned tens of millions of dollars on null results. Most infamously, the BROADEN trial, targeting the subcallosal cingulate, and the RECLAIM trial, targeting the ventral capsule/ventral striatum, both of which failed their primary endpoints.

Yet, DBS is FDA-approved for Parkinson’s treatment and is frequently used to treat OCD. Each indication is a world unto itself in how amenable it is ‘precision’ being a useful metric.

But again, this point extends beyond precision.

As a second example, consider the butcher number, a metric first coined by the Caltech neuroscientist Markus Meister, which captures the ratio of the number of neurons destroyed for each neuron recorded. Now, you’d ideally like to reduce the butcher number, because killing neurons is (probably) bad. And one way you could reliably reduce the butcher number is by simply making your electrodes thinner and more flexible. This is, more or less, at least part of Neuralink’s thesis: their polymer threads are 5 to 50 microns wide and only 4 to 6 microns thick (dramatically smaller than the Utah array’s 400-micron-diameter electrodes!) and thus almost certainly has a low butcher number.

Here’s the Neuralink implant:

And here’s the Utah array:

But does having a lower butcher number actually translate to better clinical outcomes? As far as I can tell, nobody knows! It’s largely unstudied! It’s conceivable that yes, lowering this number is useful, but surely there is a point where the priority of the problem dramatically drops compared to the litany of other small terrors that plague most neurotech startups.

The point here is not that the butcher’s number is useless. The point also isn’t that precision is useless. The point is that the relationship between any given engineering metric and clinical success (in your indication) is rarely as straightforward as anyone hopes, and it’s worth considering whether that relationship has actually been established before believing that success on the metric is at all useful.

5. Could this be done without touching the central nervous system?

Finally: something that repeated across the neurotech folks I talked to was that people consistently underestimate how extraordinarily adaptable the peripheral nervous system is. For example, a company that claims to, say, automatically interpret commands to a digital system via EEG should probably make absolutely certain that attaching an electromyography device to a person’s forearm (and training them to use it) wouldn’t wind up accomplishing the exact same thing.

In fact, there was a company that did exactly this. Specifically, CTRL-labs, a New York City-based startup. They came up over and over again in my conversations as a prime example of someone solving something very useful, in a way that completely avoided the horrifically challenging parts of touching the brain. Their device was a simple wristband that reads neuromuscular signals from the wrist (via electromyography, or EMG) to control external devices. Here’s a great video of it in action.

Now, if CTRL-labs was so great, what happened to their technology? They were acquired by Meta in 2019, joining Facebook Reality Labs. And if you look at the ex-CEO’s Twitter (who is now a VP at Meta), you can see that he recently retweeted a September 2025 podcast with Mark Zuckerberg, in which Mark says that their next generation of glasses will include an EMG band capable of allowing you to type, hands free, purely by moving your facial muscles.

Not too far of a stretch to imagine that this is based on CTRL-labs work! And, by the time I finally finished this essay, the device now has a dedicated Meta page!

What about something that exists today?

Another startup that multiple people were exuberant over was one called Augmental. Their device is something called ‘Mouthpad^’, and a blurb from the site best describes it:

The MouthPad^ is smart mouthwear that allows you to control your phone, computer, and tablet hands-free. Perched on the roof of your mouth, the device converts subtle head and tongue gestures into seamless cursor control and clicks. It’s virtually invisible to the world — but always available to you.

And here’s a wild video of a 19-year old quadriplegic using this device to interact with a computer and even code.

Isn’t this insane? I remember being shocked by the Neuralink demo videos showing paralyzed patients controlling cursors on screens. But this is someone doing essentially the same thing! All by exploiting both the tongue, which happens to have an extremely high density of nerve endings and remarkably fine motor control, and our brain, which can display remarkable adaptivity to novel input/output channels.

Now, fairly enough, a device like Augmental cannot do a lot of things. For someone with complete locked-in syndrome, there really may be no alternative to inserting a wire into the brain. And in the limit case of applications that genuinely require reading (or modifying!) the content of thought, the periphery again won’t cut it. But for a surprising range of use cases, the peripheral route seems to offer a dramatically better risk-reward tradeoff, and it feels consistently under-appreciated when people are mentally pricing how revolutionary a new neurotech startup is.

Conclusion

This piece has been in production for the last five months and, as such, lots of discarded bits of it can be found on the cutting room floor. There are lots of other things, not mentioned in this essay, that I think are also worth really pondering, but I couldn’t come up with a big, universal statement about what the takeaway is, or the point is pretty specific to a small subset of devices. I’ve attached three such things in the footnotes.1

Before ending, I’d like to repeat the sentiment I mentioned at the start: this field is complicated. A lot of the readers of this blog come from the more cell-biology or drug-discovery side of the life-sciences field, and may naturally assume that they can safely use that mental framework to grasp the neurotech field. I once shared this optimism, but I no longer do. After finishing this essay, I now believe that the relevant constraints in this domain come from such an overwhelming number of directions that it bears little resemblance to most other questions in biology, and more-so resembles the assessment of a small nation’s chances of surviving a war. The personality required to perform such a feat matches up with the archetype of individual I’ve found to work in this field, all of whom display a startling degree of scientific omniscience that, in any other field, would be considered extraordinary, but here is equivalent to competence. It would be impossible to recreate these people’s minds in anything that isn’t a seven-hundred-page text written in ten-point font, but I hope this essay serves as a rough first approximation.

Extraordinarily grateful to Milan Cvitkovic, Sumner Norman, Ben Woodington, and Adam Marblestone for all the helpful conversations, comments, and critiques on drafts of this essay.

Today, Canada officially lost its measles elimination status. Measles was previously declared eliminated in Canada in 1998, but countries lose that status after 12 months of continuous transmission.

Here are some articles about the the fact that we have lost our measles elimination status: CBC, BBC, New York Times, Toronto Life. You can see some chatter on Reddit about it if you’re interested here.

None of the above texts seemed to me to be focused on the actual thing that caused Canada to lose its measles elimination status, which is the rampant spread of measles among old-order religious communities, particularly the Mennonites. (Mennonites are basically, like, Amish-lite. Amish people can marry into Mennonite communities if they want a more laid-back lifestyle, but the reverse is not allowed. Similarly, old-order Mennonites can marry into less traditionally-minded Mennonite communities, but the reverse is not allowed.)

The Reddit comments that made this point are generally not highly upvoted, and this was certainly not a central point in any of the articles. It is a periphery point in all of the articles above at best. Toronto Life is particularly egregious, framing it like so:

“mis- and disinformation were factors in the outcome, which are partly due to pockets around the country with low vaccination rates.”

This is, ironically, misinformation: true information framed in such a way to precisely give you the incorrect view of things.

In this post I will make two arguments: first, yes, it is the Mennonites that began (and are the biggest victims of) the biggest measles outbreak of the current century, and second, thinking of them as resistant to vaccination is actively harmful to the work of eliminating measles from Canada once again.

I’ve been following the measles outbreak closely for basically its entire duration, because I have a subscription to my local newspaper, the Waterloo Record. The writers there do frequent updates on the outbreak, often with higher quality and more detail than you get in the national papers. This is because Waterloo Region has a significant Mennonite population, so shit sometimes got real scuffed.

Like, over last spring, there were fairly regular advisories about local stores we shouldn’t go into or quarantine if we did because someone with measles went in. One of them was the pharmacy across the street from the university campus, so that was fun.

The Mennonite Outbreak

Here is what the outbreak looks like, Canada-wide: Health Canada

Full offense to Health Canada: this is a terrible graphic, because if you don’t look at it carefully you will think that the provinces in dark blue have approximately the same number of cases, and this is very false. Sasksatchewan has barely over a hundred, Alberta has almost 2000, and Ontario has almost 2400 cases.

What’s the deal with Ontario, and Alberta? Some of it comes down to the numbers game; those are two of our most populous provinces. But Quebec has twice the population of Alberta, and it’s trucking on with only 36 cases in the entire province.

The answer is that it’s the Mennonites, who are overwhelmingly settled in those two provinces.1

I’ll be focusing on the outbreak in Ontario, because that’s the part of the story I’m more familiar with. If you dig into older news pieces, the Mennonite connection is corroborated by government officials:

Previously, Moore [the Chief Medical Officer for Ontario] shared that this outbreak in Ontario was traced back to a Mennonite wedding in New Brunswick, and is spreading primarily in Mennonite and Amish communities where vaccination rates lag. The vast majority of those cases are in southwestern Ontario.

Mennonites have a social structure where, once the community reaches a certain number of families, they undergo mitosis, and half the families split off to form a new community far away. Based on reddit scuttlebutt, it seems like there has recently been a daughter community that moved from southern Ontario to New Brunswick, which makes it doubly unsurprising that there were many southern Ontario attendees to the original superspreader event.

Additionally, Moore, remarked in a memo he sent out to local health bodies:

Over 90% of cases in Ontario linked to this outbreak are among unimmunized individuals. Cases could spread in any unvaccinated community or population but are disproportionately affecting some Mennonite, Amish, and other Anabaptist communities due to a combination of under-immunization and exposure to measles in certain areas.

And Global News reports:

In an April interview with The Canadian Press [Moore] reasserted that the “vast majority” of Ontario’s cases are among people in [Mennonite, Amish, and other Anabaptist] communities.

Some smaller publications have found connections in their own investigation. The London [Ontario] Free Press in March 2025 (the beginning of the outbreak) linked the outbreak in West Texas to their Mennonite population, and identified that several measles exposure sites in counties that have been heavily afflicted by measles are Mennonite in nature:

A list of measles exposure sites in Grand Erie includes a church and several private Christian schools in western Norfolk County catering to Old Colony Mennonites, and Moore’s letter confirmed the link.

A recent Washington Post article also corroborates the link, but buries it under several paragraphs of preamble about general vaccine skepticism.

Many large measles outbreaks in Canada have occurred in insular Mennonite communities in rural Alberta and Ontario, where some are skeptical of vaccines.

Outbreaks have also been reported in Mennonite communities in Mexico and West Texas.

Mennonite Geography

Public Health Ontario has infection numbers for you, broken down by geographic area (”public health units”). Here’s what that looks like when I plot them on a graph. Notice that there are five units that are responsible for basically all of the cases, and you will have heard of none of them because they include zero major population centres.

The most populous health units, such as Toronto, Ottawa, Halton, Hamilton, Peel, and York, all have three cases or fewer for the entire year, and a corresponding case rate of close to zero.

I admit that I do not have the temerity required to separate out Mennonites from like, generic rural dwellers, but something wonky is going on here! The measles outbreaks are all in sparsely populated regions while the big cities (with their big suburbs, presumably where all the anti-vaxxers would be) carry on basically unscathed.

To better visualize this, I am going to combine a bunch of charts together jankily: the geographic distribution of measles (blue), population density (red, adapted from Wikipedia, and, in lime green, the settlements of Amish and Mennonite communities I found online.2

I tried to match the map outlines in procreate by hand, which means it was done imperfectly. Okay, sorry, you will need to stare at it for a bit. The key takeaway is that the blue areas (which represents measles cases) almost perfectly avoids the most populated areas (red), and are full of green dots (where the Mennonite and Amish settlements are).

Let’s look at this another way. Here’s a Public Health Ontario COVID report from April 2022 (i.e. after vaccinations have been available for a while). Pages 8-10 include comparable charts on cases per 100,000 people broken down by Public Health Unit. It’s relatively stable between PHUs, and larger in city centres compared to rural settlements. This makes sense, because urban settlements are by definition denser, which means it’s easier for viruses to spread.

Here are the outbreaks plotted against each other, if you’re curious.3 Notice that the same five public health units no one has heard of are outliers again, which is what you would expect.

Also note the different degrees of variance in cases per 100,000 people in a health unit. COVID cases per 100k ranged from about 4,000 to 11,600 across health units, which is roughly a 3x difference. The Measles cases were actually incredibly discontinuous across units: many units had literally zero cases, some had under 30 cases per 100,000 people, then there’s a huge gap, and then there’s five regions that had over 100 cases per 100,000 people.

For the statistically inclined, the coefficient of variation (standard deviation divided by mean, expressed as a percentage) was about 25% for COVID and 193% for measles, which is almost eight times higher.4 (If these numbers mean nothing to you, don’t worry about it. The point is just that COVID spread relatively evenly and measles did not.)

Lastly, here is some incidental info from some July 2025 coverage on the outbreak in St. Thomas (a smaller city, in that vertical belt of green dots across central southern Ontario, in the PHU Southwestern Public Health which is the one with the most cases):

Five months later, around 150 to 200 of our [Mennonite] clients have had measles, and most of our Low German–speaking clients have at least had symptoms.

As of October 28, there has been 771 cases of measles in the Southwestern PHU. If 150-200 of them were Low German-speaking mennonites as of July, and most of their clients had symptoms at that time, this indicates that the Mennonites would have made up for a substantial amount of all cases in that PHU.

I rest my case! Which is not to say that it is a perfect one but here is where I put it down because I am not going to put more effort into it. I encourage others to pick it up and put more work into it if they are so inclined.

Mennonites Are Susceptible To Facts and Logic, When Presented In Low German

The general sentiment both in the reddit comments and in most of the news coverage seems to be something like “oh, they’re weird religious people, and therefore immune to logic about vaccines”, and also something something religious tolerance meaning that we can’t criticize their choices at all.5

But in reality, Mennonite parents love their children and do not want them to die of measles, and they do not want to contract measles themselves. Having looked into it, it seems to me like the largest barrier for them getting medical care and vaccination is that they are not fluent in English, they speak Low German.

In Ontario, three quarters(!!!!!!) of the 700 Mennonite community clients helped by a Low German-speaking personal support worker have agreed to be vaccinated.

In Alberta (the other large Mennonite population centre, and not coincidentally the other large site of the outbreak), there has been a 25% increase in demand for medical care in Low-German, and service has expanded from five to seven days a week.

And, like, yes, to be clear, there are loads of Mennonites who are actually anti-vaccine. I am not disputing the obvious fact that, in religious communities, many people are against vaccinations. Further, 75% still falls very short of the 92-94% vaccination rate needed for herd immunity. But a 75% vaccination rate is much, much higher than I’d have hoped for?

Here is an example of the miscommunication that can happen when one is not fluent in English:

The following morning, the [Mennonite] mother called me: her child [who recently developed a measles rash] was coughing so violently she was vomiting. I told her to go to the hospital. Later, she called me again, upset. She said that when she got to the ER, they’d told her to go home.

I couldn’t help but think something was off. The hospital doesn’t turn people away, I told her, but she insisted that they had. So I called them directly to figure out what had happened. It turned out there had been a miscommunication. Hospital staff had told her not to come in, using a “stop” hand gesture to communicate, and she had become so flustered that she failed to catch the second part of the message: that she should wait in the car while they prepared a negative-pressure room.

If your measles outbreak comes from this sort of community, the solution isn’t to fearmonger about anti-vaxxers. It is to train up and hire health care workers who can speak Low German. And to be clear, I think the PHUs that are affected are doing this, or at least the Ontario ones are, because our public health bodies are generally not disconnected from reality.

And that’s what I found most frustrating about the media coverage. It obscures something that’s genuinely very hopeful and turns it into another random culture war shitfest. But actually, it turns out that when you remove the actual language and access barriers, people make reasonable healthcare decisions for their families at pretty high rates!

So yes, Canada lost its measles elimination status today. But we can get it back in a year and a bit, if we’re serious. And if we’re serious about eliminating measles again, we need to focus more on investing in healthcare workers who speak the language of and can build relationships in communities, and less on implying that certain populations are fundamentally unreachable.

“Celebrity biohacker Josiah Zayner is under investigation for practicing medicine without a license” — “A grueling and grotesque biohacking experiment: Josiah Zayner experiments with fecal transplants” — “Josiah Zayner: the person who hacked his own DNA”

From the headlines it might be hard to believe but when I started down the career path towards becoming a scientist I didn’t even know what a Biohacker was. I know, Josiah Zayner doing boring academic science and not injected himself with shit, that’s crazy, amirite? I’m definitely not the first Biohacker. Not the first to experiment on themselves or the first to do science in my apartment. I was just lucky enough to have been around since the early days of Biohacking. I have seen it grow from when it was just a bunch of people bitching on the diybio mailing list. I learned through all this that Biohacking is more than just doing science outside traditional environments. Biohackers are building something. Creating resources so _anyone_ can participate, even crazy people like Rich Lee. Biohacking has created a participatory feedback loop that will make sure one day their numbers are far greater than traditional scientists. That’s what makes it so revolutionary. That’s what makes Biohacking a modern invention.

It’s not clear when the first “Biohack” happened. The first time say, someone genetically modified E. coli or some basic microorganism outside a lab. Maybe it was Rob Carlson circa the mid-late 2000s? There was the infamous arrest of artist Steve Kurtz in 2004 for growing bacteria and yeast but that seemed to lack the intent of Biohacking. It’s even more complicated because the definition of Biohacker is pretty fluid. Originally, the term and ethos were looked down upon so it wasn’t really used. If you ask Hank Greely, Biohacker is still a pejorative. In the early days, the term “DIY Biologist” was more en vogue because it sounded less scary. Me though, I hate that term. It implies that a scientist outside a traditional lab is somehow inferior. They can’t be a Biologist. They are a _DIY_ Biologist.

y tho?

When those inside and outside professional labs have access to the same equipment and resources. They can and do literally do the same work. So why does one group need a preface? To me the “DIY Biologist” name supports the same elitist academic science attitude that we/I are trying to destroy. What’s the point in just rebuilding a less funded academia? The DIY Biologists I know are trying to fit into the system while Biohackers are trying to break free.

I was in the system for many years I’ve seen it all and it’s ugly under the surface of what the public sees. If you don’t fit into the stereotype and fall into line your chances of academic success are nonexistent. There is an extreme lack of diversity both socio-culturally and of thought. It probably wasn’t until around 2010 that I began to see myself as a Biohacker. As someone who didn’t fit in the mold. I was working on my PhD at University of Chicago at the time so as you can imagine I was an arrogant brash asshole. I was trying to publish my first paper. And after staring into the abyss one too many times I realized Academic science wasn’t what I imagined. I wanted to work on the crazy stuff. The experiments everyone else was hesitant to do even though they were always on our minds.

What if?

And so I started doing experiments in my apartment in my spare time. I would fall asleep at the workbench in my bedroom at 4AM. I was enthralled. That freedom of expression I had through Biohacking drove me away from academia.

The sentiment of doing independent science seemed to be in the beer. It was weird because from like 2009-2010 the idea of doing science outside academia all of the sudden became kind of popular. The mailing list was blowing up with names that today are recognizable in the Biotech world. Two or three Biohackerspaces popped up across the US. It was a foundational time. People didn’t really know what to do. Should we make GFP fluorescent yogurt? Sequence anal bacteria? Try and do something groundbreaking? Whatever it was the limited availability of resources made it so Biohackerspaces were where it was at.

It’s funny that they are called Biohackerspaces considering most who venture there dislike the term Biohacker. A Biohackerspace is a lab that charges people for access, sometimes around $100-$200 a month. They have some general equipment and resources available but most supplies are self furnished. Ima be honest. Most spaces I have visited are generally more bougie than useful and are more like social clubs for tech people interested in science. I’m not against capitalism. I run a business that sells genetic engineering supplies. I get it. They need money to keep the doors open. It’s just sad that the practice of a free open lab doesn’t exist even to this day. Would technology even be where it is today without free public access to computers? I dream of the day that’s possible with science. Because really, in order to exist without gatekeepers you need to remove the walls. (*sidenote La Jolla library did actually have a public biolab for a time - according to their website it is currently shut-down).

For science to truly be accessible anyone should be able to do any experiment at anytime anywhere. That’s Biohacking. Then you’re only limited by your imagination. Well, reality also. And um’ laws, if the Feds are watching. Spoiler, they are.

I grew up during the 90s computer hacker era and was an active participant. I was part of the hacking group Legions of the Underground. 31337 I know. Everything was founded on the principles of radical access and autonomy. People don’t own knowledge and we shouldn’t be trying to control who can access technology. Innovation exploded. Science, the one field that should be forcible cramming knowledge into people’s gourds is instead one of the biggest gatekeepers of knowledge in the world. Expensive paywalls and PhDs required. How is everyone ok with that? I’m not ok with that.

Around 2014 I started my company, The ODIN. I left my job as a NASA scientist in 2016 to run it full-time (←100% this sentence is superfluous and just to tell you I worked at NASA). The goal was to have a centralized website where Biohackers could get all their supplies for the lowest price on the market. Ease of access would mean more people would start doing genetic engineering in their kitchen. Experimenting on whatever they wanted. When technologies are made accessible it leads to creation. Even better, it leads to people creating beautiful things. The printing press, the automobile, the computer, NFTs. Wait, sorry, not NFTs. Science needs a little more of “Fuck it. I’ma just do that because it’s beautiful.” That’s Biohacking.

There weren’t any genetic engineering technological breakthroughs between 2000 when there was no Biohacking and 2014 when it was common place for budding Biohackers to have genetically modified E. coli, a simple process that is a staple of modern genetic engineering labs. Putting DNA into an organism was done exactly the same way—IS STILL done exactly the same way. Once there was this nucleus of knowledge and skill things escalated quickly.

I injected myself with a CRISPR plasmid in 2017. Ostensibly, to genetically modify my muscles to make them grow bigger. I actually was trying to raise awareness about the possibility of Biohacking. It went viral. Despite being non-viral gene delivery.

I’m so sorry for that bad science joke.

Not long after, I was sitting on my couch and scrolling through Facebook. A post came into my feed, some randos we’re going to do a live injection of a gene therapy to try and cure HIV. Whatttt the fuuuckkk. I thought it would be years before someone followed in my irresponsible footsteps. Things were happening way way faster than I anticipated. The fact that, my predictions were so far off scared me a little. It meant that Biohacking was out of everyone’s control. The press were thirsty for this shit and that encouraged a whole slew of injections afterward that probably made Biohacking look more crazy than competent. That’s Biohacking.

Biohackers began to push boundaries. Medical, surgical, genetic, you name it. Normally, in a traditional academic environment human experimentation would require forms and committees, meetings and approvals. Biohackers found their niche there. You don’t need approval to test on yourself. And you don’t need an ethics committee if you are operating outside an organization. I actually love where Biohacking is going. At first I was like, “Oh fuck, what have I done.” But that time period created the Biohacking aesthetic most people pursue nowadays. One that is imperfect and flawed for sure but one that favors brashness and style above all else. Let’s be honest, if you are going to destroy the system you might as well do it in a brash and stylish way. That’s Biohacking.

To me, it all changed in 2021. Before any covid vaccines had been released to the public David Ishee, Dariia Dantseva and myself successfully made and tested a covid vaccine. We saw antibody responses which is more than the majority of people vaccinated in the world can say. Still, it was largely ignored by the mainstream press and I was banned from YouTube for life for showing people step-by-step how to do it. It was shocking. We were living in the future but the world couldn’t keep up. This is where I believe Biohackers can exist. At the intersection of the future and reality. We can create breakthroughs that are only being held back because of bureaucracy or fear. I’ve been banned, ridiculed and harassed by the government. Others will be also. That’s Biohacking.

Biohacking doesn’t seem to be slowing down and that’s hard to comprehend when it went from virtually nonexistent to even beginners being able to engineer human cells in their kitchen in under 10 years. Individuals now have the Biotech power of governments and large pharma companies. And Biohacking is different. It’s new. Despite being in the pitch-deck of every modern biotech CEO this is not like the computer revolution. No one had to create new technology for a DIY covid vaccine to exist. It is only the monopolization of knowledge and information by traditional science that has prevented biohacking from flourishing earlier and growing faster.

That should scare you.

The state of science is a tragedy. In interviews people always ask me how Biohackers will publish papers or participate in other ceremonial activities. My response is, “They won’t”. Biohacking isn’t meant to be Science 2.0. It’s evolved to become a completely new species distantly separated from its phylogenic origin. The way science is done hasn’t changed much in the past 500 years and its shriveled corpse of outdated principles has been due a funeral for a long time. The thing that is going to save us from science?

That’s Biohacking.

There is an extraordinarily important implication of Cosmological Natural Selection that I feel I haven’t made sufficiently clear in my writing so far. I MENTION it frequently, but always in passing; I’ve never stopped and shone a spotlight directly on it. But it answers three extremely important and fundamental questions, and so it needs its own standalone post; something I can simply direct people to, when they ask

• “Why do you think this theory is true?”

• “What does it imply?”

• “What does it predict?”

The failure to highlight this vital insight is partly because of the unusual circumstances of its birth: it only occurred to me in those frantic days, in June and early July of 2022, during which I was thinking through Cosmological Natural Selection from first principles. I had a hard deadline: I needed to generate, refine, and write up my predictions, then post them online, and email them to my subscribers, before the James Webb Space Telescope unveiled its first data on July 12th, 2022. That breakthrough insight became the basis of my (successful) predictions, but I didn’t have the time to explore it at length, and draw attention to it, in the way it deserved. It just became a hastily-written paragraph in a predictions post that was emailed to my 146 subscribers of the time.

This Substack now has over 10,000 subscribers, so 98.5% of them didn’t read that original breakthrough post. Yet it contains the Big Original Idea that led to Three-Stage Cosmological Natural Selection; the key insight from which all else follows. My successful predictions about the early universe emerge directly from it (and thus my grants and funding flow from it); the Blowtorch Theory of structure formation is built directly on top of it; the big piece on spiral galaxy formation that I’m currently working on starts with it; and, above all else, it’s the key piece of evidence that this theory may be true.

So, I’m going to, finally, separate it out here as a short(ish) post.

BACKGROUND: THE EARLY HISTORY OF COSMOLOGICAL NATURAL SELECTION

(Yes, I need to recap again for new subscribers – numbers have gone up a lot recently. And yes, if you know the history of Cosmological Natural Selection already, then click here to skip to the “implications” section.)

The brilliant, quirky American physicist John Wheeler (1911-2008) was co-author of the definitive textbook on gravity, and accidental co-inventor (along with a heckler at one of his talks) of the term “black hole”.

In the mid-1970s, Wheeler became fascinated by the fact that cosmology now had two mysteries involving singularities – points where the density of matter went off-the-scale, and where both of our best mathematical theories (general relativity and quantum mechanics) ceased to function, and started spitting out infinities.

One was a black hole, where mass/energy collapsed rapidly to a point, and mysteriously vanished from the bubble of space-time that is our universe.

The other was the Big Bang, where mass/energy emerged mysteriously from a point, and rapidly expanded to form the bubble of space-time that is our universe.

Wheeler pointed out that they looked awfully like the two sides of one process or event, joined at the singularity. One side simply mirrored the other, as though contraction had flipped or “bounced” at the singularity to become expansion.

(A BRIEF ASIDE: This, by the way, is not unprecedented in physics. For example, and to massively oversimplify: during the collapse of a star, the net effect of all the relevant forces flips, in a way that abruptly reverses the inward collapse of the star into the outward expansion of a supernova explosion. The forces themselves don’t flip. It’s just that their various values, and the way they interact at extreme densities, happen to be tuned to values that can generate this extraordinarily powerful reversal – which allows the periodic table of elements, generated by fusion deep in the heart of the star, to be distributed back up out of the star’s enormous gravity well and into the interstellar medium, where it can help build more stars, planets, and ultimately life. Yes, if you’re looking at our universe through an evolutionary lens, that looks awfully like the kind of fine tuning – to otherwise unlikely values, with functional consequences for reproductive success – that you would end up with after a long evolutionary process. But we are getting ahead of ourselves… END OF BRIEF ASIDE.)

As an elegant way to solve the two problems, Wheeler proposed that mass/energy in a parent universe collapsed to form a black hole (thus leaving the bubble of space-time that was the parent universe). It then “bounced” at the singularity, and expanded in a Big Bang, to form a new universe – budding off a new bubble of space-time, outside of, and separate from, its parent universe.

That is, parent universes reproduced through black holes, which “bounced” to form Big Bangs. Each Big Bang was therefore the birth of a new child universe, which grew up to produce more black holes; more child universes.

In the 1990s, another brilliant American theoretical physicist, Lee Smolin, saw a marvellous implication of Wheeler’s idea that had been completely missed by everybody (because theoretical physicists don’t usually read much evolutionary biology; Smolin had been reading some, for fun). Smolin realised that if there was slight variation in the basic parameters of matter in the child universe (rather than the large and random variation assumed by Wheeler), you would automatically get Darwinian evolution of universes. That’s because slight variation allows for inheritance; and some of those slight variations would lead to more black hole production; more reproductive success; more offspring. Those successful variations would therefore be inherited more widely than less successful variations. The successful variations would vary again (slightly) in the next generation, with some variants being even more reproductively successful (and some less); and off we go – Darwinian evolution of universes, with the evolutionary ratchet optimizing for reproductive success. Over time, the majority of universes will come to be fine-tuned for very high levels of black hole production (compared to the evolutionary starting point, where the numbers could be as low as one or two). This seems plausible: our universe has already produced over forty quintillion black holes, just from stellar collapse. (That’s 4×10¹⁹; a four and 19 zeros! For how they worked that out, see Sicilia, Lapi, et al, 2022.) Smolin’s simple and straightforward evolutionary theory of universes became known as Cosmological Natural Selection (CNS), laid out in this paper, and this book.

So, Smolin had uncovered a powerful evolutionary implication, hidden inside Wheeler’s simple model of the reproduction of universes.

But there is a further, huge, implication hidden inside Smolin’s new and improved theory. Lee Smolin didn’t see it at the time because he was a theoretical physicist, not an evolutionary biologist, and so not saturated in the logic of evolution (from which the implication emerges). He was also handicapped by the fact that astronomy had only observed half a dozen or so supermassive black holes at the time he wrote his original paper and book. It was still therefore blithely assumed, by astronomers and cosmologists and theoretical physicists alike, that supermassive black holes must simply be a bunch of (much smaller) stellar-collapse black holes merged together. It just hadn’t occurred to anybody that black holes that large might have a separate, far simpler, formation mechanism.

But that’s exactly what an evolutionary theory of universe implies. This is an extremely important point, but people usually skip over it when I make it, because it’s usually embedded in a more complex post, alongside a lot of other new information; and so it just whizzes past, as one point among many. This time, I’m going to walk through the logic in excruciating detail. And when I’m done, don’t just click through to the next thing in your scroll. Take a moment, and sit with it. Think it through. Test it in your mind.

And if, at the end of all that, you think it’s true, internalise it. Bring it to the conversation around these issues. Because if it’s true, the boundaries of various scientific fields are currently in the wrong place to explore this new knowledge. A lot of things will need to change. And you can help change them.

COSMOLOGICAL NATURAL SELECTION’S KEY IMPLICATION, AND THUS PREDICTION

The original reproductive mechanism for universes can’t have been stellar-collapse black holes, because stars are complex, orderly structures made of complex, orderly matter; and all the complex, orderly systems we know of are the result of evolution. (Just as the evolutionary history of biological life can’t have started with complex, orderly eukaryotic cells, with their sophisticated, interacting organelles; just as the evolutionary history of communications devices can’t start with the iPhone 17 Pro Max. ) Looked at through this lens, the periodic table itself, with its dozens and dozens of ever more complex elements, clearly must be the result of an evolutionary process. (It is highly suggestive that the elements get more ragged and unstable towards the upper, more complex, and thus more recently evolved, end.) The original ur-matter, in the original ur-universes, however, must have been as simple as it gets. Very pure, very simple matter, with no structure, and building no structure.

Small, efficient, stellar collapse black holes, therefore, generated at the end of a star’s rich and complex lifecycle, are – must be – a later evolutionary breakthrough.

So, think this through from first principles. If our universe is the result of a Darwinian evolutionary process, then every single one of the earlier universes in our direct evolutionary line must have reproduced successfully, by definition.

But the only thing that is unavoidably, unsimplifiably required to make a black hole (to reproduce) is for mass/energy to collapse.

The original reproductive mechanism, therefore, must have been the direct collapse of extremely primitive and unstructured matter to form small numbers of large, crude, and therefore supermassive black holes. Far more massive, and therefore far less numerous, than the forty quintillion stellar-collapse black holes we already have so far in our specific universe. (That’s unavoidable, because the more massive a black hole is, the larger the percentage of its universe it must take up, and therefore the less of them there can be.) Evolution will thus blindly drive towards the ever-more efficient production of ever-smaller black holes - with each still able to produce a full-sized universe, thanks to the fact that the positive mass energy and negative gravitational energy in a universe net out to zero, and so full-sized universes can be built for free.

But we have discovered, in the three decades since Lee Smolin published the first, wonderful, tentative (and tragically overlooked) version of Cosmological Natural Selection, that our specific universe in fact contains roughly a trillion supermassive black holes – one at the centre of every galaxy. (It sounds like a lot, but divide that trillion into forty quintillion: there’s just one supermassive black hole for every forty million stellar collapse black holes to date.) However, though relatively few in number, those supermassive black holes are often millions or even billions of times more massive than the average stellar-collapse black hole. They must, therefore, have formed by that original reproductive mechanism; direct collapse of extremely large amounts of primitive and unstructured matter.

Why? Because evolution is frugal; she doesn’t bother to come up with a much more complicated way of doing something for which she has already come up with a perfectly good mechanism, unless that innovation leads to far greater reproductive success. Yes, evolution – blindly exploring the possibility space for the basic parameters of matter, through generation after generation of universe – eventually came up with more complex forms of matter, and thus stars, and stellar-collapse black holes, and conserved that breakthrough, because they were a huge improvement on a small number of immense direct-collapse supermassive black holes. With stars, you could get millions or even billions of stellar-collapse black holes / offspring / child universes, out of the same mass that once (in earlier generations) produced just a single supermassive black hole, and thus a single child universe.

Yet supermassive black holes clearly still have a function in our universe, given that there’s one at the center of every galaxy; given that they have been conserved by evolution. (And I explore that function in my first Blowtorch Theory post.) They are clearly still required, to help build out the larger, later, more complex structures (just as, say, complex contemporary multicellular organisms still require the presence of the primitive, ancient mitochondria they long ago engulfed, in order to build and assemble themselves). If supermassive black holes have been conserved by evolution, then their formation mechanism, direct collapse, must also have been conserved. (Similarly, mitochondria still reproduce by splitting themselves in two, as they always did – in a cycle that is not synchronised with the reproduction of their host cell.)

Meanwhile, how did mainstream cosmology think supermassive black holes formed, before the James Webb Space Telescope data? Well, they had half a dozen theories, which means they had no theory. This is the understandable and inevitable consequence of having no meta-theory; no helpful and constraining framework (such as evolution) capable of extracting meaning from data, and imposing meaning (or, if you prefer, discipline) on hypotheses. Such a meta-theory gives you something against which the real-world likelihood of hypotheses could be tested. In the absence of such a meta-theory, the only test cosmologists could apply to their ideas was (and is), are they mathematically possible (that is, without mathematical contradictions). But there are countless mathematically possible formation mechanisms, especially when dark matter is added as an optional, and tuneable, ingredient.

(ANOTHER BRIEF ASIDE: Quantum mechanics, incidentally, is in a similar pickle, which is why theoretical physicists, over the past century, have come up with SO MANY utterly wild, purely mathematical theories, attempting to advance the field, without ever actually getting anywhere. As Adam Forrest Kay puts it, in his superb book Escape from Shadow Physics,

“…without pictures of hidden reality, the only guidance is mathematical rigor. This makes the search space explode, because then all steps that do not lead immediately to contradiction are on equal footing.”

– Adam Forrest Kay, Escape From Shadow Physics: Quantum Theory, Quantum Reality and the Next Scientific Revolution

But that’s a whole other story… END OF ANOTHER BRIEF ASIDE.)

And so, in the decades before the James Webb, mainstream cosmology kept coming up with exciting new hypotheses for supermassive black hole formation.

DARK MATTER MINI-HALOS GENERATE POPULATION III STARS!

The most widely accepted theory was that the seeds for supermassive black holes were formed by “Population III” stars, the annoying term for the very first stars, made of only hydrogen and helium, which were believed to be unusually massive and shortlived: after one burned out and collapsed to form a large black hole (of maybe a hundred solar masses), it would continue to swallow a huge amount of gas and grow fast. And maybe merge with other Population III black holes, if that was still necessary to get the mass right (i.e., to keep the maths from becoming contradictory). As nobody had ever seen a Population III star, or knew when they first formed (or, indeed, IF they formed), you had a lot of mathematical wriggle room with this one. Plus, oh yes, you needed a dark matter mini-halo to form them. More mathematically wiggly fun.

DENSE STAR CLUSTER MERGER MANIA!

Another popular idea was that, in dense clusters of stars, many massive stars would sink to the center, collide, and merge to build an extremely massive star that would quickly collapse to form a black hole with a mass of between a thousand and a hundred thousand solar masses. (That’s an IMBH [Intermediate Mass Black Hole], with a mass of 10³–10⁵ M☉, if you speak maths and acronym.) Several of those intermediate mass black holes would then merge, and also accrete more gas, to finally form a supermassive black hole.

PRIMORDIAL BLACK HOLES FORM IN THE FIRST SECOND!

Another option was primordial black holes: black holes which might form in the absurdly early universe, like, the first second after the Big Bang.

(A BRIEF ASIDE FOR THE MATHS BROS, SKIP IF YOU HATE MATHS: “The first second” is, in fact, surprisingly, a fairly precise term here. When radiation dominates – i.e. in the immediate aftermath of the Big Bang – the horizon mass scales ≈ 10⁵ M☉ × (t/1 s). Primordial black holes forming at t ≈ 1 s naturally sit at ~10⁵ M☉, which would make a nice, heavy seed for rapid growth into a supermassive black hole. (Earlier times make for much smaller primordial black holes; e.g., t ≈ 10⁻⁵ s → ~1 M☉.) END OF BRIEF ASIDE FOR THE MATHS BROS.)

How did these primordial black holes theoretically form? Through mechanisms for which we had no evidence, but which – given how little we know about the Big Bang itself, and how much you are therefore free to make up – you could easily make mathematically non-contradictory. Formed that early, they had a looooong time to drink gas and grow large. Also, they could merge, why not.

SELF-INTERACTING DARK MATTER GRAVOTHERMALLY COLLAPSES!

Want more? Self-interacting dark matter could undergo “gravothermal collapse” (don’t ask). I mean, it’s self-interacting dark matter! It can do anything it likes. Mathematically speaking, it could probably bake you cookies and bring them to you in bed. So, gravothermal collapse, why not.

MIGRATION-TRAP PHYSICS!

Or how about migration-trap physics – an idea borrowed from the theories used to explain how planets form in protoplanetary discs. The idea was that many small stellar-collapse black holes would get embedded in gas, which would drag them closer, pair them up, merge them, merge that merged pair with other merged pairs… and eventually you end up with a supermassive black hole.

You will note that many of these approaches to supermassive black hole formation put forward the same kind of passive, bottom-up, hierarchical, merger-driven process that was assumed for galaxy formation. This reveals the unspoken meta-theory under which cosmology was operating: it’s all arbitrary and random and means nothing. Any large complex, organised structure is staggered into, almost accidentally, through a random walk, by matter with arbitrary characteristics.

DARK MATTER ANNIHILATION HEATS THE FIRST PROTOSTARS!

Oh, but we are not finished. How about dark-star seeds! The big idea here was that dark-matter annihilation (don’t ask) would heat up the first protostars. As we know nothing at all about dark matter, you had plenty of room to make this mathematically non-contradictory.

Was there a clear winner from all this? No.

In fact, by the 2020s, you had comprehensive reviews like Inayoshi, Visbal & Haiman’s The Assembly of the First Massive Black Holes, in 2020, arguing for a multichannel model, where you had seeds forming from Population III stars, clusters smashing lots of stellar collapse black holes together, some primordial black holes, and maybe a few dark-matter-driven dark stars.

In other words, no theory. (Because, no meta-theory. No frame. Or rather, a nihilistic, meaning-denying meta-theory that couldn’t help you decide between options.)

But here’s something that makes the whole situation even more interesting, and revealing; in the years since Smolin’s original paper, Did the Universe Evolve?, some gutsy astronomers and astrophysicists – Martin Rees, Avi Loeb, Priya Natarajan, Volker Bromm, Marta Volonteri, and others – did work out that the formation of supermassive black holes by direct collapse was technically possible in our universe. They published lots of papers saying so, like this one, and this one, and this one…

However, those intrigued by direct collapse were outnumbered by those proposing all the formation mechanisms laid out above.

So direct collapse was lying there, on the table, as an option, since the 1990s. Why wasn’t it embraced?

Because cosmologists were stuck inside the old paradigm, the old meta-theory, the old frame-story: that our universe is a random one-off, with arbitrary properties, self-assembling slowly and passively and basically randomly – and certainly not fine-tuned by evolution for reproductive success through black hole production. Their meta-theory, their frame-story, therefore pushed them towards bad ideas, and away from good ones.

Three-stage cosmological natural selection, however (the key theory in the field I am starting to call Evolutionary Cosmology more generally) – provides a framework that allows you to simply and cleanly pick the killer formation mechanism – direct collapse – out of the police lineup (or identification parade, if you are in the UK or Ireland) of formation mechanism suspects.

That’s because if supermassive black holes – the earliest and most primitive structures, through which the earliest universes reproduced – are still to be found in our universe, then their production in our universe should precede that of any complex structures that evolved far later. The simple, original, reproductive mechanism shouldn’t depend on – and certainly couldn’t emerge from – later, more complex structures. That’s simply the unavoidable logic of an evolutionary history. And so, direct-collapse supermassive black holes, which must have been the earliest form of reproduction for the most primitive universes, with no structures required for reproduction, will come into existence in our specific universe before any complex structures. They will not be generated by more complex structures, and they are highly unlikely to be dependent on any more complex structures that only emerged later in the evolutionary history of universes. That’s just an unavoidable piece of evolutionary logic.

(Yes, it is possible that evolution has enriched and complexified the original mechanism of reproduction so thoroughly that it is now dependent on structures that evolved later. But the fact that there is an original mechanism of reproduction, and it is simple direct collapse, moves the odds firmly against that.)

This lets you dismiss all those formation mechanisms which rely on (early) star formation to drive (later) supermassive black hole formation. It simply cannot happen in that order. Likewise, no bottom-up, multistep, hierarchical merger model can work to explain supermassive black holes. The big primary thing can’t be built out of little secondary things. (Once formed, a supermassive black hole can later be fed stars and elephants and iPhones, sure. But it can’t be initially formed by them.)

Similarly, primordial black holes (the only other formation mechanism to make it through that filter) fails, because it requires a whole bunch of complicated (fine-tuned) tweaks to the conditions surrounding the Big Bang in order to work. And such fine-tuned tweaks can’t have PRECEDED the earliest reproduction of universes (as would be required if they ENABLE that reproduction); such fine-tuning could only be the result of later evolutionary pressure. Meanwhile, a simple, smooth Big Bang doesn’t give you – can’t give you – primordial black holes. So, that mechanism, too, fails the Evolutionary Cosmology test.

DO YOU SEE THE POWER OF THIS?

Right now, in cosmology, we don’t have a way to discriminate between mathematically plausible theories. But three-stage cosmological natural selection (and indeed the entire nascent field of Evolutionary Cosmology) gives you a way to discriminate between mathematically plausible theories, because you can do a second check – after “is it mathematically plausible?” – which is, “is it evolutionarily plausible?”; or, more precisely, “is it compatible with the evolutionary history of our universe, given that our universe reproduces in this fashion?”

This is how, back in 2022, I was able to out-predict the entire field of cosmology. Not because I’m cleverer, or more hardworking, or have any particular virtues as a scientist or thinker. But simply because I was working with, I was extending, a better meta-theory. A better framework.

It’s very, very telling that the first use of this principle, when predicting what the James Webb would see, gave such a strong positive result.

And so I think the keystone prediction of Cosmological Natural Selection, when applied to our specific universe and its development, is this: supermassive black holes should be presumed to form first, from smooth gas, by direct collapse, before galaxy formation. They should not be presumed to depend for their formation on any more complex structures, such as stars, or galaxies, which – doing far more complex things, with far more complex matter – must have evolved later in the evolutionary history of universes.

And the supermassive black holes found in our specific universe today therefore are most likely to have formed when conditions in our specific universe most resembled those found in the earliest, most primitive universes; just after the Big Bang, and before star and galaxy formation; so, well inside the first couple of hundred million years. Back when the gas was still smooth enough for direct collapses of large areas, without small local density fluctuations, and thus without breaking up into stars.

This insight, this prediction, emerges purely from evolutionary logic, not from mathematical reasoning or from physics. (Though, of course, it obeys all physical laws, and requires no new particles or physics. It’s a surprisingly conservative theory, simply applying Darwin to universes.) It was first made before the James Webb Space Telescope had sent back any data. And it appears to describe, remarkably accurately, what the James Webb is seeing in the early universe. See, for example, this astonishing paper from last year, describing a huge population of extremely early galaxies, or protogalaxies, dominated by their central supermassive black holes:

“Little Red Dots: An Abundant Population of Faint Active Galactic Nuclei at z ∼ 5 Revealed by the EIGER and FRESCO JWST Surveys”, by Matthee, Naidu, et al, 2024.

Or this great paper on UHZ1, the supermassive black hole that weighs as much as all the stars in its galaxy put together:

First Detection of an Over-Massive Black Hole Galaxy UHZ1: Evidence for Heavy Black Hole Seed Formation from Direct Collapse, by Priya Natarajan (hurray! take that victory lap!), Fabio Pacucci, et al, 2023.

Or this paper on an even larger, sleepy, early, supermassive black hole:

A dormant overmassive black hole in the early Universe, by Ignas Juodžbalis, Roberto Maiolino, William M. Baker et al, 2024.

And so on, and on.

WE ARE ALL ON THE SAME TEAM

Three-Stage Cosmological Natural Selection gives a wider conceptual framework for the pioneering work done on direct collapse black holes over the years by a handful of brave cosmologists and astronomers, and it explains why their colleagues turned out to be wrong to think of direct collapse as just one possible formation channel among many (with various ways of merging stellar collapse black holes as the leading candidates): direct collapse will not turn out to be a rare, unlikely event (even though Wikipedia still says “Direct collapse black holes are generally thought to be extremely rare objects in the high-redshift Universe”); direct collapse will turn out to be ubiquitous, because evolution has fine-tuned our universe to enable it.

The two approaches, those of the pioneers of direct collapse and mine, turn out to be totally complementary. We can only help each other. This is a delightful win/win situation, where the new paradigm doesn’t need to smash the old paradigm; it can just help the old paradigm solve all its problems. Reframe, and explain. We get to keep all the old data, gathered under the old paradigm; we will just understand it better now.

Let’s end with last month’s article in Quanta, scrambling to explain a recently discovered “naked” supermassive black hole, in the early universe, that doesn’t seem to have a galaxy around it yet.

“This new black hole, which is as heavy as 50 million suns and is dubbed QSO1, clashes with the old, provisional account of the galaxy formation process, which did not start with black holes. Black holes were thought to have come along only after a galaxy’s stars gravitationally collapsed into black holes that then merged and grew. But Maiolino and his colleagues described a solitary leviathan with no parent galaxy in sight.
The question now is how this black hole came to exist.”

– Quanta Magazine, “A Single, ‘Naked’ Black Hole Rewrites the History of the Universe”, by Charlie Wood, September 12th 2025.

Obviously, I think I know the answer. (Having predicted exactly what they are seeing now back in 2022.)

It is amusing to see that, three years into the new James Webb Space Telescope era – after three years of discovering ever-greater domination of galaxy formation by ever-larger, ever-earlier supermassive black holes – so many cosmologists, astronomers, and science journalists remain oblivious to the fact that there is a perfectly respectable and conservative theory that predicted all this...

OK, that was fun. Glad I finally got this piece written.

Now, ponder it. Stress-test the logic. See if you think it’s true. And if it is, internalise it. Then go read Blowtorch Theory, if you haven’t already, for more – much much more – information on this evolutionary approach. Or watch this video, if video is your thing.

And I’ll go back to writing my epic spiral galaxy formation piece.

OK, if I’m going to kick cosmology around the room in this post, I should probably define what it is first, and separate it out from astronomy as a whole.

ASTRONOMY VERSUS COSMOLOGY

Astronomy studies the objects we can see in our universe. (So: stars, planets, galaxies, etc…). Cosmology, however, is that subsection of astronomy which studies the origin of our universe, how it has developed so far, what its structure is, what lies in its future, and so on. (So: the big bang, the cosmic microwave background radiation, where the heck all the chemical elements come from, the ultimate heat death of the universe, etc…)

Obviously there is a great deal of overlap between the two; but, as a general rule, astronomy is largely observational: it piles up data.

Cosmology, however, is largely theoretical: building on that data, it theorises about the structure, origin, and development of the universe as a whole.

SUBTLE EVIDENCE OF SOMETHING STRANGE…

Which brings me to a subtle, but important, piece of evidence favouring the idea of an evolved universe (the idea we’re exploring in The Egg and the Rock) over a random and arbitrary one-shot universe (the mainstream assumption).

Favouring, that is, the idea that our universe behaves like an egg (a complex evolved entity undergoing a highly structured process of development), rather than a rock (dead matter, slowly and randomly losing order over time).

Which means favouring the idea that evolution, over countless earlier generations of universe, has fine-tuned the basic parameters of matter to not just allow for, but to actively encourage, enable, and enact, the stable, dynamic, out-of-equilibrium complexity we see all around us; from spiral galaxies, with their steadily glowing multi-billion-year-old stars, to a biosphere like Earth’s, with its complex, ever-evolving menagerie of interdependent creatures.

That subtle piece of evidence is the fact that, in cosmology, over many centuries now, all our errors of prediction have leaned the same way.

We have been consistently wrong about the complexity of the universe, and the efficiency of its processes; and consistently wrong in the same direction.

• The universe has always turned out to be larger than we expected; but more importantly, and more profoundly…

• Our universe has always turned out to be more complex in structure (at both large and small scales).

• Its energy production has always turned out to be more efficient than we had assumed.

• And, remarkably often, energy production that looked at first glance explosive and randomly-aimed, when looked at more closely turned out to be generating (and protecting) complexity in specific regions (such as Earth’s biosphere).

• That is, the flow of energy through the universe is such that, step by step, it builds out and protects complexity, rather than blowing all existing complexity apart.

LET’S GO BACK

Let’s go back to the very beginnings of science – what the hell, let’s go back before science – and just deal with scale. How big did we think the universe was?

For an extremely long time, we thought there was only one planet, Earth.

Then, as we begin to develop modern science, and tools like the telescope, we realised there were several planets (some with their own moons!), but thought that was an end to it: our sun, and those planets, formed the only solar system.

Then we realised – again through closer observation – that the stars weren’t small lights close by, but were distant suns – and that we were, in fact, embedded in a swirling galaxy of such suns.

Naturally we then assumed that our galaxy, the Milky Way, was the only galaxy.

So wedded were we to the idea that there could only be one galaxy that we continued to assume this even after we’d begun looking at other galaxies through powerful telescopes. We called many of them “spiral nebulae” for years (“nebula” meaning fog or cloud in Latin), and assumed right up until the 1920s that they were small clouds of gas, close by, rather than (as so many of them turned out to be) large clouds of stars, far away.

THROWING CATS INTO BLACK HOLES

Even more dramatic (and recent) is the error we made with quasars. A quasar is the blazing consequence of gas circling the drain of a supermassive black hole at the centre of a galaxy; gas that is accelerated to close to the speed of light as it falls, potentially converting up to 42% of its rest mass into energy in the process.

A NON-SCIENTIST SPEAKS: I have no idea what that means. Is that… a lot?

Yes! For comparison, atomic fusion – the marvellously efficient energy source for stars, and hydrogen bombs – only converts maaaaybe 0.7% of an atom’s mass into energy; so the conversion process driving quasars is sixty times more efficient than fusion.

A NON-SCIENTIST SPEAKS: Not to be a bore, but could you say that again in language I can actually see in my head?

Sure! Let’s borrow an example from this excellent video by Henry Reich of Minutephysics : how-much-energy-does-it-take-to-power-Norway, a large country with a population of 5 million (and some very cold winters).

If Norway got all its energy from coal, it would need to burn roughly sixty million tons of the stuff to power it for a year.

To power Norway by fission would take roughly 375 tons of uranium fuel.

Fusion would require only fifty or so tons of fuel.

But you could power Norway for a year by throwing just two and a half cats into a rapidly rotating black hole. (A link well worth clicking on, if you want to get how astonishing all this is.) And yes, the cat has, for some reason, become the standard unit of fuel for black hole energy production. I blame Erwin Schrödinger.

A NON-SCIENTIST SPEAKS: Thank you.

You’re welcome. Where were we? Oh yes.

SAME OLD THING, IN BRAND NEW DRAG

Most of the stars in a galaxy are relatively small (because the bigger stars live fast and die young). These small stars eke out their hydrogen fuel supply with (again) a startling level of precise steady-state control, making it last many billions of years. But the supermassive black hole at the centre of a galaxy is busy doing something very different, and a really big one can gulp down the mass of the earth in gas every second.

With so much gas getting converted into energy with such rapidity and efficiency, the blazing donut of circling gas (called the accretion disc) can shine dozens, hundreds, or even thousands of times brighter than all the actual stars of its galaxy added together – and it can therefore be seen from the far side of the universe.

But when we first detected quasars, in the early 1950s, we yet again assumed they were small bright objects nearby, not colossally large, astoundingly bright objects very far away indeed. We assumed – of course! – that they were inside our own galaxy (which is only a hundred thousand light years across); in fact they were often several billion lightyears away. That’s tens of thousands of times further away than we had expected. And yet again, it took us years to understand what we were seeing.

This is a human heuristic, a consistent problem with the way we think: We do not extrapolate out from our repeated, identical errors, and assume the universe is bigger and more complicated than we can see; even though we have repeatedly underestimated the size and complexity (and energy efficiency) of the universe.

Right now, we are up several more layers of complexity. We have discovered (each time to our surprise) galactic clusters, galactic superclusters, megastructures like voids and filaments… and, yet again, we have stopped, and said OK, that is it. There is just one universe, and we have mapped it.

We have reached the edge of all that is; and now we know it all.

But why would we be right this time? We never have been before…

And bear in mind, all of these colossal expansions of the universe were forced on us by observation (by better and better and better telescopes and spectrographs and probes); none of them were predicted in advance by the mainstream cosmological theories of the day.

Even the Big Bang itself was forced on us by the observation that all of the more distant galaxies were moving away from us (and the further away they were, the faster they were moving away), and so the universe must be expanding from some original point in space and time; but this wasn’t predicted, it was discovered, through observation, by Edwin Hubble in 1929. And it came as an astonishing surprise.

IF IT CAN EVEN FOOL EINSTEIN…

For an extremely clear example of this bad heuristic at work, throwing even the greatest of scientific minds off course, just look at Einstein.

When Einstein was working out his theory of general relativity (which he published in 1917), he was alarmed to discover that the theory was screaming into his face: “THE UNIVERSE IS EXPANDING!” Expanding? GROWING? (Like an egg?) Well, that was crazy talk. Everyone knew that the universe was eternal, static, and unchanging. He assumed his theory must, therefore, somehow, be wrong – and so he went back and fudged it, before publication, adding in an extra, and unnecessary, term (which he called the cosmological constant), just to force the theory to give a static, eternal, unchanging universe. (A rock.) Twelve years later (when Hubble showed the universe was indeed expanding), Einstein called this “my greatest blunder”, but he shouldn’t have been so hard on himself; it was precisely the same blunder that all mainstream cosmological theories had always made.

Everything that indicates growth, everything that indicates expansion, everything that indicates development – everything that indicates the rapid, orderly, complex, fine-tuned, directional, developmental process of an evolved universe – has always come as a huge surprise to those theories and their theorists.

That’s because all such mainstream cosmological theories have been, and still are, based on an underlying unexamined paradigm. Now, “paradigm” is one of those tricky words that can get used a bunch of different ways. By paradigm, here, I basically mean a way of looking at something. (A way of looking that is itself largely unexamined.)

That paradigm assumes our universe is an isolated one-off, made of matter with arbitrary properties, under the influence of arbitrary laws, interacting randomly.

The paradigm used to assume (as in Einstein’s day) that this one-off universe was eternal, static, and essentially unchanging. After Hubble, it had to adjust a little, but it still assumed this was a one-off universe made of matter with arbitrary properties, under the influence of arbitrary laws, interacting randomly… that now, somehow, had recently come into being out of nothing, with no explanation for that, and no history. (Yes, the paradigm was now pretty incoherent, but that’s the problem with unexamined paradigms – being unexamined, new data can force them into total incoherence without anyone noticing.)

INTO THE MULTIVERSE

Theories based on such a paradigm have always had to run hard to catch up with observation. A few brave figures have predicted a larger, stranger universe at each point, but usually as outsiders, and usually in the face of indifference or derision from the mainstream. (And of course the first person to claim that the stars might be suns like our own, complete with exoplanets, and alien forms of life, was burned at the stake – which put that particular theory back a few hundred years…)

A SMALL IRRITATING CHILD INTERRUPTS: But sir, sir, sir! Don’t many top scientists now predict the existence of infinite numbers of universes in a vast multiverse? Surely THEY are predicting a larger universe of universes, and lots more stuff that you can’t see? Huh? Huh? Huh?

Thank you, Small Irritating Child, for that irritating but excellent question.

In fact the multiverse theories you cite provide evidence for my case, not against. Because all of them are recent attempts to explain the extreme unlikeliness of our own universe; but without the evolutionary mechanism to (blindly, but effectively) guide and direct the process of generating new universes, these theories are all doomed to just give you a larger and larger pile of arbitrary universes, with no explanation for our own universe’s complexity, and no mechanism for achieving it: just a claim that, if you rummage in a large enough random pile of universes for long enough, you will probably, eventually, find a complicated one. That’s not a theory; that’s the magical thinking of mathematicians with no real understanding of the limitations, and possibilities, of material reality under the constraints of evolution.

But non-evolutionary multiverse theories (and string theories, and super symmetries, and other doomed attempts at explaining the emergence of complexity from simplicity without evolution) are rich subjects with complex histories, and deserve to be kicked around the room respectfully and at length in their own posts. I can’t do justice to them in an aside. And I want to stay focused here, and actually get a fecking post finished and published. So let’s just say, point noted, and crudely dismissed. More thoughtful dismissal to follow.

Now, let’s get back to this post.

A BALL OF HOT DIRT

The same problem – of our errors all leaning in the same direction – happens at smaller scales, too: at the level of individual stars, rather than galaxies or universes. Until relatively recently, we assumed our sun was a simple, crude, ball of hot gas. Indeed, in the final decades of the 19^th century, one of the greatest physicists of the age, Lord Kelvin, thought the sun might be made of burning coal, or simple dirt, heated only by its own gravitational collapse. (Yes, this is the same process that produces quasars… but a ball of dirt the size of the sun has a far, far less dramatic gravity well than a supermassive black hole the mass of a billion suns rotating at nearly the speed of light, and so it produces far, far less energy.) And remember, Lord Kelvin – originally William Thompson, until he was elevated to the peerage for his many scientific breakthroughs – was the guy who formulated the second law of thermodynamics, and established the absolute temperature scale, later named the Kelvin scale in his honour: nobody knew more about heat and energy than this guy!

Indeed, this was one of the great arguments made against Darwin and his theory of evolution, by the mainstream scientists of his day; the sun, being made of coal, or burning gas, or perhaps just dirt, heated by its own gravitational collapse, could not be more than a few thousand years old (if it relied on the burning of chemicals), or at the most twenty million years old (if it was heated by its own gravitational collapse), or it would have burnt out by now – and therefore Darwin’s claim that some fossils were hundreds of millions of years old had to be false.

Effortless and automatic nuclear fusion occurring deep in the heart of every star, and fusion’s parsimonious efficiency in releasing incalculable quantities of energy slowly and steadily over vast periods of time; this all came as a huge surprise.

But eventually – through direct observation, spectrum analysis, and so on – we discovered that the sun was mostly made of hydrogen (and helium); that it was in fact a plasma, not a gas (and certainly not coal); that it had a complex internal structure, in which multiple layers of plasma rotated at different speeds to generate an extremely stable and reliable (over billions of years) self-exciting dynamo, which then powered a vast electromagnetic field, which mysteriously supported a solar corona hundreds of times hotter than the surface of the sun, a corona that somehow accelerated charged particles to a significant fraction of the speed of light, causing the solar wind, and thus creating the heliosphere, a protective electromagnetic bubble around the entire solar system (keeping out the lethal rain of high-energy cosmic rays that would otherwise sterilise the planets), thus allowing fragile DNA life to develop on the protected planetary surface of our Earth (doubly protected by the electromagnetic field generated by Earth’s OWN stable, reliable, internal, liquid nickel-iron dynamo); and of course we eventually discovered that our sun generated the heat and light which powered Earth’s DNA life, not through chemical reactions, but through nuclear fusion – a process nearly four million times more energy-efficient than burning coal.

Lord Kelvin was so sure he was right that he attacked Darwin for years, forcing Darwin to take out some of his (accurate) claims about the age of the earth in subsequent editions of On the Origin of Species. (Darwin wrote to William Wallace in 1869, “Thomson [later Lord Kelvin]’s views of the recent age of the world have been for some time one of my sorest troubles.”) But Lord Kelvin was, fairly spectacularly, wrong. And he was wrong because of an unexamined paradigm that underlay his thinking.

DOGMA BITES MAN

There is a word for an underlying, unexamined paradigm, a paradigm that you are mocked (or punished) for questioning. A shorter word, a punchier word. That word is “dogma”.

And there are a number of other words for an underlying, unexamined dogma that consistently causes you to make mistakes that all lean in the same direction. Many shorter, punchier words. One is “wrong”.

A FIERCE REVOLUTIONARY YOUTH: Another is “bullshit”.

Er, you might very well think that; I couldn’t possibly comment.

A FIERCE REVOLUTIONARY YOUTH: But why doesn’t anyone notice this cascade of identical errors, and do something about it?

Well, partly because the history of cosmology, like the history of any science, is full of errors of all kinds, pointing in all sorts of directions. It is rare for anyone to stand back and look for a consistent underlying pattern in the errors, a signal in the noise.

And that is partly because the history of cosmology, and of astronomy more generally, tends to downplay just how many of these mistakes were made, and how much the truth was resisted at the time.

As Bill Bryson dryly put it, in A Short History of Nearly Everything,

“There are three stages of scientific discovery: first, people deny it is true, then they deny it is important. Finally, they credit the wrong person.”

MAN BITES WOMAN

Even the now-uncontroversial observation that the sun is in fact made of hydrogen was fiercely resisted by the mainstream, in its day, thanks to that bullshit. Sorry, dogma. Sorry, unexamined paradigm (or way of looking at things). Back at the start of the 20^th century, the sun and the earth were assumed to be made of the same stuff, in the same proportion, because (unexamined paradigm!) everybody in the sciences simply knew that the entire universe should behave like a rock, not an egg. One stuff, broken up into lumps of different sizes. That was the totally accepted, mainstream argument, made by distinguished men of science (because women, up until this point, had never even been allowed to get degrees from universities, let alone hold senior positions in the sciences); men such as the President of the American Physical Society (and first occupant of the chair of physics at Johns Hopkins University), Henry Augustus Rowland.

As the equally distinguished American astronomer, and Henry, Henry Norris Russell put it, in 1914:

“The agreement of the solar and terrestrial lists is such as to confirm very strongly Rowland’s opinion that, if the Earth’s crust should be raised to the temperature of the Sun’s atmosphere, it would give a very similar absorption spectrum. The spectra of the Sun and other stars were similar, so it appeared that the relative abundance of elements in the universe was like that in Earth’s crust.”

But, even as he was speaking, a young Englishwoman, Cecilia Payne, was battling her way into the university system, against tremendous odds.

Cambridge (the English one) let her study astronomy, but wouldn’t give her a degree (because she was a woman), so she moved to the US. In 1925, she got the first PhD in astronomy to be granted by Radcliffe College, the “female coordinate institution” for the all-male Harvard College. (Harvard’s treasurer, on the idea of females receiving actual Harvard degrees: “I have no prejudice in the matter of education of women and am quite willing to see Yale or Columbia take any risks they like, but I feel bound to protect Harvard College from what seems to me a risky experiment.”

In the process, she proved, in her thesis, that hydrogen was a million times more plentiful in the sun than it was on earth.

And if the sun was made of hydrogen, then all stars were presumably made of hydrogen.

She had discovered what the universe was made out of.

Many many years later (in 1960), the great astronomer Otto Struve called it “the most brilliant PhD thesis ever written in astronomy.”

But that wasn’t how it landed at the time.

Because the dogma of the time was that the universe can’t have dedicated, specialist, differentiated parts, like an organism. No, everything everywhere is the same; metaphorically speaking, the entire universe is all just one shattered but undifferentiated rock. Granted, some of it seems to be so hot it’s a gas, and some of it seems to be colder, so it is solid, but that’s all that’s different about it… And so she was bullied out of telling the truth by her PhD supervisor, who happened to be… Henry Norris Russell.

Now, Russell was a great astronomer (he’s the Russell behind the Hertzsprung-Russell diagram, still an extremely useful tool in astronomy), and I’m sure he genuinely, at the time, felt he was saving her from making a fool of herself. But but but… well, he was a brilliant insider, trapped inside an unexamined paradigm, and Cecelia Payne was a brilliant outsider who was not trapped by the paradigm and who could therefore see reality more clearly. Russell was wrong, and when given the right answer, he blocked it, and delayed the breakthrough for years.

Forced to remove from her thesis one of the most momentous astronomical breakthroughs of the 20^th century – and maybe the most momentous breakthrough that there had ever been, up until that point in history, in our understanding of stars, and thus the universe – Cecelia ended up sadly describing her own results as “spurious”.

What’s infuriating is that Russell eventually, a few years later, realized that Cecilia Payne was correct, when he tried a different approach, and ended up with the same results as her. He, of course, believed himself, the insider, where he had failed to believe her, the outsider, and, in 1929, published his findings. His paper made a passing mention of Payne’s earlier work (“…the most important previous determination of the abundance of the elements by astrophysical means is that by Miss Payne…”), but he still frequently gets credit for the discovery.

But that’s an aside that probably deserves a post of its own.

IT’S NOT JUST A FLAW IN THE THEORY

Let’s return to the problem that these breakthroughs alway point in the same direction: the universe is larger than we expected; the universe is more complicated than we expected; the universe is more energy-efficient than we expected; that energy is more meaningfully directed than we expected.

The fact that these repeated errors all point in the same direction is a sign that there is a major flaw in our entire approach. And that the flaw is likely to be an unexamined assumption underlying the whole theory, rather than an explicit and visible part of the theory itself.

If there were merely a flaw in the theory – a mathematical error of some kind, say – then those repeated identical errors should, by now, have given us the necessary information, the necessary feedback, to find the flaw in the theory, and fix it. Our errors should no longer all lean the same way. But with a flawed underlying assumption that isn’t explicitly articulated inside the theory, you can go wrong in the same way again and again and again, without getting useful feedback.

PHYSICS IS PHYSICS; BUT…

Of course, there is a sense in which it shouldn’t matter whether you are studying an egg or a rock; physics is physics, and the same rules apply. But there is another sense in which it matters a lot; the way physics plays out in an egg is different to the way physics plays out in a rock of roughly the same size and chemical composition. Same physics; very different outcomes.

And so my argument is that until the mainstream scientific community change from a universe-as-rock paradigm to a universe-as-egg paradigm, they will continue to be blindsided by the unanticipated complexity of our universe’s structure, by the startling efficiency of its processes, by the unexpected discovery of new, dynamic, out-of-equilibrium systems at all levels, by the surprising intricacy of its interlocking parts… In other words, by the many unanticipated ways in which the basic parameters of matter have been fine-tuned by evolution to interact, under specific developmental conditions, so as to generate structure, complexity, order, and efficiency, so as to ensure reproductive success.

And, above all, until they start using egg physics rather than rock physics, they will be blindsided particularly badly in the early universe; particularly in the first billion years – the last refuge of randomness – where they thought they would, finally, find random matter blindly obeying arbitrary laws – and where instead, again and again (as I predicted), they are finding the structure, and order, of an evolved organism efficiently and rapidly proceeding along a clear developmental path.

OK, that’s it for now! Feedback, whether positive or negative, very welcome. Add a comment, or simply hit like (or refuse to hit like! That’s feedback too!) below. Talk again soon…

(PS: Yes, I am trying to force a paradigm shift in how we think about the universe, and our place in it. Such revolutions happen one mind at a time. I suspect this post would make a pretty good, easy starting point for anyone wishing to engage with the idea of an evolved universe, so if you enjoyed it, please do pass it on to a friend you think would also enjoy it. These tiny actions will make a huge difference over time. We don’t have to be passive in the face of a bad paradigm. It’s our universe too. Also, we need to save the next generation of brilliant cosmologists, astronomers, and astrophysicists from laboriously climbing a ladder that is currently up against the wrong wall. So please, take a minute, now, to think about who you know who might find this fascinating; and pass it on. Thanks!)

For scientists interested in citation, this post can be cited in Chicago style as follows:

Gough, Julian. “The Blowtorch Theory: A New Model for Structure Formation in the Universe.” The Egg and the Rock, March 19, 2025.

THE PROBLEM

We have known since the 1970s that our universe has a complex structure. Dense nodes, packed with galaxies and gas, are connected by long, thin, filaments of galaxies and gas, all surrounded by largely empty voids. This structure resembles the neural network in a brain, and is known as the Cosmic Web. Its extraordinary scale and complexity was not predicted in advance of observation, and came as a huge surprise. So, how did all this unexpected structure form?

THE CURRENT, PASSIVE, ANSWER

The current mainstream answer to that question is called Lambda Cold Dark Matter (ΛCDM ), and it is entirely passive. Based on gravity gradually pulling everything into shape, it requires huge quantities of (unfortunately, to date, entirely theoretical) “dark matter” to work. (Plus a lot of “dark energy” – that’s the lambda bit.) Despite decades of refinement, Lambda Cold Dark Matter still can’t coherently explain all the relevant observed phenomena without making some alarmingly ad hoc, post-observation adjustments. (See the Cusp/Core problem; See the Missing Satellites problem; etc.)

Crucially, it also predicts bottom-up structure formation, with mature galaxies assumed to form gradually, and hierarchically – that is, through a long, slow, random process of repeated gravitational mergers between much smaller, messier, and unstructured clumps of stars. This means it utterly failed to predict the rapid, early, efficient formation of huge numbers of large, bright, massive, startlingly mature galaxies revealed by NASA’s new James Webb Space Telescope over the last two and a half years. Whether dark matter actually exists or not, it clearly isn’t sufficient by itself to explain what we observe.

A new theory of structure formation is therefore required.

AN ALTERNATIVE, ACTIVE, ANSWER

In this post, I will lay out, for the first time in one place, a new, active, alternative model: the Blowtorch Theory of structure formation. It argues that large numbers of extremely early, sustained, supermassive black hole jets actively shaped the universe’s structure in its first few hundred million years, largely through electromagnetic processes. These jets form vast, low-pressure cavities in the dense gas of the compact early universe, and lay down magnetic field lines, that – expanded by the universe’s growth, and shaped by later gravitational and kinetic events – go on to form the voids and filaments of the Cosmic Web we see today.

A MORE FRUGAL ANSWER

Importantly, this entire process of active structure formation by jets (assisted by gravity from ordinary matter) can be fully described without any need for dark matter. No new particles, and no new physics, are required.

SUPPORTIVE EVIDENCE

Recent evidence of unexpectedly large and early supermassive black holes (such as UHZ1 – which is as massive as all the stars in its galaxy put together), unexpectedly large and early jets (such as Porphyrion – which is a couple of hundred times longer than our Milky Way galaxy is wide, and indeed 40% longer than theory said was possible) – and, above all, large-scale extremely early rapid galaxy formation around active supermassive black holes, gives a great deal of support to this new theory.

(And even as this Blowtorch Theory post was being researched and written, a paper was published detailing an extraordinary blazar – a jet, a blowtorch, pointed straight at the earth from over 13 billion years ago, just 750 million years after the Big Bang – far earlier than Lambda Cold Dark Matter predicted, but slap-bang where the theory outlined here said we would find them. See: A blazar in the epoch of reionization, by Eduardo Bañados et al, Nature, December 17, 2024.)

The second half of this post will outline the parent theory – three stage cosmological natural selection – which successfully predicted these extremely early supermassive black holes, and their jets, plus the associated rapid early galaxy formation, in advance of the first James Webb Space Telescope data.

Blowtorch theory works, and can be explored, independently of its parent theory: however, three-stage cosmological natural selection gives an important and useful framework for more deeply understanding blowtorch theory and its implications.

But let’s start by laying out in more detail the problem our new theory solves.

INTO THE VOID

“And if you gaze long enough into an abyss, the abyss will gaze back into you.”

– Friedrich Nietzsche, Beyond Good and Evil (1886)

THE PROBLEM, IN MORE DETAIL

Cosmic voids are vast regions in our universe containing almost no stars, galaxies, or gas.

They’re not vague, blurry, density fluctuations, either. These voids are sharply bounded, with densities about an order of magnitude lower than their surroundings. In the main body, density is less than 10% of the universe’s average (often much less); near the edge, it rises to 20%, and then sharply to 100% at the walls.

We didn’t predict their existence.

We only discovered them in 1978. (Hats off to Gregory and Thompson, and, independently, Jõeveer, Einasto and Tago.)

And we were startled to discover they take up more than 80% of the universe’s volume, while containing only a tiny fraction of its matter.

3D MAPS

In the late 1970s, large-scale redshift measurements of galaxies – tracking how far their visible light shifted into the red as they moved away from us – allowed us to map the universe with depth for the first time. Before this, our maps were essentially 2D, like a sheet of paper, with near and far objects overlapping; we had no way of knowing if a vast empty space separated them.

But with these new 3D maps, we discovered that over 90% of all stars, galaxies, and gas are crammed into just 20% of the universe. In fact, the majority of stars and galaxies by mass are packed into the dense regions we call clusters and superclusters, which occupy less than 1% of the universe’s volume. These clusters form dense nodes, connected by filaments along which gas appears to travel in massive flows.

The result is a dynamic network that resembles a brain’s neurons, or a city’s transport network, far more than it does random clouds of gas.

If you have trouble with cosmological visualisation (some don’t, some do), just imagine the universe as a country, and galaxies as buildings, ranging in size from huts to gigafactories: nearly all the buildings, especially the largest, are packed tightly together in isolated villages and towns (dense nodes containing clusters and superclusters), taking up only 1-2% of the land. An extensive network of roads, motorways, and canals (the filaments) connects these hubs, spreading over ten to twenty times as much land, mostly to transport gas for the star-making factories in the towns. The remaining 80% of land is almost empty. (Voids.)

These complex, network-like structures came as a huge shock to astronomers. 1981 was the year it became a crisis, when Robert Kirshner, Stephen Gregory, and Paul Schechter discovered the Boötes Void. It was 250,000,000 light years in diameter – so you could line up 2,500 copies of our own Milky Way galaxy, side by side, inside it. Such a vast space should contain thousands of galaxies – but the Boötes Void (now often known fondly as the Great Nothing) only contained roughly sixty…

This wasn’t just a failure of prediction; it was a revealing failure of observation. We had been looking at the entire universe – through a huge range of telescopes, across all frequencies, from radio to x-rays – for decades. Yet we had completely failed to see the actual structure of what we were observing.

This failure wasn’t entirely the fault of the astronomers; voids, after all, are empty, making them easy to overlook. But discovering the scale and sharpness of these voids threw into high relief a deeply revealing incorrect assumption that underlay all cosmology.

Astronomers had assumed, right up until this point, that our universe, on the larger scales, could be treated like a simple gas in equilibrium, with galaxies behaving like gas molecules, all spread out randomly. Sure, theorists like the great Belarusian physicist Yakov Zeldovich (father of the Soviet hydrogen bomb) had done some interesting work, in the early 1970s, on how matter might collapse asymmetrically under gravity – but prior to the redshift surveys of the late 1970s and early 80s, any astronomer or cosmologist would have told you that our universe couldn’t contain anything like the Böotes Void – let alone something like the Sloan Great Wall (discovered in 2003), a filament stuffed with galaxies and gas that’s 1.4 billion light-years long. From a distance, statistically, the universe was assumed to be entirely smooth, with no structure. Simple matter, obeying simple gas laws.

This incorrect assumption emerged from an even more fundamental unexamined assumption: that our universe was a one-off, with no history, in which randomly distributed matter with arbitrary characteristics blindly obeyed arbitrary laws, with random consequences.

That the first basic assumption had turned out to be utterly, eye-wateringly wrong should have led to some introspection in cosmology, astronomy, and astrophysics about the validity of the second, and even more fundamental, unexamined assumption. But it didn’t.

THE MAINSTREAM SOLUTION – LAMBDA COLD DARK MATTER – AND ITS FAILURE

Instead, they back-engineered a new theory from scratch, to fit the startling new data. Fair enough – when you’ve got something completely wrong, you need a new theory. But this “new” theory tried to fix the problem as simply as possible (since the deeper unexamined assumption was still that the development of our universe was just the simple addition of random processes), and thus relied on gravity alone. That might seem odd to you (and it should certainly seem odd to you after reading this post), given that electromagnetism is an astonishing 36 orders of magnitude stronger than gravity. That means the electromagnetic force exerted by a single electron is 1,000,000,000,000,000,000,000,000,000,000,000,000 times stronger than the gravitational force it exerts. (Thus a modest fridge magnet, or some static from your hair, can cause an object to defy the gravitational pull of the entire Earth.) How can you leave electromagnetism out of the picture? But, in fact, this decision made total sense (given their unexamined foundational assumptions), and was entirely uncontroversial at the time.

That’s because gravity is simple, and additive. Gravity can’t be blocked, reversed, or cancelled. Add more matter; it bends spacetime further; you get more gravity. That’s it, done. A large mass of matter will therefore always attract more matter gravitationally, and get bigger over time. But any large positive electromagnetic charge will attract nearby negative charges, which will immediately neutralise it. And so, in a random, chaotic, structureless environment, gravity is self-reinforcing: but electromagnetism is self-cancelling.

And so they understandably (but as we shall show, wrongly) assumed that electromagnetism couldn’t have any effect at cosmic scales.

Zeldovich got back to work and, in papers like Giant Voids in the Universe (written with Einasto and Shandarin, 1982), and The Large Scale Structure of the Universe 1. General Properties. One- and Two- Dimensional Models (with Arnold and Shandarin, also 1982) found a way that gravitational collapse could give you something that approximated, roughly, to filaments and voids. Sure, the voids he got in these papers were too small, and too round (compared to what we were seeing), and there were a lot of other problems (it couldn’t really explain why galaxies were stable, and why everything didn’t just continue collapsing). But it was a promising approach, and we didn’t have any other theories. So, other cosmological theorists adopted it (Zeldovich himself died in 1987), and got to work developing it, adding an “adhesion model”, and a few other tweaks, to the original Zeldovich approximation, to try and make it work (i.e, get matter to stick together and actually produce galaxies).

There was one problem, however. When they added it all up, the matter they could see – the stars, galaxies and gas made from all the standard particles we know about, called for short “baryonic matter” – didn’t have enough mass to pull everything into these extreme shapes by their gravity alone. But gravity was believed to be the only force capable of shaping anything at cosmic scales! This forced the theorists to introduce massive quantities of dark matter – imaginary particles that don’t exist in the standard model of particle physics, invented purely to patch up issues like this one (and so also the galactic rotation problem, certain gravitational lensing anomalies, some odd features of the Cosmic Microwave Background, etc.)

Since no one’s ever seen any dark matter, it can be given any properties you want. And that, regrettably but understandably, makes it extremely seductive.

Sprinkle enough of this marvellous stuff you’ve just invented precisely where you’d like it, give it the properties you most desire, and – what a wonderful surprise! – you can coax everything into the shapes we observe, using gravity alone. Kind of. If you squint. And if that doesn’t quite work, just add in Dark Energy (first proposed in 1998 - yeah, that’s the Lambda again). Or maybe Dark Flow (first proposed in 2008). Or Dark Radiation (first proposed in 2009). Or my personal favorite, Dark Fluid! (First proposed back in 2005). It has “negative mass” – yes, just like phlogiston (another imaginary substance that mainstream science once firmly believed in, for reasons which made a lot of sense at the time, but which turned out to be wrong). Whole Dark Sectors now exist: rich landscapes of imaginary quantum fields, across which wander escaped zoos of hypothetical particles, none of which have ever been observed. Dark photons! Sterile neutrinos! Axions! A new “dark” gauge group, why not, that is completely unconnected to the actual Standard Model gauge group! True, many of these later suggestions are fringe theories, with little support. But such new theories keep being proposed, because Lambda Cold Dark Matter, for all its undoubted successes (and you don’t get to be the leading mainstream theory without some big successes), keeps running into new problems, and therefore needs all the help it can get.

As you can see, with dark matter, like black tar heroin, once you’ve started, it’s hard to stop. And as time passes, you need more and more to get the same hit. After fifty years of tirelessly tweaking this wonderful, seductive, never-quite-there theory, how are we doing? Well, to drag everything into shape using only gravity now requires five times more dark matter than all the actual, observable, baryonic matter in our universe put together.

That’s the current mainstream explanation for filaments and voids. Five invisible universes worth of dark matter, with ever-changing properties that we never seem to be able to quite nail down, piled up on top of the one we can see. That – bizarrely – is considered the respectable, conservative theory.

Great. There’s only one real problem: this respectable new theory is based on the same set of deep, foundational, flawed assumptions which led to the original embarrassing failure of prediction. The theory is, therefore, unfortunately, wrong.

PROBLEMS WITH ΛCDM

Does it approximate to the data? Sure, but so did Ptolemy’s earth-centred model of the universe (once you’d added enough epicycles). But there are, unfortunately, ongoing problems.

The Bullet Cluster was held up as evidence that dark matter is cold and collisionless. But the Trainwreck Cluster was held up, for a while, as evidence that dark matter is not cold, and not collisionless – that it’s warm, and self-interacting. (Warm dark matter has also been used to explain the cusp/core problem, etc.) If you keep changing the very fundamentals of what dark matter is, to suit each new, ambiguous, observation – if there is simply nothing solid there to hang onto – then you don’t have a theory.

Dark matter is basically a tiny, beautifully embroidered scientific duvet you can pull into position to cover any exposed area – but only at the price of exposing a different area.

For example, if you tweak your dark matter to suit big galaxies, it stops working for small galaxies. (The vexéd cusp/core problem.)

Meanwhile, adjust it to perfectly explain the numbers and sizes of the galaxies we see all around us today, and it massively overpredicts the number of extremely small dwarf galaxies in the early universe.

And the theory didn’t predict, and can’t explain, the recently discovered Ultra Diffuse Galaxies, DF2 and DF4 – which behave as if they contain no dark matter at all.

And now the new James Webb Space Telescope has finally shown us what’s happening in the first billion years after the Big Bang… and Lambda Cold Dark Matter predicted none of it.

STRUCTURE, STRUCTURE, EVERYWHERE…

The early universe turns out to be rich in structure – including huge numbers of mature galaxies when the universe was just 4 or 5% of its current age, some containing supermassive black holes as massive as all the stars in their galaxy put together. (So, not the small, messy, random clusters of stars, and slow, bottom-up galaxy formation, that ΛCDM predicted.) But don’t worry, theorists are now busy “explaining” these, after the fact, by fiddling with their numerous free parameters (some models have up to ten) – essentially adding epicycles.

This doesn’t mean dark matter doesn’t exist (though I personally believe it does not, or certainly not in the form, and quantity, ΛCDM suggests) – but that, even if it does, it’s clearly grossly inadequate, on its own, to explain structure formation in our universe. With or without dark matter, there’s something else going on.

Back here on Earth, fifty years of increasingly expensive attempts to directly detect it have turned up nothing at all. (Except, of course, more funding for bigger versions of the same failed experiments.)

If the theory keeps failing at this scale, it might not be the universe that’s wrong, it might be the theory. Perhaps we’ve reached a point where we should stop trying to modify the universe with more and more imaginary matter, and start looking for a new theory.

The trouble is… It’s too late. For the past twenty years, the largest and most costly computer simulations of structure formation in our universe – running from the original Millennium, Millennium II, and Millennium XXL models, right up to the more recent Uchuu and AbacusSummit simulations – only include dark matter. Yeah, actual baryonic matter – the entire visible universe – is treated as a rounding error, and left out. Thousands of published papers are based on these simulations.

Are there some simulations that include baryonic matter as well? Recently, yes, sure – EAGLE, SIMBA, IllustrisTNG, Magneticum, Horizon-AGN, FLAMINGO – but they typically cover much smaller areas, with some just modelling the formation of a single galaxy. (FLAMINGO is the exception here, in trying to model large-scale stuff too.) This is in many ways understandable: the behaviour of baryonic matter (i.e. the real world) is absurdly complex compared to that of dark matter (which is a mathematician’s dream of simplicity – hi again, Zeldovich!). It’s therefore much harder to model, and far more costly to compute, which is another reason everybody left it out until computers got fast enough to handle it.

That’s not the real problem, though. The real problem is that none of these new simulations that incorporate baryonic matter simulate it from first principles: they simply adjust a whole bunch of feedback parameters that, in the real world, we do not know (black hole feedback, star formation rates, gas dynamics), until they get a result that resembles observations. Hey, if I tune all these dials enough, the result looks like a galaxy! Cool! Are the feedback parameters chosen accurate to reality? Er, nobody knows! They are just another bunch of tuneable parameters, piled on top of the six tuneable parameters for dark matter. This is extremely expensive CGI, not science. It makes ΛCDM completely unfalsifiable.

Or, as the great Von Neumann put it,

“With four parameters I can fit an elephant, and with five I can make him wiggle his trunk.”

– John von Neumann

Always reactive, never predictive: the dog of theory is simply chasing the car of observational data. The car turns left, the dog follows. Car turns right, dog follows. It’s a worrying sign, that the dog never knows where the car will go next.

TWO CHEERS FOR COLD DARK MATTER

However, I don’t want to be too dismissive of Lambda Cold Dark Matter as a theory. No theory gets to dominate its field without some juicy, solid wins. Back in the early 2000s, Lambda Cold Dark Matter accurately predicted the Cosmic Microwave Background power spectrum – that is, it predicted the height of the soundwaves sloshing about in the tiny, ultra-hot, ultra-dense, fluid, early universe – in the brief era before light and matter uncoupled and released seventy percent of all the photons in the universe today in a single, universe-wide flash. (That flash, redshifted by the expansion of the universe, now forms the Cosmic Microwave Background.) NASA’s WMAP (the Wilson Microwave Anisotropy Probe), and the European Space Agency’s Planck space observatory, both confirmed waves of the specific heights predicted by ΛCDM – heights that didn’t make sense with just baryonic matter, but did with a lot of dark matter (and even more dark energy). That was a BIG win; after that, Cold Dark Matter became the “standard model” of cosmology.

MORE MATTER, OR LESS GRAVITY? (COKE, OR PEPSI?)

But if I’m going to mention that success, then I should also mention the successes of the main rival theory to ΛCDM, Modified Newtonian Gravity, or MOND. If ΛCDM tries to fix all the problems in the universe with more matter, then MOND tries to fix all the problems in the universe with less gravity. MOND’s also been around since the early 1980s, but, in 2021, it finally developed a model – the Aether-Scalar-Tensor framework, or AeST – which ALSO maps perfectly onto the acoustic peaks revealed by WMAP and Planck. (It does it by proposing a new vector field and scalar field that duplicate the effects of Cold Dark Matter in the early universe – see “Aether scalar tensor theory: Linear stability on Minkowski space”, by Constantinos Skordis and Tom Złośnik, and the more recent “Aether scalar tensor theory: Hamiltonian Formalism,” by Marianthi Bataki, Constantinos Skordis, and Tom Złośnik.)

So you can get a fit to the data with an imaginary, invisible new particle; but you can also get a fit to the data with an imaginary, invisible new field. I worry that the real truth we are uncovering here is that imaginary, invisible new things, with a bunch of free parameters, can always be made to fit the data.

ΛCDM and MOND are in fact now beautifully balanced in terms of success: ΛCDM is the most empirically successful for big-picture cosmology (except recently, as we have seen, in the early universe), while MOND is the most empirically successful on small, single-galaxy scales. And both are pretty bad at what the other is good at.

But MOND arrived at a full cosmological theory second, and so – despite being just as successful (and just as unsuccessful) as ΛCDM – must play Pepsi to ΛCDM’s Coke.

And Coke, as we have seen above, now dominates the world.

The fact that we have two theories with almost identical success and failure rates – yet one is built into every model, and the other completely marginalised – is a sign that what is going on here is sociology, rather than science.

But anyway, as a result of all this, dark matter is now baked into every mainstream model, as the single explanation for everything problematic, making it impossible to fix the actual underlying problems. They can’t even see what they’re looking at.

And so a lot of brilliant people remain trapped inside a faulty paradigm.

By contrast, the model I’m exploring actually predicts shit in advance.

Feel free to check: Predictions here.

Confirmation here…

…And in more technical form here…

…And more confirmation in accessible form here…

…And in more technical form here…

…And even more confirmation in accessible form here…

…And in more technical form here…

So, is there a theory that can explain voids, filaments, and their magnetic fields, using only the matter we can observe – the particles in the Standard Model of particle physics – and the established laws of nature?

Yes.

It’s called Blowtorch Theory. No, you haven’t heard of it, because this is the first time it’s been laid out in full in public like this. (Yes, as some of you know, I’ve been working on it, behind the scenes – talking to scientists, researching, getting feedback – for months; building on ten years of earlier research for a book on cosmological natural selection. The specific, original inspiration for this post/paper, however, was this terrific paper in Nature, back in September 2024, by Martijn S. S. L. Oei, Martin J. Hardcastle, Roland Timmerman, et al, on extremely large, sustained jets: Black hole jets on the scale of the cosmic web.) But first, an uncharacteristically modest note, giving praise where it’s due…

WOBBLING HEROICALLY ON THE SHOULDERS OF GIANTS

Blowtorch theory ultimately emerges from cosmological natural selection (as I’ll explain later), and therefore grows from the American theoretical physicist Lee Smolin’s seminal ideas (which were in turn influenced by innovative work on evolutionary biology by the great Lynn Margulis and Stephen Jay Gould), as well as excellent later work by Clemént Vidal, John Smart, Louis Crane, Michael E. Price and others. Of course their work, like mine, owes a profound debt to earlier work on black holes by those titans of cosmology, John Wheeler, Kip Thorne, Roger Penrose and Stephen Hawking, and to breakthroughs in magnetohydrodynamics by Steven Balbus and John Hawley (plus many others). I’m also borrowing from Priya Natarajan, Volker Bromm, Avi Loeb, and Marta Volonteri’s pioneering work on direct-collapse supermassive black holes. And of course, I’m building on the legacy of countless astronomers; far too many to list here individually. (Thank you, thank you, thank you!)

So, that was the background – both the science and the sociology, because both are important if you’re to fully understand the peculiar situation we are in.

Here’s the theory.

THE BLOWTORCH THEORY OF STRUCTURE FORMATION

I’m calling it the Blowtorch Theory, because new theories need a punchy name to break into the broader conversation. (But, if you’d be more comfortable with something respectably scientific-sounding, feel free to call it the Directed Plasma Overpressure Model, or the Localized Radiative Jet Confluence Hypothesis, or some other bullshit no one will remember.)

It starts with two observations.

The first: every one of the trillion or so galaxies in our universe that’s big enough or near enough for us to observe closely, appears to have a supermassive black hole at its center, with masses ranging from millions to billions of times that of our sun. And they have a power-law-like distribution: a few of the bigger ones, lots of the smaller ones.

The second: the Cosmic Microwave Background Radiation – the light emitted by all that hot gas shortly after the Big Bang – is incredibly smooth. We know that the only density fluctuations in that gas are very subtle, less than 0.001%: we think they come from early quantum events, blown up vastly in size by inflation since the Big Bang. These density fluctuations occur at all scales, and have a power-law-like distribution: a few big ones, lots of smaller ones. Oh, and at the range of scales that would match the volumes of gas required to make the earliest, and therefore smallest, versions of all those supermassive black holes (they’ll grow in size later), there are, to a very rough approximation – to within an order of magnitude – around a trillion of them.

Let’s put those together.

Now, the smoothness of the Cosmic Microwave Background Radiation shows that the initial gas cloud in the early universe was much too smooth for easy star formation. (You need small, dense, pockets of gas in order to form stars.) This is a problem for the current, mainstream, bottom-up theories of star formation and galaxy formation, which start with the formation of scattered individual stars, somehow, from this smooth gas.

But this smoothness (with its occasional large-but-subtle, not small-and-extreme, density variations) is actually ideal for direct-collapse supermassive black hole formation. When the smooth gas cloud collapses across the entire large-but-subtle density fluctuation, there are no local, small, dense pockets that could otherwise break up into stars. It’s this smoothness which allows the entire vast field of gas to collapse directly into a single supermassive black hole.

And so blowtorch theory argues that all of these supermassive black holes must form now, inside the first hundred million years, and almost certainly within the first 50 million after the Big Bang. And they must form by direct collapse, semi-simultaneously, in a rapid, abrupt phase transition from what was, until then, an ultrasmooth cloud of gas.

AN OBJECTION CONSIDERED

A perfectly understandable mainstream objection would be that the fluctuations are far too subtle, too small – just 0.001%! – to trigger such a direct collapse of such a large area of gas. My argument, however, is that the conditions in the early universe are fine-tuned so that such fluctuations nonetheless do trigger such collapses, at a specific point in the expansion of the smooth gas.

I do understand mainstream resistance to this kind of argument-by-consequences, which can seem lacking in rigour. But please note that, similarly, in 1953, the British astronomer Fred Hoyle, using precisely this logic, predicted a highly unlikely resonance frequency in the carbon-12 atom, that would allow for the extremely efficient, complex (and unlikely) fusion mechanism of the triple-alpha process (where, deep inside a star, three helium-4 nuclei – alpha particles – are turned into carbon). He didn’t predict this using mathematical logic: he predicted it because, when he looked around our universe, he saw a shit-ton of carbon everywhere, and it had to come from somewhere. Fusing three helium nuclei would do it, however unlikely that seemed. Hoyle had to urge his more cautious colleague, the nuclear astrophysicist William A. Fowler, to look for the necessary resonance. Grumbling, Fowler looked. And found it. (Amusingly, Fowler got a Nobel for this, in 1983; Fred, who had actually come up with the idea, didn’t, partly because he had annoyed too many people with some of his other, sillier ideas. Yes, I am aware that it’s a thin line between genius and just being spectacularly wrong! But walking that line is where all the fun is to be had…)

Similarly, in 1982, Dan Shechtman discovered quasicrystals with non-repeating, five-fold (or ten-fold) rotational symmetries were possible. This did not just require an unlikely level of fine-tuning: it was explicitly ruled out by the established laws of crystallography at the time. His paper was repeatedly rejected, the head of his lab told him to “go read the textbook,” then asked him to leave the research group for “bringing disgrace” on the team. When the paper was finally published, two years later, the most famous chemist in the world, Linus Pauling pissed all over him, saying, “There is no such thing as quasicrystals, only quasi-scientists.” But, when people finally actually looked, they discovered Shechtman was right. (He got an apologetic Nobel in 2011.) And the aperiodic, yet ordered, structures he discovered have novel mechanical, thermal, and photonic properties that classical crystals lack. Again, a highly unlikely level of fine-tuning that allowed for a complex outcome turned out to be true.

I predict it will, again, here, for reasons that will soon become clear.

PHASE TRANSITION

The ultrasmooth gas, expanding after the Big Bang, resembles a supersaturated cloud, which on a hot summer’s day can transition extremely rapidly from not-a-single-raindrop to a torrential shower, comprising millions of drops, all produced semi-simultaneously.

As the smoothness of the gas ensures nearly uniform conditions across the early universe, and as the subtle density variations are very thoroughly distributed across the universe, many regions should hit the critical threshold at the same time. In that phase transition, you should therefore see multiple, semi-simultaneous direct collapses – just like rain abruptly beginning in many spots within a saturated cloud.

It’s likely that the shockwave from each of those initial direct collapses would compress nearby gas, pushing it past the collapse threshold, and thus setting off further collapses, in a cascading wave.

So you would see chain reactions in many localised regions, till they all meet – rather as, say, crystallisation spreads during the phase transition from water to ice.

WHY MUST ALL THE SUPERMASSIVE BLACK HOLES FORM EARLY?

The supermassive black holes all need to form together, now, at this early stage, from the smooth gas, because they won’t be able to later. Why? The gas will no longer be smooth enough – because these new, massive black holes are about to fuck up that smoooooooooth gas real good.

SWITCHING ON THE BLOWTORCH

Remember, each supermassive black hole starts with a mass of tens or hundreds of thousands, maybe millions, even hundreds of millions of suns, depending on the size of the fluctuation that triggered it. (The bigger ones are much less numerous. But they’ll all grow larger than this, later, as they eat more gas, and as some galaxies merge.) This intense concentration of mass begins to gravitationally attract enormous quantities of gas.

Why doesn’t this gas orbit the black hole endlessly, and frictionlessly, like planets orbit stars? Because the shock of the collapse has ionized some of the surrounding gas, into positive protons and negative electrons. This ionized state makes it an electrically-charged plasma, generating electromagnetic fields and currents that cause turbulence. As the American astrophysicist Steven Balbus and his colleague John Hawley worked out in 1991, this turbulence acts as a remarkably efficient magnetic brake, transferring the plasma’s angular momentum outward, allowing the plasma itself to move rapidly inward, closer and hotter, forming a ferociously hot accretion disc – a blazing donut – locked tightly around the black hole. Some gas falls into the small mouth of the (dense, but tiny) black hole, bulking it up, but most of the gas just spirals closer for now, getting ludicrously hot from magnetic reconnection events (rather than mere particle-to-particle friction). Basically, a huge, hot, increasingly frantic queue builds up outside the exclusive, narrow, VIP entrance to the black hole.

“And the amazing thing is that even a very weak magnetic field would be enough to completely disrupt the stability properties of the gas. And that’s what a lot of people had a hard time getting their head around.”

– Steven Balbus, co-discoverer of the magnetorotational instability. (Also known as the Balbus–Hawley instability.)

(Yes, yet another example of the Fred Hoyle / Dan Shechtman phenomenon. Evolved systems turn out to be extremely fine-tuned, at the level of the basic parameters of matter, so as to enable unlikely outcomes that lead to complex processes and structures…)

THE DYNAMOS POWER UP

Meanwhile, conservation of angular momentum means all the rotational energy of the initial vast, smooth cloud of gas is now concentrated into the much smaller black hole. (It’s supermassive, but it’s not big; millions of times the mass of our sun, all now crammed into a space far smaller than our solar system.)

This concentration forces the supermassive black hole to spin ridiculously rapidly, at close to its absolute maximum theoretical limit (so, pretty close to the speed of light), at the core of a hot, dense cloud of charged particles, the closest of which are now also spinning at nearly lightspeed. The whole system becomes a colossal dynamo, generating an absurdly powerful magnetic field from the rotating plasma.

This magnetic field now blasts two jets of hyper-energetic charged particles from the black hole’s poles: one north, one south. The dynamo electromagnetically accelerates these particles to almost light speed. (There goes all that shed angular momentum!) As they move this fast, you see relativistic effects – time dilates, particle masses increase relative to some guy watching from far away, all the fun stuff predicted by Einstein’s General Relativity. (Hence, “relativistic jets.”)

The jet is also collimated, meaning it stays coherent and narrow as it beams into the surrounding gas. They still can’t fully EXPLAIN how it remains so tightly collimated over such long distances; theory says Kelvin-Helmholtz instabilities – turbulence, basically – should cause it to break up. But the further away in space they look (and thus the further back in time), the more the mainstream theory fails: Porphyrion, the 23-million-light-years-long jet described recently in Nature is, after all, 40% longer than theory said was possible. Yet again, fine-tuning means that highly unlikely orderly structure emerges from what a naive, everything-is-arbitrary, approach assumes should be random chaos. (Detect a theme?)

So, what does this tightly-focused, high-speed jet of charged particles do? A lot. It first sheds angular momentum from the hot gas “donut” around the black hole, stopping the donut from disintegrating and allowing more gas to fall in. But I’ll focus on the three most important later structural impacts of these jets: galaxy formation, void formation, and magnetic field formation (in galaxies, filaments, and voids).

1.) Galaxy Formation

Immediately surrounding the supermassive black hole is a hot, tight, fast-spinning donut of ionized gas. Further out lies a much larger, cooler, slower sphere of smooth gas, which has been drawn in by gravity. As it moves closer, this gas speeds up (thanks to conservation of angular momentum) and begins to flatten out into a much thinner, denser, disc. (Yeah, Zeldovich’s dear old pancake again.) The relativistic jet from the black hole blasts through this disc like a bullet through a water balloon, sending out pressure waves that disrupt the smoothness of the gas, causing pockets of density – and thus star formation. This cascade of star formation builds a tight, compact galaxy around the supermassive black hole.

I predicted this process in more detail two years ago, before the James Webb Space Telescope released its first data. (And I was right.) I won’t recap here, as I want to focus on the other two important structural consequences: void formation and magnetic fields.

2.) Void Formation

Many of these early relativistic jets just keep on going, reaching distances ten times, a hundred times, longer than the width of the galaxy forming at their base. And remember, I’m arguing that there’s very roughly a trillion of these jets, in assorted sizes, nicely spaced out. As they streak through the pristine gas of the early universe at close to light speed, the jets generate massive shockwaves that push matter out of 80% of the universe.

This is not fanciful. We know that, when fed gas, supermassive black holes fire jets from both magnetic poles. We know that, over time, these jets form huge, low-pressure cavities in the gas north and south of their galaxy. We see evidence of such lobes today – like the Fermi bubbles (which we only discovered in 2010), north and south of our own Milky Way galaxy, probably caused recently by a brief firing of jets from our galaxy’s own supermassive black hole (as some gas, or a star, fell in).

We know that older cavities tend to be larger, because they formed when gas in the vicinity of the supermassive black holes was denser and more plentiful, providing the fuel for longer, more powerful jets. (For example, a marvellous 2024 paper by Francesco Ubertosi, Simona Giacintucci, Tracy Clarke et al, on galaxy cluster Abell 496, shows cavities within cavities within cavities, with the oldest being the largest.)

My prediction here is that the dense gas of the early universe feeds sustained, powerful jets, which generate vast, early cavities that continue expanding, eventually detaching from their jet source (which eventually runs out of fuel, and switches off). These cavities, formed by an electromagnetic jet, will have (I’m arguing) a weak but effective electromagnetic boundary – rather like a soap bubble’s skin – so when they encounter other cavities, their electromagnetic skins could simply merge, forming larger voids. The universe’s expansion should then hugely expand them over time.

Blowtorch theory argues that is what generated today’s voids, with their blurry balloon-animal shapes.

A trillion dynamos, switching on a trillion blowtorches, early.

In pushing gas out of 80% of the universe, these jets are building dense gas reservoirs in the remaining 20%: the narrow spaces, or filaments, between voids. Reservoirs the galaxies will draw on later, using the third thing the electromagnetic jets are building. Something they are particularly suited to…

3.) Electromagnetic Fields! (and thus Filament Formation!)

The electromagnetic jets in this high-powered, early era – jets sometimes ten, fifty, even a hundred times longer than the galaxy at their base – are also laying down an electromagnetic circulatory system. This is important: Lay down magnetic field lines early enough, and you constrain the future development of the entire universe.

MAGNETIC FIELDS CONSTRAIN AND GUIDE IONISED GAS

Here’s why. Magnetic fields constrain and guide ionized gas (plasma). Within a magnetic field, plasma can’t move freely in all directions: it moves much more easily along magnetic field lines than across them. Processes like heat conduction, viscosity, and diffusion also become highly directional in a magnetic field. So, movement, heat distribution, energy flow – everything in a magnetic field follows the path of least resistance along those field lines.

In this way, magnetic fields act like tram tracks for charged particles (or arteries, if you’d prefer a 3D analogy), organizing the energy flow, and mass flow, that structures the system.

“Anytime astronomers figure out a new way of looking for magnetic fields in ever more remote regions of the cosmos, inexplicably, they find them.”

— Natalie Wolchover, Quanta Magazine, from The Hidden Magnetic Universe Begins to Come Into View, July 2020.

True, though I’d remove the word “inexplicably.”

EXPLAINING MAGNETIC FIELDS IN VOIDS

We recently discovered that even voids contain faint electromagnetic fields. Currently there is no generally accepted theory for how those fields got there. Blowtorch theory argues that they were left by the jets that sculpted the voids.

EXPLAINING MAGNETIC FIELDS IN FILAMENTS

The magnetic fields in the filaments connecting denser regions? Those are the electromagnetic ghosts of the original ultra-powerful, ultra-long jets. Basically, electromagnetic pipes. Their field lines link back to the centers of the galaxies that produced them, establishing a cosmic plumbing system that will, over time, drip-feed back to those galaxies the gas they previously pushed away into dense reservoirs.

EXPLAINING THE ABRUPTNESS OF THE DENSITY GRADIENTS

This also explains the abrupt tenfold change in density between filament and void. (Which is a major problem for Lambda Cold Dark Matter.) Without electromagnetic containment, gravity alone should, over time, simply blur out this abrupt density transition. Gravity drops off smoothly with distance, not abruptly, by an order of magnitude, at a specific point. Similarly, gas, even in open space, should slowly diffuse from high-pressure areas to low. Yet that, too, does not seem to be happening to the extent that a purely gravitational theory predicts. The maintenance of such hard boundaries over billions of years makes perfect sense, electromagnetically: it doesn’t make any sense gravitationally (with or without dark matter).

However, I don’t want to give you the impression this whole process is ultra-efficient, with totally clearcut boundaries. The formation of galaxies, voids, and filaments is messy (as processes evolved through blind Darwinian evolution tend to be); galaxies crash into each other, some gas gets stranded, filaments combine, and tangle. It’s more like Darwin’s entangled bank than a smoothly operating machine.

But that’s fine, because, from evolution’s point of view (and I’ll expand on this shortly), the important thing is to get the gas moving now, and to lay down magnetic field lines that can capture it and direct that movement. Galaxies are hurriedly laying cable, laying pipes, building reservoirs and filling them, while everything is still close together; while it is still easy. (Just as an embryo grows its arteries and veins while still in utero.)

Once in place, these magnetic field lines essentially “freeze” into the plasma. As the universe expands, the field lines thus stretch and expand along with the plasma itself (albeit weakening due to flux conservation).

SOPHISTICATED, POWERFUL, EARLY, ELECTROMAGNETIC STRUCTURE FORMATION

The key point: Early structure formation in our universe is primarily electromagnetic, not gravitational. Gravity does the grunt work of pulling (particularly pulling together the clusters and superclusters), but electromagnetism does the sophisticated work of shaping and directing.

And, note, this model needs only baryonic matter – the stuff we can see. No dark matter required. (You can leave it in the picture if you wish, but why? It isn’t necessary.)

There will be a couple of obvious questions, so let me answer them here.

Q: But the voids we see today are up to a hundred million light-years wide, or even larger! And the jets we see today are far smaller than that, peaking at a few million light-years in length! So how can such small jets make such large voids?

A: Because those voids were made early, when the universe was still extremely dense and compact. At redshift 20 – roughly 180 million years after the Big Bang – our universe was 20 times smaller in diameter than it is now, with gas therefore nearly 9,300 times denser. (The average density of matter in the universe scales as the cube of the inverse of the scale factor – or, less technically, if you shrink the universe in all three dimensions, you squeeze the living fuck out of the matter.)

By redshift 10 – or 500 million years after the Big Bang – the universe was still 10 times smaller, and gas was 1,300 times denser than today. So those early jets didn’t need to travel vast distances to have massive effects. Jets were pushing between a thousand and ten thousand times as much gas out of the way as they would over a similar distance today. And, over time, the universe’s expansion would make these early voids ten or twenty times longer, and wider, and taller – and thus thousands of times larger in volume – than they were at redshift 10 or 20… but they began relatively compact.

Moreover, there was much more, and much denser, gas in close proximity to supermassive black holes in the early universe. Enough available gas to feed their dynamos, and thus sustain long, powerful jets, over extended periods – tens or even hundreds of millions of years; far longer than what we see today in our vastly expanded, far more diffuse cosmos.

As in most evolved systems, key structures form remarkably early, while it’s still easy. A human embryo, for example, has decided which end will eventually be head, and which feet, by day five, when it’s a mere 64-cell blob, less than 0.2 millimetres in diameter. Likewise with our universe.

Q: But if the jets from the core of the new galaxy push gas out of the voids, why don’t they also push gas out of the galaxy, switching off star formation?

A: Because this is a complex, evolved system doing a complex, evolved job. The same dynamo effect that drives gas away from the galaxy’s poles (perpendicular to its disk) is simultaneously, by removing angular momentum from the gas at its equator, allowing that gas to spiral rapidly inward, forming new stars and feeding the jet. Gravity, electromagnetic fields, and pressure gradients combine to create a vast cosmic pump, steadily drawing in fresh gas. This pump can operate continuously for tens or even hundreds of millions of years.

Currently, the mainstream theory is, indeed, that jets quench star formation by driving gas out of the galaxy. But that view came from observing mature, relatively nearby galaxies, many billions of years after the Big Bang, where these pumps are older, erratic, running out of fuel, and beginning to switch off, coinciding with the end of rapid star formation. Correlation; not causation. Meanwhile, pre-James Webb Space Telescope, we had no data on the early, high-power phase.

What I predicted back in 2022, and what the James Webb is starting to see strong evidence for, is that, early on, these jets do the opposite – they’re driving star formation, not quenching it.

Q: You say Blowtorch Theory shows we don’t need dark matter. But isn’t dark matter required to explain galactic rotation, and disc galaxy stability, and gravitational lensing, and the galactic dynamics of dwarf galaxies, and acoustic peaks in the Cosmic Microwave Background, and the Universal Rotation Curve, and…

A: Currently, yes, simply because every outstanding observational problem has dark matter thrown at it, as the only possible explanation. This is partly because dark matter is so fuzzily defined, with so many free parameters, you can massage it to “solve” almost any problem. (Why did your grandmother slowly but steadily move UP the stairs, in defiance of the laws of gravity? Because there was a large halo of dark matter at the top of the stairs, pulling her. See? Problem solved. No need, now, to look for a more complex, evolutionary, dynamic-systems explanation.)

I’m not arguing that blowtorch theory also solves all the other outstanding problems currently explained by dark matter. It doesn’t. I’m making a broader point: that an evolved-universe approach will generalise to solve the other outstanding anomalies. They may well all have quite different explanations. But, as with structure formation, the answers are likely to lie in the sophisticated behaviour of the things we can see, not the simplistic behaviour of a thing we can’t. By the way, this makes Blowtorch Theory an excellent fit with David Wiltshire’s Timescape cosmology. (See his important recent paper, “Solution to the cosmological constant problem”.) Timescape deals with Dark Energy’s contribution to structure formation; Blowtorch Theory deals with Dark Matter’s contribution to structure formation. Indeed, it is quite likely that, as we find better, more sophisticated explanations for each of these different anomalies, dark matter will slowly fade away, without any big “eureka!” moment. (Or whatever the negative of “eureka!” is.)

DARK MATTER IS THE MODERN MIASMA

This is a classic example of the Single Variable Fallacy, or maybe better The One-Missing-Piece Illusion. A very human cognitive shortcut that assumes there must be a single missing factor that can explain all discrepancies. We made this mistake before, when we blamed all disease on “miasma” – also known as “bad air” and “night air”. “Miasma” almost-but-not-quite explained everything. (The Black Death! Cholera! Chlamydia! Malaria, whose name literally means bad air in Medieval Italian! Oh, and obesity was caused by inhaling the odor of food…)

“Miasma” was the dominant theory of disease, worldwide, for two thousand years, until the late 19^th century. But the actual answer to “what causes disease?” involved bacteria, and viruses; airborne transmission, and water transmission; vitamin deficiencies, and food spoilage; some bugs growing in the presence of oxygen, some in the absence of oxygen; mosquitos, and fleas; immune system responses, including both under-reactions and over-reactions – and on and on… In other words, we hadn’t remotely explored the behavioural possibility space of the air and water and food we could see. A special, extra kind of air to one-shot all the problems was not required.

Likewise, today, the idea that we have completely explored the possibility space occupied by baryonic matter, and must instead postulate an entirely new, and totally invisible, form of matter to explain baryonic matter’s behaviour… you can see the absurdity, right? You can see that we are making precisely the same mistake?

Evolved biological systems are complex, and self-organising. Their behaviours feature emergent properties, feedback loops, and interactions across multiple scales. If our universe is itself an evolved system, then it’s unsurprising that we’re making exactly the same mistake, trying to explain its complex emergent properties using linear causal thinking. Evolved systems resist single-cause explanations. Their behaviors emerge and adjust as the rolling result of weird, unpredictable, interacting networks.

Right now, cosmology suffers from a lethal combination of factors – it has baked ΛCDM into every simulation; it uses it as a background assumption that now underlies, and distorts, every paper in astronomy; and it insists that any alternative solution explain every outstanding observational anomaly in one go, single puzzle-piece style – or if you prefer, miasma-style. This fatal triple-combo has crippled the field’s ability to self-correct.

And again, let me emphasise: whether or not dark matter exists, blowtorch theory clearly explains many phenomena, in the specific area of early structure formation, which that theory cannot.

WHAT’S NEW, PUSSYCAT?

So – for those unfamiliar with the field – what’s novel, and what’s not, in this theory?

Not novel: Supermassive black holes exist at the centers of galaxies, some with accretion disks that emit relativistic jets. That’s now mainstream cosmology. Active Galactic Nuclei (AGNs), quasars, blazars, Seyfert Galaxies, Fanaroff-Riley Radio Galaxies – all these powerful, inexplicable phenomena eventually turned out to be supermassive black holes, doing things with jets. (Hey, it’s almost as if these black holes, and their jets, might be ubiquitous and important!)

What’s novel in my theory is the idea that all the supermassive black holes must form first, by direct collapse – before galaxies form, and indeed before there’s any significant number of stars, or (probably) any stars at all. (This emerges directly from the application of Darwinian evolutionary logic to universes. It’s not predicted by any other theory, and if I’m wrong, my theory wobbles badly and a wheel falls off. So the theory is falsifiable. But the evidence from the James Webb Space Telescope so far is very, very encouraging.) Yes, this phase transition occurs incredibly early and activates many powerful, sustained, directional jets very strongly, very quickly. But – no new physics are needed to explain it. In fact, Priya Natarajan, Volker Bromm, Marta Volonteri, and others showed 18 years ago that direct-collapse supermassive black holes are mathematically possible, in brilliant work that was largely ignored at the time. (They, too, will get a late, apologetic Nobel Prize.)

So that’s blowtorch theory. But it emerges from – it is both predicted by, and explained by – a larger theory about the universe: three-stage cosmological natural selection. Yes, a theory in desperate need of a shorter, slangier, more vivid, less abstract name. (I find myself calling it the Eggiverse theory – in opposition to the mainstream’s Rockiverse – because it describes our evolved universe rapidly developing upward in complexity over time, like an egg, rather than simply disintegrating slowly, like a rock.)

And three-stage cosmological natural selection – Eggiverse theory – builds on Lee Smolin’s original version of cosmological natural selection.

BACKGROUND: LEE SMOLIN’S ORIGINAL THEORY OF COSMOLOGICAL NATURAL SELECTION (DRAWING ON JOHN WHEELER)

A quick recap for new readers…

Cosmological natural selection assumes that universes reproduce via black holes and big bangs. How? A large mass in a parent universe collapses under its own gravity to form a black hole singularity – an infinitesimal point – that then “bounces”, in a Big Bang, to form a new, expanding, separate bubble of spacetime: a child universe, that exists outside of, and separate from, the parent universe. (The expansion of our own universe since the Big Bang is simply the growth and development to maturity of such a child universe.) If each new universe varies slightly in the fundamental parameters of matter (things like the mass of the electron, or the strength of the strong nuclear force), this will lead to that universe producing either more or fewer black holes, and thus to greater or lesser reproductive success. Over time, this reproduction-with-variation-and-inheritance leads to a Darwinian evolution of universes.

There are analogies with biological evolution: changes in the basic parameters of matter between parent and child universe act like the changes in DNA between biological parents and children: any such variations, or mutations, in the basic parameters of matter will lead to changes in the phenotype of the child universe. (Let’s call it the cosmotype.)

An evolved universe, therefore, constructs itself according to an internal, evolved set of rules baked deep into its matter, just as a baby, or a sprouting acorn, does.

The development of our specific universe, therefore, since its birth in the Big Bang, mirrors the development of an organism; both are complex evolved systems, where (to quote the splendid Viscount Ilya Romanovich Prigogine), the energy that moves through the system organises the system.

But universes have an interesting reproductive advantage over, say, animals.

We know that in our own universe, the (positive) mass energy of matter and its (negative) gravitational energy net out to zero. But that means the amount of energy required to build a universe like ours is essentially… well, none. (See Laurence Krauss’s book, A Universe from Nothing; or Stephen Hawking and James Hartle’s 1983 paper, The Wave Function of the Universe, and Hawking’s later book with Leonard Mlodinow, The Grand Design, etc.) Child universes are therefore effectively free to produce. Thus, over time, evolution will favor making more (and smaller) black holes from the same amount of matter. Yes, the black holes will get smaller – but the universes they give birth to, through Big Bangs, will still be full sized. And universes aren’t constrained by a shared environment with limited resources – newborn universes aren’t all competing in a valley with a limited amount of grass. Indeed, each new universe is entirely self-contained and self-sufficient, being both organism and environment – it supplies its own energy for its own development, efficiently and frugally, through stellar fusion, gravitational collapse, et cetera. This means evolution should ultimately favour runaway black hole production. (And that checks out: in the most recent estimate, in 2022, by a team from the International School of Advanced Studies in Trieste, our universe was estimated to have already generated up to a trillion supermassive black holes – and 40 quintillion stellar-collapse black holes.)

A key implication of all these assumptions (which I first pointed out on July 8^th 2022 – this theory is undergoing rapid development!), is that direct-collapse supermassive black holes must have been the reproductive mechanism for the earliest, most primitive universes. Primitive universes, like early prokaryotic bacteria, reproduced and nothing else – no complex structures, just straightforward reproduction.

This is a simple black hole/big bang reproductive cycle. Or, if you like, a (contracting) black hole to (expanding) white hole cycle.

Therefore, the mass-energy – the primitive matter – expanding from a Big Bang in an early universe would have quickly collapsed back into a small number of extremely large black holes (supermassive or even ultramassive), without forming stars, galaxies, planets, or even complex elements. Those things all evolved much later along the evolutionary timeline that eventually led to our universe.

Such a basic, ancient reproductive strategy would be conserved by evolution, for the obvious reason that any universes that failed to do this – reproduce – would, in evolutionary terms, die out. (That is, such sterile universes would continue, individually, to exist, growing endlessly older inside their own individual bubble of space-time, but would have no offspring.) Similarly, humans, although far more complex than our jawless fish ancestors, still reproduce as they originally did, by fertilizing an egg – a fundamental reproductive strategy necessarily conserved by evolution.

These conserved, direct collapse, supermassive black holes are, in several important ways, nothing like the, far more common, stellar-mass black holes we also find in our universe (which form when a large star runs out of fuel and collapses under its own gravity). Supermassive black holes are hundreds of thousands, millions, or even billions of times more massive than stellar-mass black holes; I argue that they must have a completely different formation mechanism (direct collapse); and, as we have seen, they perform totally different, and vital, functions in structuring the early universe.

Clearly seeing these differences is vital to understanding our universe.

Yet, for decades, mainstream cosmology (trapped inside Lambda Cold Dark Matter, a paradigm that only allowed for bottom-up structure formation) simply ignored the possibility of direct-collapse supermassive black holes. Instead they argued that the supermassive black holes we kept finding at the centre of galaxies were just a bunch of stellar-mass black holes in a trenchcoat. That the first stars must have been large; formed large stellar-collapse black holes; these quickly merged, and merged again, and again, then happened to drift to the galactic centre thanks to gravity, and then drank in huge amounts of gas. (Yeah, more random, arbitrary, bottom-up structure formation.) Voilà, an accidental, late-to-form, supermassive black hole…

But, as our ever-improving telescopes get us more and more data, from earlier and earlier, it gets harder and harder to make the maths add up. We are now able to observe supermassive black holes shockingly close to the Big Bang, when the universe was just 3% of its current age. There simply isn’t enough time for them to have formed from lots of smaller stellar-mass black hole mergers.

And so the supermassive black holes in our universe today presumably formed the same way as in our distant ancestral universes: by direct collapse of a massive, smooth gas cloud – without needing to build stars first – very soon after the Big Bang.

But why isn’t our universe still fine-tuned by evolution to turn all its gas into supermassive black holes immediately after the Big Bang? Because, over time, it evolved a new, sophisticated, multi-stage reproductive cycle that takes billions of years. There are many analogies for this in biology; we know this is what evolution does. Primitive bacteria, for example, reproduce a single all-purpose cell in 20 minutes; complex beings like humans reproduce trillions of specialised cells, in a more complex process that takes decades. Both are successful, but in their own evolutionary contexts.

So here, the unit of selection isn’t a single galaxy; it’s the entire universe – a trillion galaxies contained within a single, expanding membrane of spacetime, like the trillions of cells contained by a human’s skin. And what matters is reproduction (black hole production) over its full lifespan.

THAT THIRTEEN-SYLLABLE THEORY AGAIN

TIMELINE: To fully understand this, you’ll need the three-stage model of cosmological natural selection. (Good old Eggiverse theory.)

• 1990s: Smolin’s original theory was a limited, single-stage model because, back then, it was still assumed all black holes, whatever their size, were made out of stellar-collapse black holes.

• 2000s: Vidal, Smart, Price and Kane expanded this to a two-stage model, adding technologically-produced small black holes. (This explained why evolved universes might generate life – a brilliant conceptual breakthrough, again overlooked at the time.)

• 2020s: I develop the three-stage model, incorporating direct-collapse supermassive black holes – the model which led to blowtorch theory, and which generates the predictions we’re seeing here.

THE THREE-STAGE MODEL OF COSMOLOGICAL NATURAL SELECTION

STAGE 1: DIRECT COLLAPSE SUPERMASSIVE BLACK HOLES

Right at the start, in a phase transition shortly after the big bang, a relatively small number of supermassive black holes quickly form by direct collapse. (Very roughly a trillion, of assorted sizes.)

That is the original reproductive mechanism for universes, conserved by evolution.

STAGE 2: STELLAR-MASS BLACK HOLES

However, these supermassive black holes only use up a fraction of the gas in the universe. They then shape and direct much of the remaining gas so as to form galaxies around themselves – one galaxy (containing hundreds of millions to hundreds of billions of stars) per supermassive black hole. These galaxies thus eventually generate many more stellar-mass black holes than the initial set of supermassive ones. (Thus, overall, far more black holes – offspring – per unit mass.) This more complex, highly efficient method of reproduction evolved later in the history of universes and has likewise been conserved, because it leads to…

STAGE 3: TECHNOLOGICALLY PRODUCED SMALL BLACK HOLES

In a more sophisticated stage 3 universe, after two or three rounds of star formation, galaxies have built out (through fusion) and distributed (through supernova explosions) the periodic table of all the elements.

This allows planets and moons to form around third-round stars like our own sun. (And stanets and ploons to form in open space… but that’s another story.) Sunshine and/or gravitational energy then drive the development of the complex organic chemistry of biology: Biology is the first half of the periodic table coming to life. Eventually, that life, rapidly self-complexifying on countless billions of habitable worlds per galaxy, attains intelligence and technological ability. (It has swiftly done so on our own perfectly average planet – so we know this happens.) Technology is the second half of the periodic table coming to life. At which point, those intelligent, technology-wielding lifeforms begin creating vast quantities of extremely small black holes for energy production.

Biology is the first half of the periodic table coming to life.

Technology is the second half of the periodic table coming to life.

-Me (trying to boil this whole theory down into t-shirt memes)

Q: Why black holes?

A: Because you can just chuck any matter at all into them, and they can convert up to 42% of its mass into energy, making them the most efficient energy source in our universe. (And, as my friend Yogi has pointed out, the most sustainable – you could keep a civilization and its ecosystem alive by this method long after all the uranium’s been used up; indeed, long after all the stars have gone out.) By comparison, nuclear fission converts only 0.1% of mass into energy, and fusion 0.7%. So, any intelligent species optimizing for energy efficiency (and sustainability) will eventually converge on small black hole production. (Optimal size is roughly Mount-Everest-mass.) As the production of countless small black holes by technology-wielding lifeforms means huge reproductive success for that universe, such life will be intensely selected for, and strongly conserved. Most universes along our evolutionary branch should, by now, generate such lifeforms. Again, that’s because such Stage 3 universes will produce, over their lifetime, far more black holes per unit mass than will either Stage 1 or Stage 2 universes.

HOW THE BLOWTORCH THEORY MAKES EVOLUTIONARY SENSE

This third reproductive mechanism is far more complex than the first two (it’s more highly evolved!), and it requires the individual universe to have a much longer active lifespan. In particular, it needs multiple rounds of star formation to take place within the individual universe, over billions of years, to build out and distribute the entire periodic table – and thus planets, moons, life, and eventually, technology.

If our universe were to use up all its gas early, just to make a few more first-round, low-metallicity stars and simple stellar-mass black holes, we’d never reach this third and most reproductively successful stage.

So that pure hydrogen gas left over from the Big Bang has to be carefully rationed and delivered to spiral galaxies over billions of years. But it also needs to be enriched – because you can’t easily make stars out of pure hydrogen and helium alone. You need heavier elements – rather annoyingly dubbed “metals” by astronomers (even when, like nitrogen and oxygen, they’re not metals). In spiral galaxies, first-round stars blow up and distribute these “metals” back into the pure hydrogen that’s streaming in from the filaments. That enriched gas then gathers in the specialised regions in the arms of spiral galaxies we call “stellar nurseries”, ready for the next round of star formation. (Stellar nurseries… Interesting name, huh? I suspect that, at a deep subconscious level, many astronomers already know this is an efficient, evolved, quasi-biological process.)

So spiral galaxies can store, channel, enrich, and drip-feed this primal hydrogen gas back to their stellar nurseries, where carbon and oxygen from past supernovae help refrigerate it (and enrich it), allowing it to collapse to form next-round stars. This absurdly intricate process is driven by the delicately balanced interplay of gravitational and electromagnetic forces. It’s clearly a fine-tuned system, evolved for optimal star-planet-moon-life-technology creation. Why “clearly fine-tuned?” Because such a result is unbelievably unlikely – yet, look around you, that’s what it does. (Just as Fred Hoyle predicted that the fusion process in stars was fine-tuned to produce and distribute the elements – because look around you, that’s what it does.)

EXCEPTIONS PROVE RULES: ELLIPTICAL GALAXIES

But let’s take a quick look at elliptical galaxies. These are 10-15% of all galaxies, with roughly spherical shapes, intense but early star formation – and long-term struggles to replenish their gas. Lacking the intricate gas management system of spiral galaxies, ellipticals burn out quickly, typically failing to form the large amounts of the full suite of elements essential for planets, life, and tech. Ellipticals make SOME elements, and distribute them, obviously, as their larger stars run out of fuel and blow up as supernovae – but they don’t go through multiple rounds of star formation, over billions of years, methodically building out ever-greater amounts of heavier elements by recycling and refining them, in the way spirals do. And yes, some ellipticals are still active, as other galaxies crash into them, bringing fresh gas – but most are red and dead.

So what’s the story with ellipticals? Well, let me speculate: if our universe is the result of the evolutionary process I outlined above, then they’re probably the vestiges of an earlier cosmic age, like our own vestigial tailbone, or the fur in our armpits. All galaxies would’ve been ellipticals (or similar) at Stage Two, when only rapid star-making mattered. (Ellipticals are brilliant at extremely rapid star-making!)

EVOLUTIONARY TRACES: TAU, MUON, ELECTRON

Evolution moves on, but it drags elements of its evolutionary past around with it, it can’t simply make a clean break. Similarly, the three-stage model of cosmological natural selection explains why three kinds of electron can be produced in our universe – but only one usually is, under normal conditions. Extremely heavy tau electrons (roughly 3,500 times the mass of a normal electron), and heavy muons (roughly 200 times heavier), are presumably fossil particles from stage one and stage two respectively. (Interestingly, the Nobel-Prize-winning Japanese physicist Yoichiro Nambu intuited this as far back as 1985, in his wonderful essay Directions of Particle Physics, but had no theory to explain it.) Just to make it clear, this is much more speculative than the theory it’s based on. But it’s a speculation I find intriguing, and you might find interesting.

• Evolutionary Stage 1: An extremely heavy electron (the tau) made evolutionary sense when all you had to do was maximise immediate direct-collapse black holes. Mass was king! The positive and negative particles probably had roughly the same mass back then, but it didn’t matter; they weren’t building anything complicated.

• Evolutionary Stage 2: A less heavy (but still pretty heavy!) electron (the muon) would have been ideal if you still needed to make some supermassive black holes, but were now optimising for, more numerous, stellar-collapse black holes. You only needed to be able to make a couple of primitive elements, through primitive fusion.

• And Evolutionary Stage 3: Our modern, lightweight electron was the evolutionary breakthrough that allowed the assembly, by fusion, of many more (more complex) elements, and thus complex chemistry, life, and technology – optimising for FAR more numerous technologically-produced black holes…

All universes, like all biological DNA-based organisms, are forever a little messy; forever transitional between what they once were, and what they might yet be.

For example, the carbon, nitrogen and oxygen that originally evolved to drive efficient stellar fusion in Stage 2 universes, through the CNO cycle, were later exapted (adapted for a new purpose by evolution) as the basis for biological life in Stage 3 universes.

Likewise, just look through a telescope, and you can see handfuls of the red-and-dead elliptical galaxies that must have dominated those Stage 2 universes.

Fire up a particle accelerator, and you can bring to life ghost particles like the tau and muon, still implicit in our quantum fields, even though they have long been banished from our everyday world by the ongoing evolution of the other parameters of matter.

Those traces of our evolutionary past, embedded in our present, mean that we can trace the evolutionary history of universes, even though we only have a single specimen to examine – the universe we are embedded inside, that we are part of. The universe that recently generated us, as part of its developmental process. The universe that is coming to know itself through us; act on itself through us. The universe in which we are currently helping sand to think.

Okay. As Tyler Cowen gently pointed out to me, when I showed him an earlier draft of this, a good theory makes predictions.

So here are some.

PREDICTIONS

Supermassive black holes form first, in a phase change that precedes stars and galaxies

We will see a wave of direct collapse supermassive black hole formation (numbering a trillion or so), well inside the first 100 million years after the Big Bang. (And almost certainly inside the first fifty million.) This wave precedes star and galaxy formation.

Those early supermassive black holes will be found to spin at close to the Kerr limit

As all the angular momentum in the entire vast cloud is conserved, the spin rate of these direct collapse supermassive black holes will approach the Kerr black hole spin limit (very close to the speed of light). Later mergers, plus infills of randomly oriented gas and stars, etc, will often lower that spin rate, making them less efficient: but they are born at close to maximum efficiency.

Those extremely fast-spinning, efficient supermassive black holes then generate stars and galaxies

It is the supermassive black holes which generate the galaxies around themselves, by attracting, shocking, and enriching the surrounding gas to precipitate waves of star formation.

Meanwhile, they generate the cosmic web

A trillion quasars switch on immediately, blasting out powerful, sustained relativistic jets, which create both the low-pressure, lightly-magnetized cavities and the high-pressure, highly-magnetized pipes that merge and expand to form the voids and filaments of the cosmic web.

But it’s simply hard to see anything at all from the first fifty million years after the Big Bang, let alone the detailed picture I’ve just described – even with the James Webb Space Telescope. So here are some specifics we can look for, using a variety of methods, some available right now:

Gravitational waves: The Popcorn Signature

Gravitational waves are ripples in spacetime, caused by large masses moving asymmetrically. As our gravitational wave detectors (like the wonderful Pulsar Timing Arrays I wrote about here) grow more sensitive, and able to locate events further and further back in space and time, we will find the gravitational-wave traces of that original brief era of direct collapse supermassive black hole formation coming from all over the sky, in lots of faint, overlapping, low-frequency gravitational waves. It’s true that any initial perfectly symmetrical collapses wouldn’t generate waves (no rough edges = no splash = no ripple), but as the smoothness breaks down, the later, slightly less symmetrical collapses should, as should any early mergers.

Being a phase transition, it should take place over a relatively tight period of a few million years (and probably at some point inside the first 50 million years). As such, the overall signal (the combination of all the signals) should have a steep attack; a short, extremely intense peak; and a slightly longer, less steep decay: Pop… pop… poppopPOPPOPPOPpop… pop… pop… … … pop. Rather like the production of a batch of popcorn. Let’s call that the Popcorn Signature: The gravitational wave counterpart to the Cosmic Microwave Background.

Speaking of which, I also predict…

(Takes a deep breath….)

A statistical relationship between cosmic microwave background quantum fluctuations and supermassive black holes

If fluctuations indeed seed direct collapses, there should be a statistical relationship between the range of sizes and relative numbers of the blown-up quantum fluctuations we see in the cosmic microwave background, and the range of sizes and relative numbers of the supermassive black holes we see in the very early universe. (Before any later mergers distort the relationship.) Also…

Signatures of chain-reaction collapses

If the direct collapse of one region of gas can compress neighboring regions into collapse (similar to the raindrop-cascades in clouds here on Earth), we might be able to detect “clustered timing” effects: bursts of supermassive black hole formation in short intervals, slightly offset in space and time as shockwaves push adjacent regions over the collapse threshold. Basically, look for non-random clustering of the earliest quasar ignition, in redshift slices. (“Collapse cascades.”) And…

Extremely early, rapid, curved starbursts triggered by jet-driven shocks

Look for shock-geometry features in extremely early high-redshift galaxies: arcs, circles, or cones of newly formed stars, in rings that are centred on the active, or recently active, supermassive black hole. There will be more of this the earlier you get. (Later, once the jets switch off, ongoing star formation will occur more frugally in the spiral arms, as the full gas circulatory system establishes itself.)

OK, let’s move on to filaments and voids.

Uniform field orientation along filaments

If powerful jets – rather than slow, random, gravitational infall – formed filaments, then we should see highly correlated magnetic field orientations over incredibly long stretches of filament. Not the arbitrary, patchy, turbulent, broken-jig-saw-puzzle fields of a filament randomly assembled by gravity.

Rifle barrel filaments

Let me add a twist to that. Additionally, the rapidly spinning central supermassive black hole – because it causes spacetime to rotate at speeds close to the speed of light near the event horizon – should impart a continuous, ongoing twist to the magnetic fieldlines of the jet as it accelerates away. (As first suggested by Roger Blandford and Roman Znajek in 1977, and since confirmed, in 2010, by Keiichi Asada et al, in a very smart observational paper.) This should result in a helical, or spiral, magnetic field, which stabilises the jet of plasma, leading to less turbulence. (Similarly, the spiral groove which lines a rifle barrel causes the bullet to spin, stabilizing it.)

I predict this helical magnetic field – this rifled gun barrel design – will be inherited by the filament, allowing plasma to flow along it more swiftly, and crucially with far less turbulence, than you would expect in a filament randomly assembled by gravity alone. We are already seeing signs that gas is mysteriously low in turbulence while flowing along filaments, only becoming turbulent as it leaves the pipe of the filament and enters the open space of the cluster (the node). See, for a great example, the recent eROSITA X-ray observations of the Abell 3667 cluster of galaxies, where a HUGE filament of gas, feeding into the cluster,​ delivers gas gently, with very low turbulence, until it actually spills out of the filament into the cluster proper, and becomes extremely turbulent.

Blowtorch theory explains why. So, look for helical fields in filaments.

Weak internal fields in voids, surrounded by stronger bubble-like surface fields, with “seams”

A slightly more technical one, but a nice one: If voids are made from the merger (and later expansion) of cavities or bubbles formed early by powerful electromagnetic jets, voids should therefore have weak internal magnetic fields, surrounded by stronger electromagnetic “skins” with bubble-like field topologies. Additionally, we might see thin transition zones, with a different field topology, along the seam where the bubbles meet and merge. (Imagine the lines where the balloon skins meet in a balloon animal.) Very sensitive polarised-light surveys, like the Square Kilometre Array (SKA), might detect abrupt changes in Faraday rotation across these seams, which should show shell-like or patch-boundary field structures. (Faraday rotation is what happens when polarised light passes through a magnetised plasma: those polarised lightwaves get rotated, and the longer their wavelength, the more they get rotated. So light should be rotated differently either side of the seam... and VERY differently by the seam itself.)

Residual shocks along the boundary arcs where bubbles meet and merge might also still be detectable as extremely faint synchrotron or X-ray relics.

Cosmic rays might follow filament field funnels

If filaments are rifled magnetic pipes, they might guide and accelerate some cosmic rays. Future cosmic-ray observatories could look for higher cosmic ray numbers aligned with known filament structures. I predict this will turn out to be the explanation for the unexpected transparency of the universe to very-high-energy (VHE) gamma rays. They are spending much of their journey being accelerated down the barrel of an electromagnetic cannon.

Distinct void density profiles and sharp boundaries

This leads to another prediction: the transition between void interior and filament wall should be more abrupt than the gently sloping, radial density gradient any purely gravitational simulation typically produces. (With or without dark matter.) Right now our instruments do not have the resolution to see such a sharp transition, and are assuming, rather than observing, relatively smooth density gradients at the void/filament boundaries. So, more formally (clears throat): blowtorch theory predicts a near step-function slope over a smaller transition zone than ΛCDM-based void analyses.

Statistical relationship between supermassive black hole size and adjacent void sizes

A rather obvious one, but… Statistically, the larger supermassive black holes should tend to be found at the far ends of the larger voids (along their longer axis), as they generated those voids. (And smaller ones at the far ends of smaller voids, yeah.)

Caution: this won’t be a simple one-to-one mapping. Contemporary voids are generally formed from the earlier mergers of a number of cavities. Also, bear in mind, later mergers of not just voids, but galaxies, and supermassive black holes – plus spacetime expansion, gravity, and kinematics of all kinds – will make these relationships less obvious over time.

OK, that should be enough predictions, even for Tyler.

But I’ll just note that the theory also potentially explains a bunch of other puzzling phenomena, such as the observed large-scale asymmetry in galaxy spin directions, and why quasar polarisation appears to be coherently aligned over billions of lightyears. (Coherent early processes lead to coherent later outcomes! Also, black holes give birth to spinning universes!) Plus it potentially solves the missing satellites problem, the cusp/core problem, the galaxy rotation problem, and the Hubble tension… I am itching to write another page speculating WILDLY BUT BRILLIANTLY on each of these, but this post is already too long. (Shades of Fermat’s “I have a truly marvelous demonstration of this proposition, which this margin is too narrow to contain”, I know, I know – but subscribe, now, for free, and I’ll email you those posts as I write them.)

To sum up:

Lambda Cold Dark Matter was, and is, a brave attempt to solve a lot of difficult problems in one go. Many great people gave it the best years of their lives. (And, to any of you reading: I deeply respect your work, and I totally understand how every decision seemed to make sense at the time, as a will-o-the-wisp led you deeper and deeper into the swamp. I don’t want to fight you, I want to liberate you. Join me.)

But, tragically, after fifty years of development, ΛCDM remains an improvised, incoherent, after-the-fact, ad hoc description of a random bunch of puzzling stuff that, it now turns out, may not even have a common explanation. It’s an infinitely flexible framework, not a predictive theory. When it has made predictions, they have been consistently proved wrong by observation. Have a couple turned out OK? Sure! (I’ll give you the CMB power spectrum, nice one.) But its overall hit rate is still terrible. In general, most of its predictions have been made in haste, with a bit of parameter tweaking, based on a stream of slowly emerging data, rather than genuinely preceding and predicting that data.

ΛCDM has six free parameters – total matter density, dark energy density, Hubble constant, amplitude of matter fluctuations, spectral index of initial fluctuations, and optical depth from re-ionization – and every single one of these has been repeatedly tweaked in the light of emerging new data. That’s fine, that’s what you have to do… but it’s been fifty years now, and the tweaking never stops. There’s simply nothing solid there. Meanwhile, every failure of prediction is immediately rebranded as “an exciting opportunity for new physics!” Ah, new physics. Dark matter, after fifty years, still requires new particles, and their new physics, but can’t seem to catch a glimpse of even one. (Recall that the Large Hadron Collider found the Higgs boson exactly where predicted; but it didn’t find any dark matter, despite exploring all the energy ranges where such stuff was predicted to be.)

By contrast, three-stage cosmological natural selection, and the blowtorch theory that emerges from it, is the coherent description of an equally coherent functional complex system. It has already made extraordinarily successful predictions, since confirmed by observation. Indeed, it has now just made a whole bunch of new predictions, which you can go check on: it is not afraid of being tested, and, unlike ΛCDM, is happy to risk being falsified. It has enormous explanatory power. It requires no new particles and no new physics. Does it need a lot of work, to mathematicize, test, and validate it? Sure! But it’s doing pretty good so far, for a three-year-old theory (based, yes, on an earlier theory that is, itself, only half the age of ΛCDM), developed by a handful of people occasionally talking over coffee, and WhatsApp, with essentially no institutional backing, or funding – particularly when its opponent has had a fifty-year head-start, and all the resources of all the world’s best scientists across every relevant field.

There is a story told by the famous physicist Victor (Viki) Weisskopf, which I would like you to read, and really internalise, really think about, for a minute, particularly if you work in the fields of cosmology, or astrophysics, or astronomy. (By a poignant coincidence, it takes place at the very observatory from which Stephen Gregory and Laird A. Thompson first discovered voids.)

“Several years ago I received an invitation to give a series of lectures at the University of Arizona at Tucson. I was delighted to accept because it would give me a chance to visit the Kitts Peak astronomical observatory, which had a very powerful telescope I had always wanted to look through. I asked my hosts to arrange an evening to visit the observatory so I could look directly at some interesting objects through the telescope. But I was told this would be impossible because the telescope was constantly in use for photography and other research activities. There was no time for simply looking at objects. In that case, I replied, I would not be able to come to deliver my talks. Within days I was informed that everything had been arranged according to my wishes. We drove up the mountain on a wonderfully clear night. The stars and the Milky Way glistened intensely and seemed almost close enough to touch. I entered the cupola and told the technicians who ran the computer-activated telescope that I wanted to see Saturn and a number of the galaxies. It was a great pleasure to observe with my own eyes and with the utmost clarity all the details I had only seen on photographs before. As I looked at all that, I realized that the room had begun to fill with people, and one by one they too peeked into the telescope. I was told that these were astronomers attached to the observatory, but they had never before had the opportunity of looking directly at the objects of their investigations.”

– Victor Weisskopf, from The Joy of Insight.

OK, stop reading for a second, and soak that up.

(If you didn’t stop; seriously, stop, and soak. Maybe even read it again.)

Victor finishes that story with the words,

“I can only hope that this encounter made them realize the importance of such direct contacts.”

Yeah, me too, Victor. Me too.

OK, time to choose.

1.) Dead matter, with random properties, blindly obeying arbitrary laws, in a one-shot universe, that appeared from nowhere, for no reason, 13.8 billion years ago – before which nothing at all existed. Oh yeah, and it just stumbled from being some hot gas to being VERY VERY COMPLICATED INDEED JUST LOOK AT YOUR DOG OR A TREE OR YOUR PHONE completely by accident, with no backstory, no preparation, no explanation; no evolutionary history. Oh, and none of this means anything.

Or…

2.) Sophisticated matter fine-tuned, by a long evolutionary process, to swiftly and efficiently come to life, self-assembling upward into complexity in multiple integrated steps, due to fine-tuned laws; the latest in a long line of ever-more-sophisticated universes, born from a parent universe 13.8 billion years ago, and going somewhere, with us as the growing tip, the point where the universe comes to know itself, take control of its own growth, and direct it out into an infinitely rich possibility space.

Cold dark matter, or evolutionary cosmology?

Rockiverse, or Eggiverse?

Look inside you. Look around you. Don’t think about a theory – theirs or mine. Actually look at the universe of which you are part.

Now choose your fighter…

NOW WHAT?

So, what can you do about this? Well, this is a classic paradigm shift. So what you can do depends on who you are.

If you’re a science journalist: Get in touch. Every article on the early universe, for the past two and a half years, has quoted some scientist saying “Nobody predicted this!” That has to stop. Three-stage cosmological natural selection – dear old Eggiverse theory! – predicted all of this. Let’s get it into the conversation, where it belongs. The reasons for its exclusion have been sociological, not scientific. I realise there are no bad guys here, it’s not a conspiracy. Cosmologists, understandably, don’t have the biology background to apply an evolutionary theory to their work, and evolutionary biologists, understandably, haven’t even heard of a theory at the fringes of cosmology. But that has left it in a knowledge shadow: help me get it out into the light.

If you’re a cosmologist, astrophysicist, or astronomer: I need you. (Also! Evolutionary biologists, chemists, geologists, systems scientists: the implications run right through the sciences.) Let’s work together. I’ll be moving to the Bay Area soon, to set up an Evolutionary Cosmology working group of domain experts to develop this theory, and break it into the mainstream. You can think of the working group as...

• A startup, to disrupt cosmology.

• A heist team, to crack astrophysics.

• An Ocean’s 11 for astronomy.

Doesn’t that sound more fun than spending 70% of your time doing paperwork for National Science Foundation grants that can be abolished on a governmental whim halfway through the project? Join me!

If you’re a billionaire intrigued by all this: Reach out. So far, this project has run on small grants from the Irish Arts Council; Tyler Cowen’s Emergent Ventures fund; Jim O’Shaughnessy’s O’Shaughnessy Ventures; and the support of paid subscribers. (Plus my, inherently unreliable, publishing royalties to keep us going between grants.) I’d rather not waste so much time chasing funding; I need a Medici who’s willing to back a remarkably affordable intellectual revolution. (If you’re a Medici in a hurry, the PayPal link is here; if you’re not in a hurry, let’s talk.)

And if you’re not a journalist, a scientist, or a billionaire, just a reader who finds this work exciting? Tell everyone you know, that you think might be interested. Forward this post to friends. Talk about it in pubs, and labs, and cafés. At parties, and bus stops, and conferences. Think about it, and throw ideas into the comments.

We have discovered a new intellectual continent: it needs explorers. You can be one.

(Thanks to Jamie Rumbelow, John Caldemeyer, Sophie Gough Fives, Tyler Cowen, Digivijay Singh, Jenny Wagner, Kevin Kelly, PJ King, Rohit Krishnan, Solana Joy, Yogi Jaeger, Ananyo Bhattacharya, Ben Yeoh, and Adam Mastroianni for suggestions, and for reading drafts. I acted on some of their excellent advice, but also ignored much excellent advice – so nothing bad, wrong, or irritating that remains is their fault.)

Are bees smart?

To answer that question, here’s a crab spider:

Sadly, this is not a review of a book called The Mind of a Crab Spider. But as you crab spider lovers know, crab spiders and bumble bees are natural rivals.

Both bees and crab spiders are well-matched for strength and speed, and in the Rumble with the Bumble, the crab spider doesn’t necessarily win. Bees can often evade the spider, and live to pollinate another day. Lars Chittka, who wrote The Mind of a Bee, decided to build fake robotic crab spiders, and had them really robotically attack bumble bees when they visited flowers.

I remember when I was randomly attacked by robotic crab spiders for a day, and I didn’t enjoy it much. The bees shared my opinion. Not only did the bees have a bad time, their behavioural patterns totally changed. They began to approach the flowers differently. They began inspecting flowers via quick scanning flights before landing on them, and would occasionally reject flowers even if there was no crab spider present. They seemed more nervous.

If you want to see if humans are optimistic or pessimistic, you point at a glass of water that is halfway filled and ask them to describe it. Similarly, you can do the glass half-full versus half-empty test on bees, where you give them an ambiguous stimulus - it might be sucrose, which bees love, or it might be quinine, which they hate - and see if they want it.

If they want it, they’re likely a happy-go-lucky bee with nothing on their mind. If you simulate the bee being attacked by a predator right before this test, they are much less likely to fly to the solution and much more likely to fly into the container labelled ‘Therabee’.

Does that mean bees feel emotions? If they feel emotions, would that mean bees have conscious states? Or are these all just instinctive responses?

There are all manner of creatures and living beings whose experiences we remain fundamentally uncertain about. Bees fall into this hinterland of consciousness. Some readers will likely enter the book believing that bees do not have conscious experiences, and Lars Chittka does a good job disabusing these people of their certainty in this belief, if not the belief altogether.

Almost every chapter discussed in this book in some way reflects on this question of complex cognition versus instinct. There are instructive questions here. What does it mean for a response to be instinctive? Likewise, what sort of response is evidence of complex cognition? Is this a useful distinction? And if you think those questions are polarising, you’re starting to see like a bee.

I’ll reflect more on these questions later, but for now, let’s just talk bees.

WAGGLE YOUR BEE BOOTY

The ‘waggle dance’ was discovered by Karl von Frisch in the months after the end of World War II. Frisch worked in Nazi Germany during the war. Frisch was fractionally Jewish and his colleagues accused him of “bigoted opposition towards antisemitism” in his writings. He seems to have bowed to the pressure, penning a nasty tract on “racial hygiene” and recommending sterilisation of mentally unwell people. A charitable interpretation of these events would argue that he hoped to protect himself and other Jewish researchers by doing so.

Nonetheless, the Nazis planned to remove him from his university post. But Frisch was saved by Nosema, a single-celled gut parasite, that wiped out several hundred thousand beehives and began causing issues for food security. Martin Bormann, chief of the Nazi party chancellery, decided that Frisch’s dismissal should be postponed to the end of the war, which meant that Frisch stayed on, which meant that Frisch was able to discover the waggle dance. So I suppose we owe a small debt to that nasty parasite, and also to Nosema.

The waggle dance is a line dance performed by honey bees - an individual bee discovers a food source, and boy, is she excited (the vast majority of bees are female). She arrives back in the hive, and begins ferociously waggling while running in a line, then does a semicircular loop and starts the dance again. The angle that the bee runs in relation to the vertical is the angle that the food source is relative to the sun. If a bee runs straight upwards, the food source is in the direction of the sun. The distance the bee runs is proportional to the distance to the food source.

As you may have surmised at some point, the sun moves. The dancers factor this in, and as they spend longer performing the dance, they change the angle of their dance in order to factor in the movement of the sun. And if the sun is behind a cloud? Bees are sensitive to polarised light, and can infer the location of the sun from this light.

An important detail here is that it is generally dark inside hives. Other bees can’t actually see the dance. Instead, bees that dream of foraging put their feelers on the dancer’s abdomen, and hold them there while the dancer shimmies. It’s probably very arousing if you’re a bee. Bees have specific dialects of their waggle dance language, and some bees can learn the dialects of other bee languages if they spend enough time on Duolingo and don’t get killed by the foreign bees.

Chittka tells you all of this, and then tells you that for most bees, the dance is totally pointless.

If you tilt the hive, the bees can adapt and continue to use the sun as a reference. But if you only give bees diffuse light, they lose their reference point. They keep waggling, but the dances become basically randomised, even with just one dancer. Other bees try to follow along, but it’s essentially a waste of time - they’re just getting a distance, with no direction. And yet, if you compare the foraging performance of bees that can’t use the dance with those who can, there is no difference.

This is because the vast majority of bees evolved in tropical Asia, where they lived in large tropical forests, which present many more problems for navigation than the comparatively open European landscapes. Bees that can waggle dance in a tropical forest are about seven times as successful at foraging as those that can’t. If you’re planning on moving your bee colony to Thailand, have them brush up on their cha-cha-cha. But sadly, the hallowed bee dances in old Bavarian beehives have basically no function.

HONEY, I’M COMB

Honeycombs house larvae and store food. That’s the reason bees build them.

The honeycombs we’re most familiar with are built specifically by honey bees - there are a lot of bee species (~20,000), and not all of them are social. Of the species that build honeycombs, not all of them build in the same way. For instance, honey bees build hexagonal cells, whereas bumble bees build round cells. But round cells waste more space in the hive, since round things do not tessellate.

You could build square or triangular cells, but apparently larvae have to be raised in cells that aren’t square or triangular. Chittka does not explain why this is the case, but if you watch bees in action, it looks like a hexagon is a trade-off between the shape of a bee - roughly circular - and the waste of space problem above.

Mathematically, honey bees have done a great job. The Honeycomb Conjecture states that a “regular hexagonal grid or honeycomb has the least total perimeter of any subdivision of the plane into regions of equal area”, and it was proven in 1999 by Thomas Hales. This was to the delight of honey bees everywhere, who to celebrate, constructed a small hexagonal sculpture of Thomas Hales with a humongous perimeter.

Honey bees are the only bee species that builds double-sided hexagonal combs. The bottom of each cell has the shape of a pyramid, and the two sides of the comb are connected through the pyramid-shaped base of the cell. Honey bees also build their combs vertically, so that honey doesn’t leak out. But they don’t build them purely vertically - they keep it at an angle so that the honey’s viscosity and adhesion keep it inside the vessel. The cells are tilted slightly downwards.

That’s the structure. But as you read the next part, keep in mind the battle of cognition and instinct, the prize fight of cerebral might; it’s time to put your money where the honey is.

The first row of honeycomb cells is different from the others - it’s a foundation. It may seem like worker honey bees just use their body as a template to create the cell - but worker bees build cells for drone bees, and drone bees are larger than worker bees, by about 30%. There are pillars and cross beams that stabilise the comb. The queen gets a differently shaped cradle structure. Different workers will continue work where other workers have left off, so the bee isn’t just implementing a schematic of a cell and following it to completion. Bees will make alterations to the construction of the cells of other bees. If one bee misplaces wax, other bees will correct it.

François Huber, Marie-Aimée Lullin and François Burnens, working in the late 18th and early 19th centuries, started a malicious and fascinating program of messing with bees’ beeswax to see how they’d respond. Bees usually build downwards, by attaching honeycomb to the ceiling of the hive. If you stop them doing this, like the mean old team of François² and Marie, they reverse all their motor sequences and build a tower like in Tower Bloxx. If you stop them building up or down, they go from side to side.

If you put glass in the way of the ‘comb - bees hate attaching their ‘comb to glass - bees will rotate their construction. And they seem to notice the glass is there, because they usually rotate their construction prior to reaching the glass wall. If you keep doing it, the bees keep rotating, ducking and diving and moving their hive around you. This changes all the dimensions of the cells, and you get cells that are wider on the outside. How do bees ‘agree’ on their ‘comb dimensions, or the new directions to take? It’s not necessarily clear.

If you keep putting glass in, by the way, the bees eventually just attach to glass. Better glass for the ‘comb than nothing. In winter, bees stop foraging and conserve energy. But in winter, in one of Huber’s glass hives, some comb broke off the ceiling. The bees awoke, fortified the dislodged comb with pillars and crossbeams, and then reinforced the other combs attached to the ceiling, in case they got dislodged. The bees, having experienced a crisis, invested in additional security.

Bees can also do it in space. There were bees on a Challenger mission in 1984 (two years before the tragedy), and zero-g-bees constructed honeycombs with cells of normal dimensions, combining that with other trivial details like learning to fly in space. The difference between space ‘combs and earth ‘combs? The bees got rid of the slight angle downwards - there’s no gravity in space, and thus no need for the angle.

Bees that are raised alone can build honeycombs, but their diameters and cell structures aren’t great. Bees that don’t go through the acadebee are not as good at building as those who do. It is easy to say that bees are just following some program, but it’s very difficult to feel that as you see the way that they build.

BEE BRAINED

This is what a bee’s brain looks like, as stolen liberally from this paper:

Central body (CB) and one of the pair of lobulas (Lo), medullas (Me), antennal lobes (AL), mushroom body calyces (MBC) and mushroom body lobes (MBL).

If you look at the image on the right above, the lobules, medullas and antennal lobes all broadly handle sensory information. Learning takes place in the mushroom body calyces and lobes. Bees utilise something known in machine learning as a ‘fan-out, fan-in’ architecture. This involves spreading out tasks to multiple destinations, where they can be computed in parallel, before sending those tasks to the similar final destinations.

There are hundreds of neuronal links between the antennal lobe and the visual system, and these ‘fan-out’ to ~170,000 Kenyon cells (a type of cell specific to the mushroom body), which then ‘fan-in’ to 400 mushroom body extrinsic neurons, which then connect back to the brain where appropriate behavioural responses are selected.

It is possible to build computerised models of this circuitry, and from these relatively simple circuits produce much more complex learning responses than you might expect.

Incidentally, before a bee leaves the hive to forage, which they do at about 2-3 weeks old, their brains enlarge drastically - their mushroom bodies grow between 15-20%, presumably in order to memorise a load of information about flower locations and the spatial environment. I remember friends at school who displayed a similar ability before they were about to enter an exam hall.

BEE BRAIN WAVES

The oscillation of the brain is somewhat understudied. As this book review about brain waves points out, that there are basically no books on brain waves. Similarly, when I was doing my masters in neuroscience, brain waves came up a bunch, but mostly in a “and this elicits an alpha wave response which we’ve put into our big list of alpha wave responses and look how pretty our list is” way. I never really felt like I got an intuitive understanding of the significance of an alpha wave, or a theta wave, or the Mexican brain wave.

One research team studied Drosophila, or flies, when they were asleep, and found that they also have varying brain wave oscillations when they’re sleeping (which makes you wonder if those flies ended up with human-shaped sleep paralysis demons). Broadly, brain waves are linked to conscious experiences.

To expand on that, measuring consciousness is hard. Duh. As Anil Seth points out in Being You, one easy bit of consciousness to measure is awareness. We have different conscious experiences when we are under anaesthetic (i.e., none), when we’re asleep, and when we’re awake. These correlate with different patterns of brain waves.

Bees need to sleep. If they don’t get their beauty sleep, their dance moves get worse. If you expose bees to odours while they are in a phase of deep sleep, they begin to consolidate memories from the previous day, and their memories improve. Rats do a similar thing, repeating activation patterns from the previous day in their hippocampus. Does this suggest these creatures are dreaming?

Humans have a pattern of brain activity known as the default mode network, which occurs when our minds wander, or we daydream. Insect brains also exhibit patterns of activity akin to having a default mode network. None of this necessarily means insects ‘think’ or that they have ‘attention’ in the same way that we think and pay attention to things. But it all suggests that they might have similar processes that generate similar, if maybe more rudimentary, forms of those experiences.

NOT A FAN?

One of the most commonly cited impressive bee-behaviours is that they seemingly act in an organised way without any general command structure. But this is an emergent phenomenon. Take ‘fanning’. Bees have to keep the hive ventilated to avoid suffocation. If the hive is poorly ventilated, bees start fanning their wings to increase air circulation. If it is very poorly ventilated, all of the bees do this.

The book is filled with these scenarios, and the best way to read it is to try and solve the problems yourself. Imagine for a second you’re not sitting in your bee-tagging, brain-scanning ivory tower getting readings of how a microscopic part of the mushroom calyx responds to the sound of a frog jumping into a pond.

You’re a scientist in the past, who has to saddle horses and wait for bees to swarm so you can try to follow them, leaping fences and hurtling o’er hill and dale in a wild ride after nature - François Huber actually did this to figure out where bees were going when they were swarming!

When it comes to bee fanning, what’s the solution? How do the bees decide how many of them should be fanning? What do you think François Huber guessed? There’s no communication, but as the ventilation gets worse in the hive, more and more bees start fanning their wings. How would you design bees to solve this problem? You don’t want every bee fanning their wings 24/7 or they’re wasting time, but a nice ratio of ‘bees fanning’ to ‘bees not fanning’ that adapts in order to hit your ventilation criteria.

When Huber examined the fanning problem, he came up with an elegant theory. He suggested that bees are differentially sensitive to noxious smells. So as the noxious smells get worse, the sensitivity threshold of more and more bees is reached, and more of them begin fanning until ultimately the entire hive is fanning. Termites likely have a similar set of proclivities, as they will close and open entrances into their mounds in response to humidity changes within.

Note that there are lessons for specialisation and other forms of social organisation here - if you are more sensitive to noxious smells, you spend more time fanning, and you get better at it. This may cement differences in job allocation. Small differences in preference become permanent differences between individuals.

This sort of thing probably happens in humans - if you dislike dirty dishes lying around more than your slobby roommate Jeremy, you get better at doing the dishes. Jeremy doesn’t. This means you start to resent Jeremy and leave him passive aggressive post-it notes around the house.

In another example, Karl von Frisch observed that some bees were very picky about the sweetness of foods they would tolerate, and (many years later) Robert Page and others found that differences in sensitivity to sugar are present when bees are hours old, and determine whether they become pollen or nectar foragers.

None of that to my mind requires particularly complex cognition. It’s just a simple instinctive response to a problem that generates a relatively complex set of solutions. Other properties which bees have that we consider complex may also be emergent in a similar way.

SO ARE BEES SMART? DO YOU HAVE THE ANSWERS?

No. Sorry. Not really. It’s tricky. Bees seem to have a form of metacognition, where they will leave an experiment if they are unsure about what will happen, as opposed to making a decision. I might try a similar tactic.

You have to be careful when measuring performance in things that can’t communicate. Scientists used to think that bees that were tested on colour discrimination performed to the best of their abilities. Nope. Instead, it turns out bees deploy a speed-accuracy tradeoff, where they will make rapid decisions if there is no downside to doing so.

So, for decades, scientists assumed that bees could not differentiate squares and triangles, because in their experimental paradigms where they asked bees to look at squares and triangles, there was no downside. Bees just blitzed the test to get that sweet, sweet sucrose. Once penalties were introduced for mistakes, bees started performing much better and could easily differentiate squares and triangles.

This illuminates one large problem with measuring cognition in vertebrates and invertebrates. If an animal can’t do something in one test, it tells you a lot more about that test than the animals’ generalised skillset. This is really frustrating. Many animals don’t pass the mirror test, where you place something on their head and see if they try to remove it. This is frequently used to see if animals are aware of themselves. Bees don’t generally pass the mirror test. But why would they? The majority of bees have pretty similar faces.

On the other hand, Polistes wasps live in small colonies, and invest heavily into face recognition. This is because to determine their place in the colony hierarchy, Polistes wasps have fights. It’s useful to be able to recognise faces to learn the hierarchy.

Say I’m Wasp A, and I have a fight with Wasp B, and it’s close, but I lose. I then watch Wasp B get absolutely pancaked by Wasp C. It is helpful to recognise Wasp C so I know not to fight that hulking behemoth of a wasp.

Bees do have some awareness of their bodies, because before they fly through gaps they often make scanning flights to see if they need to approach the gap diagonally, sideways, or go around.

Here’s a ridiculous diagram of this, from this paper:

Does this count as self-awareness? For me, these tests all seem to fall short of the quiddity of consciousness. But they can’t all be ignored, can they?

You may be familiar with William Molyneux’s Problem:

Suppose a man born blind, and now adult, and taught by his touch to distinguish between a cube and a sphere of the same metal, and nighly of the same bigness, so as to tell, when he felt one and the other, which is the cube, which is the sphere. Suppose then the cube and the sphere placed on a table, and the blind man made to see: query, Whether by his sight, before he touched them, he could now distinguish and tell which is the sphere, which the cube? To which the acute and judicious proposer answers: ‘Not. For though he has obtained the experience of how a globe, and how a cube, affects his touch; yet he has not yet attained the experience, that what affects his touch so or so, must affect his sight so or so...’.

Put simply, if you give a cube and a sphere to a blind man, let him feel them, and then give the blind man sight - can he tell them apart?

It’s a hard question to answer in human subjects, because they can often get the information from elsewhere. It’s hard to restrict humans sufficiently to determine if they’ve actually passed this test. But bumble bees can identify shapes in the dark that they have only seen previously, or vice versa. Bees seem to have little difficulty transferring information between sensory modalities. If bees appear to hold mental representations of objects, does that take them further along the spectrum of consciousness towards higher bee-ings?

In an interview with Tyler Cowen, David Deutsch once discussed the idea of understanding, or of having explanatory power for the actions you take. For Deutsch, this distinguishes us from almost everything else that is alive.

Deutsch:

[Animals] have genes which contain knowledge, but it is fixed knowledge, and it is not the kind of knowledge that constitutes understanding. Understanding is always explanatory. You can write a book on canine behavior and look in chapter 37 and it will tell you what a dog will do when such and such happens to it. Sometimes it will say, “Some dogs will do this; some dogs will do that.” There is no such book for humans because chapter 37 will be blank. It’ll say, “Humans are going to do something that neither we nor you can predict.”

It’s unlikely bees could explain their actions, if we could even find a way for them to do so. But is their cognition just a “program enacted by their genes”? Deutsch also mentions squirrels:

You know squirrels bury nuts so they can dig them up later. Well, some people did a very cruel experiment. They put a squirrel, given some nuts or something (I don’t know how they set up the experiment), on a concrete floor. The squirrel did exactly the same behavior with its hind legs with the nuts and put the nuts there and so on. Even though it was having no effect whatsoever. We see the point of scrabbling with your hind legs and then nudging the nuts over there and so on, but it doesn’t. It’s just a program being enacted by its genes.

There’s no link to the lets-mess-with-squirrel-minds study, so it’s difficult to evaluate it. But it may have been the case that they tested the wrong squirrels.

Here’s what I mean: Chittka did an experiment with bees where he placed flowers with sucrose under a glass screen. The flowers were connected to a rope. If you pulled the rope, you could get at the sucrose. Bees were unleashed upon the flowers.

Many of the bees did not figure out the sucrose-rope-pulley system. But a handful of bright bees figured out that they could pull the rope. Tool use in bees! But here’s the really sweet part - remember those idiot bees from before that didn’t figure it out? Some of them watched the brainy bees, and started pulling the rope to get at the sucrose.

There are natural variances in intelligence in bee populations, as in every population. There are also variances in intelligence between colonies. It seems so advantageous to be a fast learner, that Chittka was curious why, evolutionarily, there are still slow learners. The main argument he finds is that bees who learned faster seemed to burn brighter, but were active for less long. This produced a pronounced effect where slower learners actually gathered more resources over their lifespan.

Animals are wily and irritating to test. Simply because one set of squirrels cannot figure out the concrete problem, it doesn’t necessarily mean all squirrels cannot. It may well be the case that all squirrels cannot, but you need to do a bunch of repeated studies, and then grant application boards start asking you why you need another boatload of cash to keep being a dick to squirrels.

This all makes it really difficult to say that animals ‘cannot’ do things, and this is really annoying and I don’t have a good solution other than to say in the short run, let’s keep being dicks to squirrels.

PUTTING THE ‘SCIENCE’ IN ‘CONSCIENCE’

Remember right at the start where I talked about anxious bees? Chittka says that his work suggests bees feel something:

Natural selection might not look kindly upon individuals that do not know fear, mothers who are indifferent to the loss of their offspring, or social animals for whom it does not “feel rewarding” to be in their social setting. In other words, having at least a range of basic emotions might be part of most animals’ “survival tool kit”.

Feelings are not the same as understanding or explanatory power. Emotion doesn’t necessarily equate to consciousness, and Chittka knows this:

There has never been a formal proof of consciousness in any animal, and in this book I have not supplied a formal proof for bees, either. Critical readers might counter that every single psychological phenomenon, every intelligent behavior, described in this book could somehow be replicated by a computing algorithm or a robot, and therefore could in theory be accomplished without any form of conscious awareness. They would be right.

You could design a robotic system for planning honeycomb construction, you could build robots that behave as if they experience pain when damaged, and you could of course mimic the counting abilities of bees quite easily in silico. And the list goes on. But, first of all, if you wanted to build an automaton that could do everything I’ve described in this book - the dozens of “innate” as well as learned and innovated behaviors - you would have to equip your robot with a very long list of detailed instructions, and your machine would still be able to cope only with what you have programmed it to cope with. It would be helpless with any novel challenge for which you have not written any code.

If or when we get to forms of AI consciousness and intelligence, and particularly early forms of that consciousness, they might look a little like what we see from the rest of life on earth; complex behaviours that could easily be explained away by some hand-waving, but perhaps shouldn’t be. Perhaps they won’t.

As a question, it needs to be taken seriously, and we should probably be examining a much wider spectrum of animals - unfamiliarity with the creature seems to preclude a lot of science funding. Chittka notes how difficult it is to work on bees, and that his field lacks prestige, and academic recognition, and funding. And that’s bees! People know bees! People love bees!

Imagine being an oarfish specialist. Oarfish could exhibit really unique and weird forms of consciousness that are for some reason easy to study. But we wouldn’t know, because oarfish are a nightmare to find, and we aren’t funding trips to the underwater kingdom of the oarfish, and we aren’t making up a prestigious Rowers Medal for oarfish scientists to win, because no-one really cares about oarfish. I have maybe gone too deep into the oarfish analogy, but I hope you get my point.

One day, people much smarter than me will figure all of this out, but for now, this is where I fly out of the experiment chamber.

Instead, I’ll leave you with one last titbit - bumble bees used to be called humble bees. The linguistic change wasn’t particularly exciting, but I like the old name. It wasn’t because the bees in days of yore were modest, god-fearing types - it was because they hummed. Cute, right?

If you’ll permit me to block quote one last time, I’ll leave you with Charles Darwin talking about his humble-bees and his hive bees ‘cheating’ while gathering nectar:

One day I saw for the first time several large humble-bees visiting my rows of the tall scarlet Kidney Bean; they were not sucking at the mouth of the flower, but cutting holes through the calyx, and thus extracting the nectar. And here comes the curious point: the very next day after the humble-bees had cut the holes, every single hive bee, without exception, instead of alighting on the left wing-petal, flew straight to the calyx and sucked through the cut hole; and so they continued to do for many following days. Now how did the hive-bees find out that the holes had been made? Instinct seems to be here out of the question, as the Kidney Bean is an exotic.

The holes could scarcely be seen from any point, and not at all from the mouth of the flower, where the hive-bees hitherto had invariably alighted. I doubt whether they were guided by a stronger odour of the nectar escaping through the cut holes; for I have found in the case of the little blue Lobelia, which is a prime favourite of the hive-bee, that cutting off the lower striped petals deceived them; they seem to think the mutilated flowers are withered, and they pass them over unnoticed. Hence I am strongly inclined to believe that the hive-bees saw the humble-bees at work, and well understanding what they were at, rationally took immediate advantage of the shorter path thus made to the nectar.

So what do you think? Are bees smart?

An 81-year-old woman with advanced Alzheimer’s lies in an Icelandic retirement home. She hasn’t recognized her family or spoken a word to them in over a year. Her son Lydur sits at her bedside, working on a crossword puzzle—not expecting any interaction, just being present.

Suddenly, she sits up. She looks him directly in the face.

“My Lydur,” she says. “I am going to recite a verse to you.”

And then she does, loudly and clearly:

Oh, father of light, be adored.
Life and health you gave to me,
My father and my mother.
Now I sit up, for the sun is shining.
You send your light in to me.
Oh, God, how good you are.

After which she lies back down and never talks again, dying soon thereafter.

This case, reported in Nahm et al., (2011), is an example of terminal lucidity: an unexpected return of cognitive function in patients with severe dementia (or other brain damage) occurring shortly before death.

As a neuroscientist, my first thoughts when encountering reports like this are that they can’t possibly be real. By the time patients with severe dementia actually die, their brains are catastrophically damaged. They typically show no signs of recognizing family members. They often haven’t responded meaningfully to their environment in months or years. Their brains are riddled with plaques and tangles. And they’ve lost 20-50% of their synaptic connections—so much that their brains have visibly shrunk on MRI scans.

On top of this existing damage, the dying process itself is hardly optimal for brain function. Most dementia patients die from infections, such as pneumonia from inhaling material into their lungs, or sepsis originating from bedsores. As bacteria multiply and the patients’ weakened immune systems struggle to respond, their brains suffer oxygen deprivation and exposure to inflammatory toxins. Everything we know about neuroscience says this should make cognition worse, not better.

And yet, terminal lucidity keeps being reported. In retrospective reports, 60-70% of dementia caregivers say they’ve witnessed at least one patient return to lucidity in the days before death (though given caregivers may observe many deaths over the years, the actual per-death incidence is potentially much lower). The only prospective study, which followed 100 hospice deaths, found it in 6% of cases. That’s not ubiquitous, but nor is it rare—in the US alone, it would mean around ten thousand cases per year.

As Michael Nahm, one of the key researchers in the field, puts it: “terminal lucidity… is more common than usually assumed, and reflects more than just a collection of anecdotes that on closer scrutiny emerge as wishful thinking.”

Ignoring trillion-dollar bills on the sidewalk

It might not be immediately obvious, but if terminal lucidity is real, it has profound implications for neuroscience and medicine.

We increasingly understand how memories are physically stored in the brain. Forming new episodic memories requires the hippocampus. Retaining memories depends on neural circuitry maintaining particular synaptic connection strengths. In animal models, selectively destroying these synapses can erase specific memories. In humans, as dementia progresses and these neural structures physically deteriorate - neurons dying, synapses disintegrating, the hippocampus shrinking - memory reliably worsens.

So when a woman who hasn’t recognized her son in a year suddenly calls him by name and recites a childhood poem, where have those memories been physically stored? If the neural substrate we thought was necessary for memory is degraded, but the memories can still be accessed, then either our understanding of memory storage is fundamentally wrong, or there’s some resilient substrate we haven’t identified yet.

This isn’t just academically interesting—understanding terminal lucidity could revolutionize dementia treatment. Right now, we’re already spending over a trillion dollars annually caring for 60 million people with dementia worldwide. Despite this, our best treatments like Lecanemab cost around $30,000 per year, in exchange for, at best, a minuscule reduction in the rate of disease progression (let alone actually reversing the decline).

But if terminal lucidity is real, it means there are naturally occurring conditions under which even severely damaged brains can have their function restored. Instead of trying to prevent amyloid accumulation—an approach that has largely failed despite billions invested—we could study what’s happening during these episodes. If a dying brain can indeed spontaneously access “lost” memories, perhaps we could induce a living brain to do the same.

The implications for brain preservation are equally significant. Severe dementia suggests irreversible identity loss, and if the person you were is already permanently gone, there’s little point in subsequently preserving your brain for attempts at future revival. This creates a terrible dilemma: anyone serious about preservation needs to consider undergoing the procedure before dementia progresses too far, raising obvious legal and personal challenges.

But if terminal lucidity is real, it suggests dementia might not be primarily a disease of memory loss, merely one of memory retrieval. The information is still in there, just inaccessible under normal conditions. For preservation advocates, any technology sophisticated enough to perform revival would likely be able to solve retrieval problems too. It would certainly be a great relief to already have precedent for people being recovered from brains that appear catastrophically damaged.

Motivated reasoning or reasonable belief?

Still, it would be easier to believe terminal lucidity was real if there weren’t such emotionally compelling reasons to want it to be real.

Watching a parent slowly lose themselves without ever having a final chance to say goodbye is devastating. People crave the deathbed scenes from books and movies: the elderly relative surrounded by family, pronouncing meaningful final words before peacefully expiring. But in reality, there’s often a long gap between when someone loses the ability to communicate and when they die—days, weeks, months, or even years in severe dementia.

This creates exactly the conditions for observer bias to flourish. Family members sit at bedsides desperately hoping for one last moment of connection, watching for any sign—however small—that their loved one recognizes them or is trying to communicate. They have time to rehearse what they wish they could say, and dream of what they wish they could hear back.

Consider this case from a 2024 study:

“[A 12-year-old girl with cancer, who had been unresponsive for months, was declared dead]. Just before this event, her mother noted that the patient opened her eyes, and mouthed “I love you” and “I’m ready to go home,” and blinked her eyes to answer yes/no questions. Her mother commented that she had not “seen her eyes” in several months and appeared to consider these moments as signs of life, despite her daughter’s medical condition.”

No one wants to ask a grieving mother about how confident she truly is that her dying daughter told her she loved her. But from a scientific perspective: was the girl actually forming words, or were those involuntary muscle movements interpreted as speech? The same question applies to a grandmother who seems to smile and nod (postural reflex?), or a dying spouse who squeezes a hand when told “I love you” (muscle spasm?).

The problem is that nearly all reports of terminal lucidity come from exactly these sources: emotionally invested family members or caregivers with strong motivations to perceive communication where there might be none, no scientific training to distinguish meaningful behavior from reflexes, and no blinding to prevent bias.

The only study I’ve found that actually tries to characterize what these episodes look like is Griffin et al., (2024), which surveyed family caregivers who reported witnessing lucid episodes in dementia patients. They found 51% of episodes involved communication that “made complete sense,” with another 32% making “some sense.” Although, keep in mind that 48% of episodes occurred at least six months before death, with only 18% happening in the final week. Most of these weren’t ‘terminal’ lucid episodes at all—they’re potentially just the normal cognitive fluctuations that occur throughout dementia progression.

So far then, we’re left almost entirely with anecdotal reports from biased observers, collected retrospectively, with no objective verification, mostly describing ambiguous behaviors that could arguably be explained by wishful interpretation.

And yet.

The phenomenon has been consistently reported across centuries and cultures. Although anecdotal, case reports of patients who haven’t spoken in years suddenly engaging in complex conversation are hard to dismiss as mere muscle spasms or observer bias. The single prospective study found it in 6% of deaths, suggesting something is happening beyond pure confabulation.

I genuinely don’t know whether terminal lucidity exists. Given that, the appropriate response is neither to credulously accept every report nor to dismiss the phenomenon entirely either—it’s to actually study it properly. Which brings us to what might be happening if any of this is real.

Brazen speculation

No one has actually studied the neuroscience of terminal lucidity while it’s happening —no EEG recordings, no brain imaging, no blood work during episodes. So any explanation is necessarily speculative. But we can make educated guesses based on known and related phenomena:

Brain swelling reduction

Dying patients often develop brain swelling from multiple causes: electrolyte imbalances, low blood protein, kidney or liver failure, inflammation, or damage to the blood-brain barrier. This swelling compresses brain tissue and reduces blood flow, starving neurons of oxygen and glucose. The result is the familiar progression of terminal delirium: confusion, then sleepiness, then coma, then death.

But in their final days, many patients stop eating and drinking entirely. The resulting dehydration could reduce brain swelling, allowing blood flow to increase and temporarily restoring some cognitive function—a brief window of lucidity before the dying process continues. It’s a simple mechanism: lower intracranial pressure means better perfusion, means more oxygen and glucose in the brain, which in turn means more functional neurons.

This is testable. If swelling reduction is the mechanism, we’d expect terminal lucidity to correlate with dehydration, and potentially with certain causes of death more than others.

Paradoxical network effects

Alternatively (or additionally) something stranger might be happening with how brain networks respond to extreme conditions.

Normal consciousness requires precise neurochemical balance. Too much excitation causes seizures. Too much inhibition causes unconsciousness. Wakefulness depends on brainstem neurons continuously releasing neuromodulators like acetylcholine, noradrenaline, and histamine into the cortex. When this system breaks down in a damaged or dying brain, people typically become progressively less responsive.

But sometimes, perturbing a broken system can temporarily fix it. If you’ve ever gotten a glitchy electronic device to work by hitting it, you’ve seen how pushing a malfunctioning system into an unusual state can accidentally restore function.

Maybe something similar happens in dying brains, where under extreme or unusual conditions, damaged neural networks can sometimes spontaneously reorganize in ways that temporarily restore function. One research paper speculates that dying brains might generate “spontaneous network integration manifesting as lucid behavior” through “rapid and nonlinear synchronisation.” Translated from Academic into English, this means “we think weird things might happen in a dying brain that briefly improve function, but we don’t really know what.”

Elucidating terminal lucidity

If terminal lucidity is real, it should be one of the most intensively studied phenomena in neuroscience. Instead, it’s barely studied at all. This is wholly unacceptable, given it’s far from impossible to investigate. The studies we need are straightforward, if ethically sensitive:

With informed consent from patients and families, video record supposed lucid episodes in people with advanced dementia or other terminal conditions. Have researchers who are blinded to each patient’s medical status and proximity to death score the recordings for evidence of lucidity. Compare these to baseline videos of the same patients, and to control recordings from other dementia patients who don’t experience supposed lucid episodes.

If independent, blinded raters confirmed that genuine improvements in cognition were occurring, follow-up studies could investigate what’s actually happening physiologically: neuroimaging to measure brain activity patterns, blood work to check for electrolyte changes, continuous monitoring to track the lead-up to episodes. If we could identify reliable precursors or correlates, we might be able to induce similar states safely in living patients.

Yes, this requires asking grieving families for permission to record and monitor their loved one’s final days. That’s difficult. But patients routinely consent to organ donation and whole-body donation to science—procedures far more invasive than video cameras and EEG caps. If anything, families motivated by the possibility of helping future dementia patients might welcome the chance to contribute to research.

And to be fair to researchers in this field, systematic surveillance with blinded scoring was proposed at a 2018 National Institute on Aging workshop specifically convened to examine terminal lucidity. A 2021 Guardian article even reports there are six funded studies ongoing, and that one of these, led by Sam Parnia at NYU Langone Medical Center, plans to monitor 500 dementia patients at end of life using continuous EEG with synchronized video recording.

Yet as of 2025, we still lack the rigorous evidence needed to answer the basic question of whether terminal lucidity is real, let alone understand its mechanisms. Meanwhile, we continue pouring billions into marginally effective amyloid-clearing drugs and performing study after study on Alzheimer’s mouse models that fail to translate to humans. A fraction of that funding redirected to studying terminal lucidity—a potentially naturally occurring phenomenon that might reveal entirely new therapeutic approaches—seems like an obviously better investment.

Note: just as I was finalising this post, the results of one of these new studies was published. Gilmore-Bykovskyi et al. (2025) prospectively monitored 20 hospice patients with advanced dementia using continuous video recording and found lucid episodes in 3 of them. This is great to see, but it does come with significant caveats: the study wasn’t blinded (the same family and clinicians who witnessed events also scored them) and the sample was tiny. Still, it strengthens the case for terminal lucidity being real, and provides strong justification for larger, blinded studies.

To me, the gross neglect that is terminal lucidity having been ignored for so long reflects a deeper issue: death and dying are drastically understudied across medicine and neuroscience.

Sure, from a traditional medical perspective, there’s little reason to investigate what’s happening in a terminal patient’s brain during their final hours, or how quickly neural decay occurs after blood flow stops. The patient is already doomed at that point, and palliative care’s focus on comfort and dignity doesn’t require detailed scientific understanding.

But terminal lucidity is a rare case where understanding what happens during dying matters to more than just brain preservation advocates. If severely neurologically damaged patients can spontaneously access “lost” memories and abilities under certain dying conditions, we need to know what those conditions are, why they work, and whether we can replicate them safely. However unlikely it is that terminal lucidity is truly real, something with such broad implications—for both dementia-hating normies and brain-preserving weirdos—shouldn’t be ignored.

Note: this essay was originally published in August, 2022. See this footnote1 for Claude’s assessment of how the predictions of this future history are holding up 2.5 years later (TLDR: “Remarkably prescient”).

The apparent rate of biomedical progress has never been greater.

On the research front, every day more data are collected, more papers published, and more biological mechanisms revealed.

On the clinical front, the pace is also rapid: the FDA is approving more novel therapeutics than ever before, buoyed by a record number of biologics approvals. A flurry of new therapeutic modalities—gene therapies, cell therapies, RNA vaccines, xenogeneic organ transplants—are slated to solve many hard-to-crack diseases over the coming decade.

It seems we are finally on the cusp of ending biomedical stagnation.

Unfortunately, this account isn’t quite correct. Though basic biological research has accelerated, this hasn’t yet translated into commensurate acceleration of medical progress. Despite a few indisputable medical advances, we are, on the whole, still living in an age of bio-medical stagnation.

That said, I’ll make a contrarian case for definite optimism: progress in basic biology research tools has created the potential for accelerating medical progress; however, this potential will not be realized unless we fundamentally rethink our approach to biomedical research. Doing so will require discarding the reductionist, human-legibility-centric research ethos underlying current biomedical research, which has generated the remarkable basic biology progress we have seen, in favor of a purely control-centric ethos based on machine learning. Equipped with this new research ethos, we will realize that control of biological systems obeys empirical scaling laws and is bottlenecked by biocompute. These insights point the way toward accelerating biomedical progress.

Outline

The first half of the essay is descriptive:

• 1. The Biomedical Problem Setting and Tool Review will outline the biomedical problem setting and provide a whirlwind tour of progress in experimental tools—the most notable progress in biomedical science over the past two decades.

• 2. The End (of Biomedical Stagnation) Is Nigh will touch on some of the evidence for biomedical stagnation.

• 3. The Spectrum of Biomedical Research Ethoses will recast the biomedical control problem as a dynamics modeling problem. We’ll then step back and consider the spectrum of research ethoses this problem can be approached from, drawing on three examples from the history of machine learning.

The second half of the essay is more prescriptive, looking toward the future of biomedical progress and how to hasten it:

• 4. The Scaling Hypothesis of Biomedical Dynamics will explain the scaling hypothesis of biomedical dynamics and why it is correct in the long-run.

• 5. Biocompute Bottleneck will explain why biocomputational capacity is the primary bottleneck to biomedical progress over the next few decades. We’ll then briefly outline how to build better biocomputers.

• 6. The Future of Biomedical Research will sketch what the near- and long-term future of biomedical research might look like, and the role non-human agents will play in it.

Niche, Purpose, and Presentation

There are many meta-science articles on why ideas are (or are not) getting harder to find, new organizational and funding models, market failures in science, reproducibility and data-sharing, bureaucracy, and the NIH. These are all intellectually stimulating, and many have spawned promising real-world experiments that will hopefully increase the rate of scientific progress.

This essay, on the other hand, is more applied macro-science than meta-science: an attempt to present a totalizing, object-level theory of an entire macro-field. It is a swinging-for-the-fences, wild idea—the sort of idea that seems to be in relatively short supply. (Consider this a call for similarly sweeping essays written about other fields.) However, because this essay takes on so much, the treatment of some topics is superficial and at points it will likely seem meandering.

All that said, hopefully you can approach this essay as an outsider to biomedicine and come away with a high-level understanding of where the field has been, where it could head, and what it will take to get there. My aim is to abstract out and synthesize the big-picture trends, while simultaneously keeping things grounded in data (but not falling into the common trap of rehashing the research literature without getting to the heart of things, which in the case of biomedicine typically results in naive, indefinite optimism).

1. The Biomedical Problem Setting and Tool Review

Biomedical research is intimidating. At first glance, it seems to span so many subjects and levels of analysis as to be beyond general description. Consider all the subject areas on bioRxiv—how can one speak of the effect of cell stiffness on melanoma metastasis, the evolution of sperm morphology in water fleas, and barriers to chromatin loop extrusion in the same breath? Furthermore, research is accretive and evolving, the frontier constantly advancing in all directions, so this task becomes seemingly more intractable with time.

That said, biomedical research does not defy general description. Though researchers study thousands of different phenomena (as attested to by the thousands of unique research grants awarded by the NIH every year), the scales of which range from nanometers to meters, underneath these particularities lies a unified biomedical problem setting and research approach:

The purpose of biomedicine is to control the state of biological systems toward salutary ends. Biomedical research is the process of figuring out how to do this.

Biomedical research has until now been approached predominantly from a single research ethos. This ethos aims to control biological systems by building mechanistic models of them that are understandable and manipulable by humans (i.e., human-legible). Therefore, we will call this dominant research ethos the “mechanistic mind”.

The history of biomedical research so far has largely been the story of the mechanistic mind’s attempts to make biological systems more legible. Therefore, to understand biomedical research, we must understand the mechanistic mind.

Tools of the Mechanistic Mind

The mechanistic mind builds models of biology by observing and performing experiments on biological systems. To do this, it uses experimental tools.

Because they are the product of the mechanistic mind, these tools have evolved unidirectionally toward reductionism. That is, these tools have evolved to carve and chop biology into ever-smaller conceptual primitives that the mechanistic mind can build models over.

We’d like to understand how the mechanistic mind’s models of biology have evolved. However, delving into its particular phenomena of study—specific biological entities, processes, etc.—would quickly bog us down in details.

But we can exploit a useful heuristic: experimental tools determine the limits of our conceptual primitives, and vice versa. Therefore, we can tell the story of the mechanistic mind’s progress in understanding biology through the lens, as it were, of the experimental tools it has created to do so. By understanding the evolution of these tools, one can understand much of the history of biomedical research.

The extremely brief summary of this evolution goes:

In the second half of the 20th century, biomedical research became molecular (i.e., the study of nucleic acids and proteins). At the turn of the 21st century, with the (near-complete) sequencing of the human genome, molecular biology became computational. The rest is commentary.

Scopes, Scalpels, and Simulacra

That summary leaves a lot to be desired.

To make further sense of it, we can layer on a taxonomy of experimental tools, composed of three classes: scopes, scalpels, and simulacra.

• “Scopes” are used to read state from biological systems.

• “Scalpels” are used to perturb biological systems.

• “Simulacra” are physical models that act as stand-ins for biological systems we’d like to experiment on or observe but can’t.

This experimental tool taxonomy is invariant across eras and physical scales of biomedical research, generalizing beyond any particular paradigm like cell theory or genomics. These three tool classes are, in a sense, the tool archetypes of experimental biomedical research.

Therefore, to understand the evolution in experimental tools (and, consequently, the evolution of the mechanistic mind), we can simply track advances in these three tool classes. We will pick up our (non-exhaustive) review around 15 years ago, near the beginning of the modern computational biology era, when tool progress starts to appreciably accelerate. (We will tackle scopes and scalpels now and leave simulacra for later.) I hope to convey a basic understanding of how these tools are used to interrogate biological systems, the rate they’ve been advancing at, and the resulting inundation of biomedical data we’ll soon face.

Scopes

To reiterate, scopes are tools that read state from biological systems. But this raises an obvious question: what is biological state?

As alluded to earlier, in the mid-20th century biological research became molecular, meaning it became the study of nucleic acids and proteins. Therefore, broadly speaking, biological state is information about the position, content, interaction, etc. of these molecules, and the larger systems they compose, within biological systems. Subfields of the biological sciences are dedicated to interrogating facets of biological state at different scales—structural biologists study the nanometer-scale folding of proteins and nucleic acids, cell biologists study the orchestrated chaos of these molecules within and between cells, developmental biologists study how these cellular processes drive the emergence of higher-order organization during development, and so on.

Regardless of the scale of analysis, advances in tools for reading biological state (i.e., scopes) occur along a few dimensions:

• feature-richness

• spatio-temporal resolution

• spatio-temporal extent

• throughput (as measured by speed or cost)

However, there are tradeoffs between these dimensions, and therefore they map out a scopes Pareto frontier.

In the past two decades, we’ve seen incredible advances along all dimensions of this frontier.

To illustrate these advances, we will restrict our focus to the evolution of three representative classes of scopes, each of which highlights a different tradeoff along this frontier:

• extremely feature-rich single-cell methods

• spatially resolved methods, which have moderate-to-high feature richness and spatio-temporal resolution

• light-sheet microscopy, which has large spatio-temporal extent, high spatio-temporal resolution, and low feature-richness

By tracking the evolution of these methods over the past two decades, we’ll gain an intuition for the rate of progress in scopes and where they might head in the coming decades.

But we must first address the metatool which has driven many, but not all, of these advances in scopes: DNA sequencing.

Sequencing as Scopes Metatool

DNA sequencing is popularly known as a tool for reading genetic sequences, like an organism’s genome. But lesser known is the fact that sequencing can be indirectly used to read many other types of biological state. You can therefore think of sequencing as a near-universal solvent or sensor of the scopes class—much progress in scopes has simply come from discovering how to cash out different aspects of biological state in the language of A’s, T’s, G’s and C’s.

The metric of sequencing progress to track is the cost per gigabase ($/Gb): the cost of consumables for sequencing 1 billion base pairs of DNA. Bioinformaticians can fuss about the details—error rates, paired-end vs. single-read, throughput, read quality in recalcitrant regions of the genome like telomeres or repetitive stretches—but for our purposes this metric provides the single best index of progress in sequencing over the past two decades.

You’ve probably seen the famous NIH sequencing chart, which plots the cost per Mb of sequencing (1000 Mb equals 1 Gb). However, this chart is somewhat confusing—the curve clearly appears piecewise linear, with steady Moore’s-law-esque progress from 2001 to 2007, then a period of rapid cost decline from mid-2007 to around 2010, followed by a seeming reversion to the earlier rate of decline.

For the purposes of extrapolation, the current era of sequencing progress started around 2010 (when Illumina released the first in its line of HiSeq sequencers, the HiSeq2000). When we plot sequencing prices from then onward, restricting ourselves to short-read methods, we get the following plot.

Over the past 12 years, the price per gigabase on high-volume, short-read sequencers has declined by almost two orders of magnitude, halving roughly every 2 years—slightly slower than Moore’s law.

The first order of magnitude cost decline came in the 2010-2015 period with Illumina’s HiSeq line, the price of sequencing plummeting from ~$100/Gb to ~$10/Gb; this was followed by 5 years of relative stagnation, for reasons unknown; and in the past 2 years, there’s been another order of magnitude drop, with multiple competitors finally surpassing Illumina and approaching $1/Gb prices. The sequencing market is starting to really heat up, and that likely means the biennial doubling trend will hold; if it does, and if current prices are to be believed, then we can expect sequencing prices to hit $0.1/Gb around 2028-2029.

To make this trend more intuitive, we can explain it in terms of the cost of sequencing a human genome.

A haploid human genome (i.e., one of the two sets of 23 chromosomes the typical human has) is roughly 3 Gb (bases, not bytes) on average (e.g., the X chromosome is much larger than the Y chromosome, so a male’s two haploid genomes will differ in size). Therefore, sequencing this haploid genome at 30x coverage—meaning each nucleotide is part of (i.e., covered by) 30 unique reads on average—which is the standard for whole genome sequencing, results in ~90 Gb of data (call it 100 Gb to make the numbers round). So, we can use this 100 Gb human genome figure as a useful unit of measurement for sequencing prices.

In 2010 to 2011, a human genome cost somewhere in the $5,000 to $50,000 range; by 2015, the price had fallen to around $1000; and now, in 2022, it is allegedly nearing $100 (though this was already being claimed two years ago).

This exponential decline in price has led to a corresponding exponential increase in genome sequencing data. Since around 2014, the number of DNA bases added per release cycle to GenBank, the repository of all publically available DNA sequences, has doubled roughly every 18 months.

But as noted earlier, sequencing has many uses other than genome sequencing. Arguably, the revolution in reading non-genomic biological state has been the most important consequence of declining sequencing costs.

The Single-Cell Omics Revolution

Biological systems run off of nucleic acids and proteins, among other macromolecules. And because proteins are translated from RNA, all biological complexity ultimately traces back to the transformations of nucleic acids—epigenetic modification of DNA, transcription of DNA to RNA, splicing of RNA, etc. Sequencing-based scopes allow us to interrogate these nucleic acid-based processes.

We can divide the study of these processes into two areas: transcriptomics, the study of RNA transcripts, which are transcribed from DNA; and epigenomics, the study of modifications made to genetic material above the level of the DNA sequence, which can alter transcription.

Transcriptomics and epigenomics have been studied for decades. However, the past decade was an incredibly fertile period for these subjects due to the combination of declining sequencing costs and advances in methods for preparing biological samples for sequencing-based readout.

The defining feature of these sample preparation methods has been their biological unit of analysis: the single cell.

It’s not inaccurate to call the past decade of computational biology the decade of single-cell methods. The ability to read epigenomic and transcriptomic state at single-cell resolution has revolutionized the study of biological systems and is the source of much current biomedical optimism.

Applications of Single-Cell Omics

To understand how much single-cell methods have taken off, consider the following chart:

This is a plot of the number of cells in each study added to the Human Cell Atlas, which aims to “create comprehensive reference maps of all human cells—the fundamental units of life—as a basis for both understanding human health and diagnosing, monitoring, and treating disease.” The number of cells per study has been increasing by an order of magnitude a little under every 3 years—and the frequency with which studies are added is increasing too.

The HCA explains their immense ambitions like so:

Cells are the basic units of life, but we still do not know all the cells of the human body. Without maps of different cell types, their molecular characteristics and where they are located in the body, we cannot describe all their functions and understand the networks that direct their activities.

The Human Cell Atlas is an international collaborative consortium that charts the cell types in the healthy body, across time from development to adulthood, and eventually to old age. This enormous undertaking, larger even than the Human Genome Project, will transform our understanding of the 37.2 trillion cells in the human body.

The way these human cells are charted is by reading out their internal state via single-cell methods. That is, human tissues (usually post mortem, though for blood and other tissues this isn’t always the case) are dissociated into individual cells, and then these cells’ contents are assayed (i.e., profiled, or read out) along one or more dimensions of transcriptomic or epigenomic state.

Crucially, these assays rely on sequencing for readout. In the case of single-cell RNA sequencing, the RNA transcripts inside the cells are reverse transcribed into complementary DNA sequences, which are then read out by sequencers. But sequencing can be used to read out other types of single-cell state that aren’t natively expressed in RNA or DNA—chromatin conformation, chromatin accessibility, and other epigenomic modifications—which typically requires a slightly more convoluted library preparation.

The upshot is that these omics profiles, as they are called, act as proxies for the cells’ unique functional identities. Therefore, by assaying enough cells, one can develop a “map” of single-cell function, which can be used to understand the behavior of biological systems with incredible precision. Whereas earlier bulk assays lost functionally consequential inter-cellular heterogeneity in the tissue-average multicellular soup, now this heterogeneity can be resolved in its complete, single-cell glory. These single-cell maps look like so (this one is of a very large human fetal cell atlas, which you can explore here):

And the Human Cell Atlas is the tip of the iceberg—single-cell omics methods have taken the entire computational biology field by storm. These methods have found numerous applications: comparing cells from healthy and diseased patients; tracking the differentiation of a particular cell type to determine the molecular drivers, which might go awry in disease; and comparing single-cell state under different perturbations, like drugs or genetic modifications. Open up any major biomedical journal and you’re bound to see an article with single-cell omics data.

But this rapid expansion of single-cell data is only made possible by continual advances in methods for isolating and assaying single cells.

Single-Cell Omic Technologies

Over the past decade or so, single-cell sample preparation methods have advanced along two axes: throughput (as measured by cost and speed) and feature-richness (as measured by how many omics profiles can be assayed at once per cell and the resolution of these assays).

Svensson et al. explain the exponential increase in single-cell transcriptomic throughput over the past decade as the result of multiple technical breakthroughs in isolating and processing single cells at scale:

A jump to ~100 cells was enabled by sample multiplexing, and then a jump to ~1,000 cells was achieved by large-scale studies using integrated fluidic circuits, followed by a jump to several thousands of cells with liquid-handling robotics. Further orders-of-magnitude increases bringing the number of cells assayed into the tens of thousands were enabled by random capture technologies using nanodroplets and picowell technologies. Recent studies have used in situ barcoding to inexpensively reach the next order of magnitude of hundreds of thousands of cells.

That is, once a tissue has been dissociated into single cells, the challenge then becomes organizational: how do you isolate, process, and track the contents of these cells? In the case of transcriptomics, at some point the contents of the cell must undergo multiple steps of library preparation to transform RNA transcripts into DNA, and then these DNA fragments must be exposed to the sequencer for readout—all without losing track of which transcript came from which cell.

As can be seen in the graph, throughput began to really take off around 2013. But these methods have continued to advance since the above graph was published.

For instance, combinatorial indexing methods went from profiling 50,000 nematode cells at once in 2017, to profiling two million mouse cells at once in 2019, to profiling (the gene expression and chromatin accessibility of) four million human cells at once in 2020. With these increases in scale have come corresponding decreases in price per cell—even in the past year, sci-RNA-seq3, the most recent in this line of combinatorial indexing methods, was further optimized, making it ~4x less expensive than before, nearing costs of $0.003 per cell at scale.

Costs have also declined among commercial single-cell preparation methods. 10X Genomics, the leading droplet-based single-cell sample preparation company, offers single-cell RNA sequencing library preparation at a cost of roughly $0.5 per cell. But Parse Biosciences, which, like sci-RNA-seq3, uses combinatorial indexing, recently claimed its system can sequence up to a million cells at once for only $0.09 per cell. (Though one can approach these library preparation prices on a 10x chip by craftily multiplexing and deconvoluting—that is, by labeling cells from different samples, multiple cells can be loaded into the same droplet, and the droplet readout can be demultiplexed (i.e., algorithmically disentangled) after the fact, thereby dropping the cost per cell.)

Thus, like with sequencing, there’s an ongoing gold-rush in single-cell sample preparation methods, and commercial prices should only decline further—especially given that the costs of academic methods are already an order of magnitude lower.

The second dimension along which single-cell methods have progressed in the richness of their readout.

In the past few years, we’ve seen an efflorescence of so-called “multi-omic” methods, which simultaneously profile multiple omics modalities in a single-cell. For instance, one could assay gene expression along with an epigenomic modality, like chromatin accessibility, and perhaps even profile cellular surface proteins at the same time. The benefit of multi-omics is that different modalities share complementary info; often this complementary information aids in mechanistically interrogating the dynamics underlying some process—e.g., one can investigate if increases in chromatin accessibility near particular genes precede upregulation of those genes.

Yet not only can we now profile more modalities simultaneously, but we can do so at higher resolution. To give a particularly incredible example, the maximum resolution of (non-single-cell) sequencing-based chromatin conformation assays went from 1 Mb in 2009, to ~5 kb in 2020, to 20 base pairs in 2022—an increase in resolution of over four orders of magnitude. Since we’re nearing the limits of resolution, future advances will likely come from sparse input methods, like those that profile single cells, and increased throughput.

Thus, improvements in sequencing and single-cell sample preparation methods have revolutionized our ability to read out state from biological systems.

But as great as single-cell methods are, their core limitation is non-existent spatio-temporal resolution. That is, because these methods require destroying the sample, we get only a snapshot of cellular state, not a continuous movie—the best we can do is reconstructing pseudo-temporal trajectories after the fact based on some metric of cell similarity (which is perhaps one of the most influential ideas in the past decade of computational biology), though some methods are attempting to address this temporal limitation. And because we dissociate the tissue before single-cell sample preparation, all spatial information is lost.

However, this latter constraint is addressed by a different set of scopes: spatial profiling methods.

Spatial Profiling Techniques

If single-cell omics methods defined the 2010’s, then spatial profiling methods might define the early 2020’s. These methods readout nucleic acids along with their spatial locations. This information is valuable for obvious reasons—cells don’t live in a vacuum, and spatial organization plays a large role in multicellular interaction.

We’ll briefly highlight two major categories of these methods: sequencing-based spatial transcriptomics, which resolve spatial location via sequencing, and fluorescence in situ hybridization (FISH) approaches, which resolve spatial location via microscopy.

Sequencing-Based Spatial Transcriptomics

The premise of sequencing-based spatial transcriptomics is simple: rather than randomly dissociating a tissue before single-cell sequencing, thereby losing all spatial information, RNA transcripts can be tagged with a “barcode” based on their location, which can later be read out via sequencing alongside the transcriptome, allowing for spatial reconstruction of the locations of the transcripts.

These barcodes are applied by placing a slice of tissue on an array covered with spots that have probes attached to them. When the tissue is fixed and permeabilized, these probes capture the transcripts in the cells above them; then, complementary DNA sequences are synthesized from these captured transcripts, with specific spatial barcodes attached depending on the location of the spot. When these DNA fragments are sequenced, the barcodes are read out with the transcripts and used to resolve their spatial positions.

One major dimension of advance of these methods is spatial resolution, as measured by the size of the spots which RNA transcripts bind to, which determines how many cells are mapped to a single spot, and therefore the resolution with which gene expression can be resolved. Over just the past 3 years, maximum resolution has jumped by almost three orders of magnitude (image source):

These spatial transcriptomics methods produce stunning images. For instance, here’s a section of the developing mouse brain as profiled by the currently highest resolution method, Stereo-seq:

Each dot in the middle pane represents the intensity of the measured gene at that location, but plots of this sort can be generated for all the tens of thousands of genes assayed via RNA sequencing. In the left pane, these complete gene expression profiles are used to assign each dot to its predicted average cell-type cluster, as one might do with non-spatial single-cell transcriptomics.

Yet note the resolution in the rightmost pane of the above figure—it is even greater than that of the middle pane. This image uses in situ hybridization, the basis of the spatial profiling technique we’ll explore next.

smFISH

Fluorescent in situ hybridization (FISH) methods trade off feature-richness for increased spatial resolution. That is, FISH-based methods don’t assay the RNA of every single protein-coding gene like sequencing-based methods do, but in return they localize the transcripts they do assay better.

Instead of using sequencing for readout, these methods use the other near-universal solvent or sensor of the scopes class: microscopy.

That is, whereas spatial transcriptomics resolve the location of transcripts indirectly via sequencing barcodes, FISH methods visually resolve location via microscopy. They do this by repeatedly hybridizing (i.e., binding) complementary DNA probes to targeted nucleic acids in the tissue (attached to these probes are fluorophores which emit different colors of light); after multiple rounds of hybridization and fluorescent emission from these probes, the resulting multi-color image can be deconvolved, the “optical barcodes” used to localize hundreds of genes at extremely high spatial resolution.

Unlike spatial transcriptomics, which is a hypothesis-free method that doesn’t (purposefully) select for particular transcripts, in FISH the genes to probe must be selected in advance, and typically they number in the tens or hundreds, not the tens of thousands like with spatial transcriptomics (though in principle they can reach these numbers—see below). The number of genes that can be resolved per sample (i.e., multiplexed) is limited by the size and error-rates of the fluorescent color palette that is used to mark them, and the spatial resolution with which these transcripts are localized is limited by the sensor’s ability to distinguish these increasingly crowded fluorescent signals (perhaps so crowded that the distance between them falls beneath the diffraction limit of the sensor).

The general idea of FISH has been around for over 50 years, but the current generation of multi-gene single-molecule FISH (smFISH) methods only began to take off around 15 years ago. Since then, there’s been a fair deal of progress in gene multiplexing and the number of cells that can be profiled at once (images source):

But the best way to understand these advances is visually. For instance, we can look at MERFISH, a technique which has been commercialized, part of a growing market for FISH-based spatial profiling methods. Here’s what part of a coronal slice of an adult mouse brain, with more than 200 genes visualized, looks like (the scale-bar is 20 microns in the left pane and 5 microns in the right pane):

The amount of data these methods generate is immense:

As a point of reference, the raw images from a single run of MERFISH for a ~1 cm^2 tissue sample contain about 1 TB of data.

But like spatial transcriptomics, these smFISH methods have one major drawback: though they can spatially resolve extremely feature-rich signals, they lack temporal resolution—that is, they image dead, static tissues, a problem addressed by longitudinal methods like light-sheet microscopy.

Light-Sheet Microscopy

Biological systems operate not only across space, but across time.

Methods like light-sheet fluorescence microscopy (LSFM) trade off feature-richness in exchange for this temporal resolution, all while maintaining high spatial extent and resolution. The niche filled by LSFM is explained as follows:

Fluorescence microscopy in concert with genetically encoded reporters is a powerful tool for biological imaging over space and time. Classical approaches have taken us so far and continue to be useful, but the pursuit of new biological insights often requires higher spatiotemporal resolution in ever-larger, intact samples and, crucially, with a gentle touch, such that biological processes continue unhindered. LSFM is making strides in each of these areas and is so named to reflect the mode of illumination; a sheet of light illuminates planes in the sample sequentially to deliver volumetric imaging. LSFM was developed as a response to inadequate four-dimensional (4D; x, y, z and t) microscopic imaging strategies in developmental and cell biology, which overexpose the sample and poorly temporally resolve its processes. It is LSFM’s fundamental combination of optical sectioning and parallelization that allows long-term biological studies with minimal phototoxicity and rapid acquisition.

That is, LSFM gives us the ability to longitudinally profile large 3D living, intact specimens at high spatial and temporal resolution. Thus, it plays an important complementary role to moderately feature-rich spatial transcriptomic methods and extremely feature-rich single-cell methods, both of which lack temporal resolution.

Needless to say, the technology underlying LSFM has advanced quite a lot over the past two decades. The beautiful thing about spatio-temporally resolved methods is that we can easily witness these advances with our eyes. For instance, we can compare the state of the art in imaging fly embryogenesis in 2004 vs. in 2016:

Yet in just the past year, we have seen further advances still. A new, simplified system has improved imaging speed while maintaining sub-micron lateral and sub-two-micron axial resolution—and on large specimens, no less.

And recently an LSFM method was developed that can image tissues at subcellular resolution in situ, without the need for exogenous fluorescent dyes—in effect, a kind of in vivo 3D histology. The applications of this technology, to both research and diagnostics, are numerous.

Here’s the stitching together of a 3D-resolved (but not temporally resolved) slice of in situ mouse kidney:

And here’s real-time imaging of a fluorescent tracer dye perfusing mouse kidney tubules in situ:

Scopes Redux

The scopes Pareto frontier has advanced tremendously over the past two decades, and on many dimensions at a surprisingly regular rate. This is perhaps the most exciting development in all of biomedical research.

However, the ability to read state from biological systems doesn’t alone much improve our understanding of them—for that, we must perturb them.

Scalpels

Scalpels are used to experimentally perturb biological systems. The dimensions of the scalpels Pareto frontier are similar to those of the scopes Pareto frontier:

• feature precision

• spatio-temporal precision

• spatio-temporal extent

• throughput

However, the past decade has been one dominated by advances in scopes, not scalpels. One metric of this dominance is the annual Nature method of the year, which is a good gauge for what tools are becoming popular among biological researchers. Among the past 15 winners, two are scalpels (optogenetics and genome editing), two are related to simulacra (iPSC and organoids), and the rest are scopes (NGS, super-resolution fluorescence microscopy, targeted proteomics, single-cell sequencing, LSFM, cryo-EM, sequencing-based epitranscriptomics, imaging of freely behaving animals, single-cell multiomics, and spatial transcriptomics) or analytical methods (protein structure prediction).

Scalpels have simply experienced far narrower progress over the past decade or so than scopes. Advances occurred mostly along the feature-precision and throughput dimensions within a single suite of tools, which we will restrict our attention to (and therefore this section will be comparatively short).

Genome and Epigenome Editing

The most notable advance in scalpels has been in our ability to perturb the genome and epigenome with high precision, at scale. Of course, we’re referring to the technology of CRISPR, which researchers successfully appropriated from bacteria a decade ago (though there’s some disagreement about whom the credit should go to).

When one thinks of CRISPR, they likely think about making DNA double-strand breaks (DSB) in order to knock out (i.e., inhibit the function of) whole genes. But CRISPR is a general tool for using guide RNAs to direct nucleases (enzymes which cut DNA or RNA, like the famous Cas9) to specific regions of the genome—in effect, a kind of nuclease genomic homing guidance system. By varying the nuclease one uses and the molecules attached to them, a variety of functions other than knockouts can be performed: targeted editing of DNA without DSB, inhibiting gene expression without DSB, activating gene expression, editing epigenetic marks, and RNA editing and interference.

CRISPR is not the first system to perform most of these functions, and it certainly isn’t perfect—transfection rates are still low and off-target effects still common—but that’s beside the point: the defining features of CRISPR are its generality and ease of use. If sequencing and microscopy are the universal sensors of the scopes tool class, then CRISPR might be the universal genomic actuator of the scalpels tool class.

In conjunction, these scopes and scalpels have enabled interrogating the genome at unprecedented resolution and throughput.

Using Scopes and Scalpels to Interrogate the Genome and Beyond

By perturbing the genome and observing how the state of a biological system shifts, we can infer the mechanistic structure underlying that biological system. Advances in scopes and scalpels have made it possible to do this at massive scale.

For instance, one could use CRISPR to systematically knockout every single gene in the genome across a set of samples (perhaps with multiple knockouts per sample), a so-called genome-wide knockout screen. But whereas previously the readout of these screens was limited to simple phenotypes like fitness (that is, does inhibiting a particular gene produce lethality, which can be inferred by counting the guide RNAs in the surviving samples) or a predefined set of gene expression markers, due to advances in scopes we can now read out the entire transcriptome from every sample.

Over the past five years, a lineage of papers has pursued this sort of (pooled) genome-wide screening with full transcriptional readout at increasing scale. In one of the original 2016 papers, only around 100 genes were knocked out across 200,000 mouse and human cells; yet in one of the most recent papers, all 10,000+ protein-coding genes are inhibited across more than 2.5 million human cells, with full transcriptional readout.

Yet the genome is composed of more than protein-coding genes, and ideally we’d like to systematically perturb non-protein-coding regions, which play an important role in the regulation of gene expression. The usefulness of such screens would be immense:

The human genome is currently believed to harbour hundreds-of-thousands to millions of enhancers—stretches of DNA that bind transcription factors (TFs) and enhance the expression of genes encoded in cis [i.e., on the same DNA strand]. Collectively, enhancers are thought to play a principal role in orchestrating the fantastically complex program of gene expression that underlies human development and homeostasis. Although most causal genetic variants for Mendelian disorders fall in protein-coding regions, the heritable component of common disease risk distributes largely to non-coding regions, and appears to be particularly enriched in enhancers that are specific to disease-relevant cell types. This observation has heightened interest in both annotating and understanding human enhancers. However, despite their clear importance to both basic and disease biology, there is a tremendous amount that we still do not understand about the repertoire of human enhancers, including where they reside, how they work, and what genes they mediate their effects through.

In 2019, the first such massive enhancer inhibition screen with transcriptional readout was accomplished, inhibiting nearly 6,000 enhancers across 250,000 cells, an important step to systematic interrogation of all gene regulatory regions.

But inhibition and knockout are blunt methods of perturbation. To truly understand the genome, we must systematically mutate it. Unfortunately, massively parallel methods for profiling the effects of fine-grain enhancer mutations don’t yet read out transcriptional state at scale, instead opting to trade off readout feature richness for screening throughput via the use of reporter gene assays. However, we’ll likely see fine-grain enhancer mutation screens with transcriptional readout in the coming years.

Yet advances in scopes enable us to interrogate the effects of not only genomic perturbations, but therapeutic chemical perturbations, too:

High-throughput chemical screens typically use coarse assays such as cell survival, limiting what can be learned about mechanisms of action, off-target effects, and heterogeneous responses. Here, we introduce “sci-Plex,” which uses “nuclear hashing” to quantify global transcriptional responses to thousands of independent perturbations at single-cell resolution. As a proof of concept, we applied sci-Plex to screen three cancer cell lines exposed to 188 compounds. In total, we profiled ~650,000 single-cell transcriptomes across ~5000 independent samples in one experiment. Our results reveal substantial intercellular heterogeneity in response to specific compounds, commonalities in response to families of compounds, and insight into differential properties within families. In particular, our results with histone deacetylase inhibitors support the view that chromatin acts as an important reservoir of acetate in cancer cells.

However, though impressive, it’s an open question whether such massive chemical screens (and all the tool progress we’ve just reviewed) will translate into biomedical progress.

2. The End (of Biomedical Stagnation) Is Nigh

Progress in experimental tools over the past two decades has been remarkable. It certainly feels like biomedical research has been completely revolutionized by these tools.

This revolution is already yielding advances in basic science. To name but a few:

• Our understanding of longevity has advanced tremendously, from the relationship between mutation rates and lifespan among mammals, to the molecular basis of cellular senescence and rejuvenation (which could have huge clinical implications).

• Open up any issue of Nature or Science and you’re bound to see a few amazing computational biology articles interrogating some biological mechanism with extreme rigor, likely using the newest tools.

• And how can one forget Alphafold, the solution to a 50-year-old grand challenge in biology. “It will change medicine. It will change research. It will change bioengineering. It will change everything.” (Well, it isn’t necessarily a product of the tools revolution, but it is a major leap in basic science that contributes to the current mood of optimism.)

Biomedical optimism abounds for other reasons, too. Consider some of the recent clinical successes with novel therapeutic modalities:

• Gene therapy (i.e., genetically edited autologous stem cell transplants) seem poised to solve terrible blood disorders like sickle cell disease and familial hypercholesterolemia.

• RNA vaccines solved SARS-CoV-2, so maybe they’ll solve other infectious diseases, like HIV. Perhaps they’ll even solve pancreatic cancer. (Or how about the immuno-oncology double-whammy: combining CAR-T and RNA vaccines to treat solid tumors.)

It certainly feels like this confluence of factors—Moore’s-law-like progress in experimental tools, the ever-increasing mountain of biological knowledge they are generating, and a bevy of new therapeutic modalities that are already delivering promising clinical results—is ushering in the biomedical golden-age. Some say that “almost certainly the great stagnation is over in the biomedical sciences.”

How much credence should we give this feeling—are claims of the end of biomedical stagnation pure mood affiliation?

Premature Celebration

A good place to start would be to define what the end of biomedical stagnation might look like.

One of the necessary, but certainly not sufficient, conditions would be the normalization of what is often called personalized/precision/genomic medicine. Former NIH Director Francis Collins sketched what this world would look like:

…The impact of genetics on medicine will be even more widespread. The pharmacogenomics approach for predicting drug responsiveness will be standard practice for quite a number of disorders and drugs. New gene-based “designer drugs” will be introduced to the market for diabetes mellitus, hypertension, mental illness, and many other conditions. Improved diagnosis and treatment of cancer will likely be the most advanced of the clinical consequences of genetics, since a vast amount of molecular information already has been collected about the genetic basis of malignancy…it is likely that every tumor will have a precise molecular fingerprint determined, cataloging the genes that have gone awry, and therapy will be individually targeted to that fingerprint.

Here’s the kicker: that quote is from 2001, part of Collins’ 20-year grand forecast about how the then-recently-accomplished Human Genome Project would revolutionize the future of medicine (i.e., what was supposed to be the medicine of today). Unfortunately, his forecast didn’t fare well.

Though the 2000’s witnessed “breathtaking acceleration in genome science,” by the halfway point, things weren’t looking good. But Collins held out hope:

The consequences for clinical medicine, however, have thus far been modest. Some major advances have indeed been made…But it is fair to say that the Human Genome Project has not yet directly affected the health care of most individuals…

Genomics has had an exceptionally powerful enabling role in biomedical advances over the past decade. Only time will tell how deep and how far that power will take us. I am willing to bet that the best is yet to come.

Another ten years later, it seems his predictions still haven’t been borne out.

• Precision oncology fell short of the hype. Only ~7% of US cancer patients are predicted to benefit from genome-targeted therapy. And oncology drugs approved for a genomic indication have a poor record in improving overall survival in clinical trials (good for colorectal cancer and melanoma; a coin-toss for breast cancer; and terrible for non-small cell lung cancer). What we call genome-targeted therapy is merely patient stratification based on a few markers, not the tailoring of therapies based on a “molecular fingerprint”.

• There are no blockbuster “gene-based designer drugs” for the chronic diseases he mentions.

• Pharmacogenomics is the exception, not the norm, in the clinic. The number of genes with pharmacogenomic interactions is now up to around 120 (though only a subset of interactions are clinically actionable). And, as shown in the precision oncology case, oftentimes the use of genetic markers has little impact on the outcomes of the majority of patients.

Thus, despite the monumental accomplishment of the Human Genome Project, and the remarkable advances in tools that followed from it, we have not yet entered the golden-age of biomedicine that Collins foretold.

(But don’t worry: though the precision medicine revolution has been slightly delayed, we can expect it to arrive by 2030.)

Biomedical Stagnation Example: Cancer

The optimist might retort: that’s cherry-picking. Even though genomics hasn’t yet had a huge medical impact, and even though Collins’ specific predictions weren’t realized, he is directionally correct: biomedical stagnation is already ending. Forget about tools—just look at all the recent clinical successes.

Take cancer, for instance. Many indicators seem positive: five-year survival rates are apparently improving, and novel therapeutic modalities like immunotherapy and cell therapy are revolutionizing the field. Hype is finally catching up with reality.

Yes, there have undoubtedly been some successes in cancer therapeutics over the past few decades—Gleevec cured (that is not an exaggeration) chronic myeloid leukemia; and checkpoint inhibitors have improved treatment of melanoma, NSCLC, and a host of other cancers.

But in terms of overall (age-standardized) mortality across all cancers, the picture is mixed.

In the below graph, we see this (note the log-10-transformed y-axis):

Mortality from the biggest killer of the cancers, lung cancer, has fallen due to smoking cessation (the same goes for stomach cancer). The mortality rate for the second biggest killer among women, breast cancer, fell by ⅓ from 1990 to 2019—a good portion of this is attributable to improved therapeutics. Mortality from prostate cancer in men, and colorectal cancer in both sexes, have all also fallen around 30-40%, much of which is attributable to screening.

Yet mortality from pancreatic cancer hasn’t moved. Late-stage breast cancer is still a death sentence. The incidence (and mortality) of liver cancer has actually increased among both sexes, due to increased obesity. And plenty of the cancers we’ve lowered the incidence of through public health measures—esophageal, lung, stomach—still have incredibly low survival rates.

The declines in disability-adjusted life years (DALYs) lost mirror the declines in overall mortality (note again the log-scale):

And if you’re wondering if it looks any different for adults ages 55 to 59, perhaps because of an age composition effect, it does not. The declines are all roughly the same compared to the 55 to 89 group (that is, 30-40% declines for the major cancers like breast, prostate, colorectal, ovarian, etc.).

But one needn’t appeal to mortality or DALY rates to show things are still stagnant. Just look at how shockingly primitive our cancer care still is: we routinely lop off body parts and pump people full of heavy metals or other cytotoxic agents (most of which were invented in the 20th century), the survival benefits of which are often measured in months, not years.

The optimist might push back again: yes, the needle hasn’t moved much for the toughest cancers, but novel modalities like CAR-T are already having a huge impact on hematological malignancies, and they’ll someday cure solid tumors. Clearly biomedical stagnation is in the process of ending. Give it a bit of time.

To which the skeptic replies: Yes, CAR-T has shown some great results in some specific blood cancers, and there’s a lot to be optimistic about. But let’s not get ahead of ourselves. When one critically examines the methodology of many CAR-T clinical trials, the survival and quality of life benefits aren’t as impressive as its proponents would lead you to believe (as is the case with many oncology drugs). We’re likely at least a decade away from CAR-T meaningfully altering annual mortality of any sort of solid tumor.

Thus, it seems a bit premature to say biomedical stagnation in cancer has ended, based purely off some recent promising clinical trials—especially when we’ve been repeatedly sold this story before.

The war is still being fought, 50 years on. Likewise for the other five major chronic diseases, where the picture is often even bleaker.

We wanted a cure for cancer. Instead we got genetic ancestry reports.

Real Biomedical Research Has Never Been Tried

The optimist will backpedal and grant that biomedical stagnation hasn’t yet ended for chronic diseases (infectious diseases and monogenic disorders are another question). But despite biomedical progress being repeatedly oversold and under-delivered on, and despite us putting ever-more resources into it, they think this time is different. Yes, Francis Collins said the same thing 20 years ago, but forget the reference class: this time really is different. Those were false starts; this is the real deal. The end of biomedical stagnation is imminent.

There’s a simple reason for believing this: our tools for interrogating biological systems are on exponential improvement curves, and they are already generating unprecedented insight. Due to the accretive nature of scientific knowledge, it’s only a matter of time before we completely understand these biological systems and cure the diseases that ail them.

To which the skeptic might say: but didn’t the optimists make that exact same argument 10 or 20 years ago? What’s changed? Eroom’s law hasn’t: all the putatively important upstream factors (combinatorial chemistry, DNA sequencing, high-throughput screening, etc.) continue to become better, faster, and cheaper, and our understanding of disease mechanisms and drug targets has only grown, yet the number of drugs approved per R&D dollar has halved every 9 years for the past six or seven decades—with this trend only recently plateauing (not reversing) due to a friendlier FDA, and drugs targeting rare diseases, finer disease subtypes, and genetically validated targets.

Likewise for returns per publication and clinical trial. (And, by the way, more than half of major preclinical cancer papers fail to replicate.)

To which the optimist says: we’ve been hamstrung by regulation and only recently developed the experimental tools necessary to do real biomedical research. But now that these tools have arrived, they will change the game. We might even dare to say these tools, combined with advances in software, enable a qualitatively different type of experimental biomedical research. Biomedicine will become a “data-driven discipline”:

…exponential progress in DNA-based technologies—primarily Sequencing and Synthesis in combination with molecular tools like CRISPR—are totally changing experimental biology. In order to grapple with the new Scale that is possible, Software is essential. This transition to being a data-driven discipline has served as an additional accelerant—we can now leverage the incredible progress made in the digital revolution [emphasis not added].

Just wait—now that we have the right biological research tools, the end of biomedical stagnation is imminent.

The Mechanistic Mind’s Translational Conceit

But the optimist is begging the question. They assume progress in biological research will naturally lead to progress in biomedical outcomes, but we’ve repeatedly seen this isn’t the case: our biological research has (apparently) advanced tremendously over the past twenty years, yet this hasn’t translated into similarly tremendous medical results, despite all predictions to the contrary.

Yet the mechanistic mind is confident this will change. This is the mechanistic mind’s translational conceit: that accumulating experimental knowledge and building ever-more reductionistic models of biology will eventually lead to cures for disease. Once we carve nature at the joints (i.e., discover the ground-truth mechanistic structure of biological systems, expressible in human-legible terms), the task of translation will become easy. Through understanding nature, we will learn how to control it.

And if we look back in ten years and the mortality indicators haven’t budged much, then it simply means we’re ten years closer to a cure. This is a marathon, not a sprint. Stay the course: run more experiments, collect more data, and continue carving. The diseases will yield eventually. We must leave no nucleotide unsequenced, no protein un-spectrometered…

At a Crossroads

The mechanistic mind does not have a concrete model of biomedical progress. Rather, they have unwavering faith that more biomedical research leads to more translational progress. Their optimism is indeterminate. It is why they are repeatedly disappointed when amazing research discoveries fail to translate into cures but nonetheless maintain faith that more research is the answer—they can’t tell you when the cures will arrive, but at least they know they’re pushing in the right direction.

The mechanistic mind has certainly done a lot for us—there’s no denying that. But it will not deliver on the biomedical progress it has promised in a timely fashion.

However, there’s no need for despair: this time is, in fact, different. Progress in tools has created the potential for a radically different research ethos that will end biomedical stagnation. But to understand this new research ethos, we must first understand the telos of the mechanistic mind and why it is at odds with the biomedical problem setting.

3. The Spectrum of Biomedical Research Ethoses

Let’s return to our original formulation of the biomedical problem setting:

The purpose of biomedicine is to control the state of biological systems toward salutary ends.

This definition is rather broad. It doesn’t specify how we must go about learning to control biological systems. Let’s reframe the problem to make it more tractable.

The Dynamics Reframing of Biomedicine

We can first recast this problem as a search problem: given a starting state, s0, and a desired end state, s1, the task is to find the intervention that moves the system from s0 to s1. For instance, a patient is in a diseased state, and you must find the therapeutic intervention that moves them into a healthy state.

However, the space of interventions is large and biology is complex, so brute-force search won’t work. Therefore, we can further recast this search problem as the problem of learning an action-conditional model of biology—i.e., a dynamics model, to use the language of model-based reinforcement learning—to guide this search through intervention space. A dynamics model takes in an input state (or, in the non-Markovian case, multiple past states) and an action, and predicts the resulting state, f:S×A→S′.

The predictive performance of the dynamics model directly determines the efficiency of the search through intervention space. That is, the better your model predicts the behavior of a biological system, the easier it will be to learn to control it.

Thus, the biomedical control problem reduces to a search problem over intervention space, which itself reduces to the problem of learning a dynamics model to guide this search.

Learning this dynamics model is the task of biomedical research.

However, the mechanistic mind smuggles in a set of assumptions about what form this model must take:

This ethos aims to control biological systems by building mechanistic models of them that are explainable and understandable by humans (i.e., human-legible).

By interrogating this set of assumptions, we will see why the mechanistic mind is the wrong research ethos to approach the biomedical dynamics problem from.

Telos of The Mechanistic Mind

…In that Empire, the Art of Cartography attained such Perfection that the map of a single Province occupied the entirety of a City, and the map of the Empire, the entirety of a Province. In time, those Unconscionable Maps no longer satisfied, and the Cartographers Guilds struck a Map of the Empire whose size was that of the Empire, and which coincided point for point with it. The following Generations, who were not so fond of the Study of Cartography as their Forebears had been, saw that that vast Map was Useless, and not without some Pitilessness was it, that they delivered it up to the Inclemencies of Sun and Winters. In the Deserts of the West, still today, there are Tattered Ruins of that Map, inhabited by Animals and Beggars; in all the Land there is no other Relic of the Disciplines of Geography. — Jorge Luis Borges, On Exactitude in Science

A unifying telos, invariant across phenomena of study and eras of tooling, underlies the mechanistic mind.

This telos is building a 1-to-1 map of the biological territory—reducing a biological system to a perfectly legible molecular (or perhaps even sub-molecular) diagram of the interactions of its constituent parts. This is how the mechanistic mind intends to carve nature at the joints; it will unify through dissolution.

This isn’t hypothetical: a simplified version of such a map of biochemical pathways was created over 50 years ago for the major metabolic pathways and major cellular and molecular processes.

Those two maps are quite large and complex, but they are massively simplified and pre-genomic. Our current maps are far larger and more sophisticated.

These maps take the form of ontologies. For instance, Gene Ontology is “the network of biological classes describing the current best representation of the ‘universe’ of biology: the molecular functions, cellular locations, and processes gene products may carry out.”

Concretely, GO is a massive directed graph of biological classes (molecular functions, cellular components, and biological processes) and relations (“is a”, “part of”, “has part”, “regulates”) between them.

But simple ontologies are just the beginning. One can build extremely complex relational logic atop them, like qualifiers for relational annotations between genes and processes:

A gene product is associated with a GO Molecular Function term using the qualifier ‘contributes_to’ when it is a member of a complex that is defined as an “irreducible molecular machine” - where a particular Molecular Function cannot be ascribed to an individual subunit or small set of subunits of a complex. Note that the ‘contributes_to’ qualifier is specific to Molecular Functions.

But single annotations, even with qualifiers, are simply not expressive enough to encode the complexity of biology. Luckily, there’s an ontology of relations one can draw on to compose convoluted relations, which can be used to model larger, more complex biological systems:

GO-Causal Activity Models (GO-CAMs) use a defined “grammar” for linking multiple standard GO annotations into larger models of biological function (such as “pathways”) in a semantically structured manner. Minimally, a GO-CAM model must connect at least two standard GO annotations (GO-CAM example).

The primary unit of biological modeling in GO-CAM is a molecular activity, e.g. protein kinase activity, of a specific gene product or complex. A molecular activity is an activity carried out at the molecular level by a gene product; this is specified by a term from the GO MF ontology. GO-CAM models are thus connections of GO MF annotations enriched by providing the appropriate context in which that function occurs. All connections in a GO-CAM model, e.g. between a gene product and activity, two activities, or an activity and additional contextual information, are made using clearly defined semantic relations from the Relations Ontology.

For instance, here’s a graph visualization of the GO-CAM for the beginning of the WNT signaling pathway:

Conceivably, one could use these causal activity models to encode all observational and experimental knowledge about biological systems, including the immense amounts of genome-wide, single-nucleotide-resolution screening data currently being generated. The potential applications are immense:

The causal networks in GO-CAM models will also enable entirely new applications, such as network-based analysis of genomic data and logical modeling of biological systems. In addition, the models may also prove useful for pathway visualization…With GO-CAM, the massive knowledge base of GO annotations collected over the past 20 years can be used as the basis not only for a genomic-biology representation of gene function but also for a more expansive systems-biology representation and its emerging applications to the interpretation of large-scale experimental data.

The benefit of this sort of model is that it is extremely legible: the ontologies and relations are crystal clear, and every annotation points to the piece of scientific evidence it is based on. It is ordered, clean, and systematic.

And, in effect, this knowledge graph is what most modern biomedical research is working toward. Even if publications aren’t literally added to an external knowledge graph, researchers use the same set of conceptual tools when designing and analyzing their experiments—relations like upregulation, sufficiency and necessity; classes like biological processes and molecular entities—the stuff of the mechanistic mind. Ontologies and causal models are merely an externalization of the collective knowledge graph implicit in publications and the heads of researchers.

And yet, what does this knowledge graph get us in terms of dynamics and control?

Suppose we were given the ground-truth, molecule-level mechanistic map of some biological system, like a cell, in the form of a directed graph. For instance, imagine this map as a massive, high-resolution gene regulatory network describing the inner-workings of the cell.

It’s not immediately clear how we’d use this exact map to control the cell’s behavior, let alone the behavior of larger systems (e.g., a tissue composed of multiple cells).

One idea is to take the network and model the molecular kinetics as a system of differential equations, and use this to simulate the cell at the molecular level. This has already been tried for a minimal bacterial cell:

We present a whole-cell fully dynamical kinetic model (WCM) of JCVI-syn3A, a minimal cell with a reduced genome of 493 genes that has retained few regulatory proteins or small RNAs… Time-dependent behaviors of concentrations and reaction fluxes from stochastic-deterministic simulations over a cell cycle reveal how the cell balances demands of its metabolism, genetic information processes, and growth, and offer insight into the principles of life for this minimal cell.

In theory, once you’ve built a spatially resolved, molecule-level simulation of the minimal cell, you then move up to a full-fledged cell, then multiple cells, and so on. Eventually you’ll arrive at a perfect simulation of any biological system, which you can do planning over.

The most significant problem you face, obviously, is computational limitations. Therefore, it seems a perfect map of the biological territory wouldn’t make for a good dynamics model.

However, intuitively, it appears that an exact map should give us the ability to predict the dynamics of the cell and, as a result, control it. Even if simulation isn’t possible, shouldn’t understanding the system at a fine-grain level necessarily give us coarser-grain maps that can be used to predict and control the system’s dynamics at a higher level?

Unfortunately, the answer is no. The issue, it turns out, is that the mechanistic mind’s demand for dynamics model legibility led the model to capture the biological system’s dynamics at the wrong level of analysis using the wrong conceptual primitives.

This is the fundamental flaw in the mechanistic mind’s translational conceit: mistaking advances in mechanistic, human-legible knowledge of smaller and smaller parts of biological systems for advances in predicting the behavior of (and, consequently, controlling) the whole biological systems those parts comprise.

But we needn’t resign ourselves to biomedical stagnation; there are alternative forms of dynamics models which are more suitable for control. Understanding them will require a brief foray into the history of machine learning.

Three Stories from the History of Machine Learning

By tracing the development of machine learning methods applied to three problem domains—language generation and understanding, game-playing agents, and autonomous vehicles—we will develop an intuition for what direction biomedical dynamics models must head in to end biomedical stagnation.

The Bitter Lesson

To make a long story short, these three AI problem domains, and countless others, have undergone a similar evolution, what reinforcement learning pioneer Richard Sutton calls the “bitter lesson” of AI research:

The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin. The ultimate reason for this is Moore’s law, or rather its generalization of continued exponentially falling cost per unit of computation. Most AI research has been conducted as if the computation available to the agent were constant (in which case leveraging human knowledge would be one of the only ways to improve performance) but, over a slightly longer time than a typical research project, massively more computation inevitably becomes available. Seeking an improvement that makes a difference in the shorter term, researchers seek to leverage their human knowledge of the domain, but the only thing that matters in the long run is the leveraging of computation. These two need not run counter to each other, but in practice they tend to. Time spent on one is time not spent on the other. There are psychological commitments to investment in one approach or the other. And the human-knowledge approach tends to complicate methods in ways that make them less suited to taking advantage of general methods leveraging computation…

The bitter lesson is based on the historical observations that 1) AI researchers have often tried to build knowledge into their agents, 2) this always helps in the short term, and is personally satisfying to the researcher, but 3) in the long run it plateaus and even inhibits further progress, and 4) breakthrough progress eventually arrives by an opposing approach based on scaling computation by search and learning. The eventual success is tinged with bitterness, and often incompletely digested, because it is success over a favored, human-centric approach.

That is, in the long run (i.e., in the data and compute limit), general machine learning methods invariably outperform their feature-engineered counterparts—whether those “features” are the data features, the model architecture, the loss function, or the task formulation, all of which are types of inductive biases that place a prior on the problem’s solution space.

General methods win out because they meet the territory on its own terms:

The second general point to be learned from the bitter lesson is that the actual contents of minds are tremendously, irredeemably complex; we should stop trying to find simple ways to think about the contents of minds, such as simple ways to think about space, objects, multiple agents, or symmetries. All these are part of the arbitrary, intrinsically-complex, outside world. They are not what should be built in, as their complexity is endless; instead we should build in only the meta-methods that can find and capture this arbitrary complexity. Essential to these methods is that they can find good approximations, but the search for them should be by our methods, not by us. We want AI agents that can discover like we can, not which contain what we have discovered. Building in our discoveries only makes it harder to see how the discovering process can be done.

As Sutton notes, this lesson is a hard one to digest, and it’s probably not immediately apparent how it applies to biomedicine. So, let’s briefly examine three example cases of the bitter lesson that can act as useful analogs for the biomedical problem domain. In particular, I hope to highlight cases where machine learning methods superficially take the bitter lesson to heart, only to be superseded by methods that more fully digest it.

Language Understanding, Generation, and Reasoning

Cyc has never failed to be able to fully capture a definition, equation, rule, or rule of thumb that a domain expert was able to articulate and communicate in English. — Parisi and Hart, 2020

ML can never give an explicit step-by-step explanation of its line of reasoning behind a conclusion, but Cyc always can. — Cyc white paper, 2021

Here’s an abbreviated history of how AI for language progressed over the past 50 years:

Neural networks experienced a brief moment of hope in the early 60’s, but this was quickly dashed. Neural networks became unfashionable.

From then onward, the dominant approach to language-based reasoning became symbolic AI, which is modeled after “how we think we think”, to use Sutton’s phrase. This approach was realized in, for instance, the expert systems which became popular in the 70’s and 80’s—and it is still being pursued by projects like Cyc.

In the 90’s and 00’s, simple statistical methods like Naive Bayes and random forest began to be applied to language sub-tasks like semantic role labeling and word sense disambiguation; symbolic logic still dominated all complex reasoning tasks. Neural networks were outside the mainstream, but a small group of researchers continued working on them, making important breakthroughs.

But by the early 2010’s, neural networks started to take off, demonstrating state of the art performance on everything from image recognition to language translation (even though in some cases the solutions had been around for more than twenty years, waiting for the right hardware and data to come along). However, these neural networks still used specialized architectures, training procedures, and datasets for particular tasks.

Finally, in the past five years, Sutton’s bitter lesson has been realized: a single, general-purpose neural network architecture, fed massive amounts of messy, varied data, trained using very simple loss functions and lots of compute, has come to dominate not only all language tasks, but also domains like vision and speech—part of the “ongoing consolidation in AI”:

You can feed it sequences of words. Or sequences of image patches. Or sequences of speech pieces. Or sequences of (state, action, reward) in reinforcement learning. You can throw in arbitrary other tokens into the conditioning set—an extremely simple/flexible modeling framework.

The neural network naysayers were proven wrong and the believers vindicated: in 2022, massive language models are resolving Winograd schemas at near-human levels and explaining jokes—and on many tasks surpassing average human performance. And yet, they’re only able to do this because we do not hard-code our understanding of semantic ambiguity or humor into the model, nor do we directly ask the model to produce such behaviors.

Game-Playing Agents

The evolution of AI for game-playing agents parallels that of AI for language in many ways. To take chess as an example, we can divide it into three eras: the DeepBlue and Stockfish era, the AlphaZero era, and the MuZero era.

Both DeepBlue and Stockfish use massive brute-force search algorithms with hand-crafted evaluation functions (i.e., the function used during search to evaluate the strength of a position) based on expert knowledge, and a host of chess-specific heuristics. Chess researchers were understandably upset when these “inelegant” methods began to defeat humans. As Sutton tells it:

In computer chess, the methods that defeated the world champion, Kasparov, in 1997, were based on massive, deep search. At the time, this was looked upon with dismay by the majority of computer-chess researchers who had pursued methods that leveraged human understanding of the special structure of chess. When a simpler, search-based approach with special hardware and software proved vastly more effective, these human-knowledge-based chess researchers were not good losers. They said that “brute force” search may have won this time, but it was not a general strategy, and anyway it was not how people played chess. These researchers wanted methods based on human input to win and were disappointed when they did not.

Whereas AlphaZero dispenses with all the chess-specific expert knowledge and instead uses a neural network to learn a value function atop a more general search algorithm. Furthermore, it learns this all by training purely through self-play (that is, by playing with itself—it isn’t given a bank of past chess games to train on, only the rules of the game). It is so general an architecture that it can learn to play not only chess, but also Go and Shogi.

AlphaZero not only handily beats Stockfish, but does so using less inference-time compute, making it arguably more human-like in its play than its “brute-force” predecessors:

AlphaZero searches just 60,000 positions per second in chess and shogi, compared with 60 million for Stockfish…AlphaZero may compensate for the lower number of evaluations by using its deep neural network to focus much more selectively on the most promising variations—arguably a more human-like approach to searching, as originally proposed by Shannon.

Finally, MuZero takes AlphaZero and makes it even more general: it learns to play any game without being given the rules. Instead, MuZero bootstraps a latent dynamics model of the environment through interacting with it, and uses this learned dynamics model for planning. This latent planning allows it to learn tasks, like Atari games, which were previously not seen as the province of model-based reinforcement learning because of their high state-space dimensionality. In fact, MuZero is so general a learning method that it can be applied to non-game tasks like video compression. If you can give it a reward signal, MuZero will learn it.

Autonomous Vehicles

Autonomous vehicles (AV) are by far the most instructive of the analogs, because the battle is still being played out in real time and many people are mistaken about which approach will win. Additionally, the psychological barriers to arriving at the correct approach (namely, an inability to abstract) are similar to the psychological barriers end-to-end machine learning approaches in biomedicine will face.

We can divide modern AV approaches into two basic camps: the feature-engineering camp of Waymo et al., and the end-to-end approach of Comma.ai and Wayve (and, to a lesser extent, Tesla). Both camps use “machine learning”, but their mindsets are radically different.

The classical approach to AV—pursued by Waymo, Cruise, Zoox, et al. (all the big players in AV that you’ve likely heard of)—decomposes the task of driving into a set of human-understandable sub-problems: perception, planning, and control. By engineering human knowledge into the perception stack, this approach hopes to simplify planning and control.

These companies use fleets of vehicles, decked out with expensive sensors, to collect massive amounts of high-resolution data, which they use to build high-definition maps of new environments:

To create a map for a new location, our team starts by manually driving our sensor equipped vehicles down each street, so our custom lidar can paint a 3D picture of the new environment. This data is then processed to form a map that provides meaningful context for the Waymo Driver, such as speed limits and where lane lines and traffic signals are located. Then finally, before a map gets shared with the rest of the self-driving fleet, we test and verify it so it’s ready to be deployed.

Just like a human driver who has driven the same road hundreds of times mostly needs to focus only on the parts of the environment that change, such as other vehicles or pedestrians, the Waymo Driver knows permanent features of the road from our highly-detailed maps and then uses its onboard systems to accurately perceive the world around it, focusing more on moving objects. Of course, our streets are also evolving, so if a vehicle comes across a road closure or a construction zone that is not reflected in a map, our robust perception system can recognize that and make adjustments in real-time.

But mapping only handles the static half of the perception problem. While on the road, the AV must also perceive and interact with active agents in the environment—cars, pedestrians, small animals. To do this, the AV’s are trained on a variety of supervised machine learning perception tasks, like detecting and localizing moving agents. However, sometimes localization and detection aren’t enough—for instance, you might need to train the model to estimate pedestrian poses and keypoints, in order to capture the subtle nuances that are predictive of pedestrian behavior:

Historically, computer vision relies on rigid bounding boxes to locate and classify objects within a scene; however, one of the limiting factors in detection, tracking, and action recognition of vulnerable road users, such as pedestrians and cyclists, is the lack of precise human pose understanding. While locating and recognizing an object is essential for autonomous driving, there is a lot of context that can go unused in this process. For example, a bounding box won’t inherently tell you if a pedestrian is standing or sitting, or what their actions or gestures are.

Key points are a compact and structured way to convey human pose information otherwise encoded in the pixels and lidar scans for pedestrian actions. These points help the Waymo Driver gain a deeper understanding of an individual’s actions and intentions, like whether they’re planning to cross the street. For example, a person’s head direction often indicates where they plan to go, whereas a person’s body orientation tells you which direction they are already heading. While the Waymo Driver can recognize a human’s behavior without using key points directly using camera and lidar data, pose estimation also teaches the Waymo Driver to understand different patterns, like a person propelling a wheelchair, and correlate them to a predictable future action versus a specific object, such as the wheelchair itself.

The resulting perception stack produces orderly and beautiful visualizations: the maps are clean, every pedestrian and car is localized, and objects have an incredible level of granularity.

Then comes the task of planning and control: the AV is fed this human-engineered scene representation (i.e., the high-definition map with agents, stop signs, and other objects in it) and is trained to predict and plan against the behavior of other agents in this feature space, all while obeying a set of hand-engineered rules written atop the machine-learned perception stack—stop at stop signs, yield to pedestrians in the crosswalk, don’t drive above the speed limit, etc.

However, this piecewise approach has one fatal flaw, claim the end-to-end camp: it relies on a brittle, human-engineered feature space. Abstractions like “cone” and “pedestrian pose” might suffice in the narrow, unevolving world of a simulator or a suburb of Phoenix, but these abstractions won’t robustly and completely capture all the dynamic complexities of real-world driving. Waymo et al. may use machine learning, but they haven’t yet learned the bitter lesson.

Rather, as with other complex tasks, the ultimate solution will be learned end-to-end—that is, directly from sensor input to motion planning, without intermediate human-engineered abstraction layers—thereby fully capturing the complexity of the territory. Wayve (not to be confused with Waymo) founder Alex Kendall explicitly cites large language models and game-playing agents as precedents for this:

We have seen good progress in similar fields by posing a complex problem as one that is able to [be] modeled end-to-end by data. Examples include natural language processing with GPT-3, and in games with MuZero and AlphaStar. In these problems, the solution to the task was sufficiently complex that hand-crafted abstraction layers and features were unable to adequately model the problem. Driving is similarly complex, hence we claim that it requires a similar solution.

The solution we pursue is a holistic learned driver, where the driving policy may be thought of as learning to estimate the motion the vehicle should conduct given some conditioning goals. This is different to simply applying increasing amounts of learning to the components of a classical architecture, where the hand-crafted interfaces limit the effectiveness of data.

The key shift is reframing the driving problem as one that may be solved by data. This means removing abstraction layers used in this architecture and bundling much of the classical architecture into a neural network…

Likewise, Comma.ai founder George Hotz is pursuing the end-to-end approach and thinks something like MuZero will ultimately be the solution to self-driving cars: rather than humans feature-engineering a perception space for the AV to plan over, the AV will implicitly learn to perceive the control-relevant aspects of the scene’s dynamics (captured by a latent dynamics model) simply by training to predict and imitate how humans drive (this training is done offline, of course).

Hotz paraphrased Sutton’s bitter lesson as the reason for his conviction in the end-to-end approach:

The tasks themselves are not being learned. This is feature-engineering. It’s slightly better feature-engineering, but it still fundamentally is feature-engineering. And if the history of AI has taught us anything, it’s that feature-engineering approaches will always be replaced and lose to end-to-end approaches.

And the results speak for themselves: the feature-engineering camp has for years been claiming they’d have vehicles (without safety drivers) on the road soon, yet only recently started offering such services in San Francisco (which has already caused some incidents). Conversely, the end-to-end approach, which uses simple architectures and lots of data, is already demonstrating impressive performance on highways and in busy urban areas (even generalizing to completely novel cities), and is slowly improving—the AVs are even starting to implicitly understand stop signs, despite never being explicitly trained to do so.

The Full Spectrum of Biomedical Research Ethoses

These three analogs all share the same lesson: there’s a direct tradeoff between human-legibility and predictive performance, and in the long-run, low-inductive-bias models fed lots of data will win out.

But it’s unclear if this lesson applies to biomedicine. Is biology so complex and contingent that its dynamics can only be captured by large machine learning models fed lots of data? Or can biomedical dynamics only be learned data-efficiently by humans reasoning over data in the language of the mechanistic mind?

Throwing out the scientific abstractions that have been painstakingly learned through decades of experimental research seems like a bad idea—we probably shouldn’t ignore the concept of genes and start naively using raw, unmapped sequencing reads. But, on the other hand, once we have enough data, perhaps these abstractions will hinder dynamics model performance—raw reads contain information (e.g., about transcriptional regulation) that isn’t captured by mapped gene counts. These abstractions may have been necessary to reason our way through building tools like sequencing, but now that they’ve been built, we must evolve beyond these abstractions.

We can situate one’s views on this issue along a spectrum of research ethoses. On one pole is the mechanistic mind, on the other is the end-to-end approach, or what we might call the “totalizing machine learning” approach to biomedical research. These two poles can be contrasted as follows:

Mechanistic mindTotalizing ML mindPrimary valueHuman understandingControlAsks the question(s):How? Why?What do I do to control this system?Model primitiveDirected graph model (transparent, legible)Neural network model (opaque, illegible)Primary dynamics model useReason overPredict, controlProblem settingSchismatic: N separate biological systems to map, at M separate scalesConvergent: 1 general, unified biological dynamics problemProblem solutionPiecewiseEnd-to-endEpistemic moodOrdered, brittleAnarchic (anything goes), messy, robustFailure modeMechanism fetish, human cognitive limitsInterpolation without extrapolation, data-inefficiency

I’ll argue that in the long-run, the totalizing machine learning approach is the only way to solve biomedical dynamics. We will now explain why that is, and attempt to envision what that future might look like.

4. The Scaling Hypothesis of Biomedical Dynamics

Someone interested in machine learning in 2010 might have read about some interesting stuff from weirdo diehard connectionists in recognizing hand-written digits using all of 1–2 million parameters, or some modest neural tweaks to standard voice-recognition hidden Markov models⁠. In 2010, who would have predicted that over the next 10 years, deep learning would undergo a Cambrian explosion causing a mass extinction of alternative approaches throughout machine learning, that models would scale up to 175,000 million parameters, and that these enormous models would just spontaneously develop all these capabilities? — Gwern, The Scaling Hypothesis

Science is essentially an anarchic enterprise: theoretical anarchism is more humanitarian and more likely to encourage progress than its law-and-order alternatives. — Paul Feyerabend, Against Method

The scaling hypothesis is the totalizing machine learning mind’s solution to biomedical dynamics.

At the highest level, the scaling hypothesis for biomedical dynamics is the claim that training a massive, generative, low-inductive-bias neural network model on large volumes of messy, multi-modality biological data (text, omics, images, etc.), using simple training strategies, will produce an arbitrarily accurate biomedical dynamics prediction function.

Just as large models trained on human language data learn to approximate the general function “talk (and, eventually, reason) like a human”, large models trained on biological data will learn to approximate the general function “biomedical dynamics”—without needing to simulate the true causal structure of the underlying biological system. Show a model enough instances of a system’s dynamics, and it will learn to predict them.

To give a simple example: if one wanted to develop a model of single-cell dynamics, they’d take a large neural network (the architecture doesn’t matter, as long as it’s general and has low inductive bias) and feed it all the scientific publications and experimental and observational omics data they could get their hands on (how you feed these data in doesn’t matter much either—you could probably model them as text tokens, though you’d need to treat them as a set, not a sequence). These data contain information about the correlational structure of single-cell state space and the dynamics, both action-conditional and purely observational, on this state space. By masking these data and training the model to predict the masked tokens, the model will implicitly learn the dynamics underlying the data. This knowledge will be stored in the weights of the model and can be accessed by querying it. This learned dynamics model can then be used for single-cell planning and control.

Then extend this approach to all forms and scales of biological data.

Why The Scaling Hypothesis Wins Out

In the short-term, feature-engineering approaches will have the advantage in biomedical dynamics, as they did in other domains—for instance, the highest-performing single-cell dynamics model of today would likely be a smaller machine-learning model with lots of inductive biases (e.g., a graph inductive bias to mirror the graph-like structure of the gene regulatory networks generating the single-cell omics data), trained on heavily massaged data, not a massive, blank-slate neural network. But in the long-run, general, low-inductive-bias neural networks will win out over narrower, specialized dynamics models, for three reasons: dynamics incompressibility, model subsumption, and data flows.

Dynamics Incompressibility and Generality

The richness and complexity of biomedical dynamics exceed human cognitive capacity, and can therefore not be fully captured by human-legible models. Human-legible models must compress these dynamics, projecting them onto lower-dimensional maps written in the language of human symbols, which necessarily involves a tradeoff with dynamics prediction performance—simply put, legibility is a strong inductive bias. This tradeoff also applies to machine learning models insofar as they are legibility-laden.

Whereas large neural networks do not make this tradeoff between human understanding and predictive performance; they meet the territory on its own terms. By modeling raw data of all modalities, they can approximate any biological function without concern for notions of mechanism or the strictures of specialized, siloed problem domains. They can pick up on subtle patterns in the data, connections across scales that human-biased models are blind to. Feed them enough data, and they’ll find the predictive signal.

Subsumption

The mechanistic mind will continue to produce artifacts explaining biomedical dynamics—publications, figures, knowledge bases, statistics, labeled datasets, etc. Large multimodal neural networks, because they’re agnostic about what sorts of data they take in, will ingest these mechanistic models, improving dynamics modeling. This is how the neural network will initially bootstrap its understanding of biological dynamics, coasting off the modeling efforts of the mechanistic mind instead of learning a world model from scratch, and then moving beyond it.

However, the converse is not true. The mechanistic mind is incapable of subsuming the totalizing machine learning mind. A neural network’s weights can model an ontology, but not vice versa.

Data Flows and Messiness-Tolerance

Large neural networks will ingest the massive amounts of public biological data coming out in the form of omics, images, etc. These data are messy, but the models are robust to this: they tolerate partially observable, noisy, variable-length, poorly-formatted data of all modalities. The model’s performance is directly indexed against these data flows, which will only become stronger over time, due to cost declines in the tools that generate them.

Whereas narrow, specialized methods that rely on cleaner, more structured data will not experience the same tailwind, for they simply cannot handle the messy data being produced. As Daphne Koller, founder of Insitro (which develops such narrow, specialized models), says:

We also don’t do enough as a community to really establish a consistent set of best practices and standards for things that everybody does. That would be hugely helpful, not only in making science more reproducible but also in creating a data repository that is much more useful for machine learning.

I mean, if you have a bunch of data collected from separate experiments with a separate set of conditions, and you try to put them together and run machine learning in that, it’s just going to go crazy. It’s going to overfit on things that have nothing to do with the underlying biology because those are going to be much more predictive and stronger signals than the biology that you’re trying to interrogate. So I think we need to be doing better as a community to enable reproducible science.

Because of the nature of academia, these coordination problems won’t be solved. Therefore, Koller and others will have to rely on smaller datasets, often created in-house:

[W]hile access to large, rich data sets has driven the success of machine learning, such data sets are still rare in biology where data generation remains largely artisanal. By enabling the production of massive amounts of biological data, the recent advancements in cell biology and bioengineering are finally enabling us to change this.

It is this observation that lies at the heart of insitro. Instead of relying on the limited existing “found” data, we leverage the tools of modern biology to generate high-quality, large data sets optimized for machine learning, allowing us to unleash the full potential of modern computational approaches.

However, these in-house datasets will always be dwarfed by messy public data flows. The method that wins the race to solve biomedical dynamics must be able to surf this coming data deluge.

An Empirical Science of Biomedical Dynamics

All these arguments seem rather faith-based. Yes, given infinite data, the end-to-end approach will learn a superior biomedical dynamics model. But in practice, we don’t have infinite data.

But we can qualify and quantify our faith in the scaling hypothesis, because the performance of large neural networks obeys predictable functions of data and computation. This is the secret of the scaling hypothesis: scaling laws.

More precisely, scaling laws refer to the empirical finding that a neural network’s error on a task smoothly decreases as a power-law function of the size of that neural network and how much data you train it on (when one is not bottlenecked by the other).

These power-law trends can be visualized as straight lines on log-log plots, where each increment on the x- and y-axis denotes a relative change (i.e., a percentage change), rather than an absolute change. The exponent of the power law is equivalent to the slope of the scaling trendline on this log-log plot.

Scaling laws are relatively invariant across model architectures, tasks, data modalities, and orders of magnitude of compute (which likely reflects deep truths about the nature of dynamical systems).

They were originally found in language. Then they were found to apply to other modalities like images, videos, and math. Then they were found to even generalize to the performance of game-playing agents on reinforcement learning tasks. They show up everywhere.

We might say that scaling laws are a quantitative framework for describing the amount of data and model size needed to approximate the dynamics of any (evolved) complex system using a neural network.

This is the upshot of the totalizing machine learning approach to biomedicine: by treating the biomedical control problem as a dynamics modeling problem, and learning these dynamics through large neural networks, biomedical progress becomes a predictable function of how much data and computation we feed these neural networks.

Objections to the Scaling Hypothesis

However, we face two problems if we naively try to apply the scaling hypothesis to biomedicine in this way:

1. Biomedical data acquisition has a cost. Unlike chess or Atari, you can’t simply collect near-infinite data, particularly experimental data, from your target biological system, humans. We don’t have access to a perfect simulator. In the real world, we must do much of our research in non-human model systems.

2. In biomedicine, data doesn’t miraculously fall from the heavens like manna. The keystone of biomedical research is experimentation—we must actively seek out data to update our models. Brute-forcing our way to a dynamics model by collecting random data would be woefully inefficient.

By investigating these two obstacles to the scaling hypothesis, and thinking more seriously about how training data is acquired, we will arrive at an empirical law of biomedical progress.

5. Biocompute Bottleneck

A problem besets all approaches to learning biomedical dynamics: human experimental data is hard to come by. This poses a particular problem for the scaling hypothesis. How are we supposed to learn a dynamics model of our target biological system if we can’t acquire data from it?

Of Mice, Men, and Model Systems

We must instead learn a dynamics model through experimentation in model systems.

The difficulty of working with model systems can best be understood through analogies to the AI problem domains we reviewed earlier.

One analogy is to the autonomous vehicle domain: in biomedicine, like in autonomous vehicles, you can’t train your driving policy online (i.e., in a live vehicle on the road), so you must passively collect data of humans driving correctly and use this to train a policy offline via imitation. The only problem is that in biomedicine we aren’t trying to imitate good driving, but train a policy that corrects driving from off-road (i.e., disease) back on-road (i.e., health); therefore, we don’t have many offline datasets to train on. So, we have to settle for training our policy on a different vehicle (a mouse) in a different terrain (call it “mouse-world”) that obeys different dynamics (and for many diseases a good mouse model doesn’t even exist), and then attempt to transfer this policy to humans. The benefit of mouse-world is that you can crash the cars as often as you want; the downside is that you’re learning dynamics in mouse-world, which don’t transfer well to humans.

Yet even if we were unconstrained by ethics and regulation, and therefore able to experiment on humans, we would still be limited by speed and the difficulty of isolating the effects of our experiments. Mice and other model systems are simply quicker and easier to experiment on.

This is the defining tradeoff of model systems in biomedicine: external validity (i.e., how well the results generalize to humans) necessarily comes at the expense of experimental throughput (i.e., how many experiments one can run in a given period of time, which is determined by some combination of ethical, regulatory, and physio-temporal constraints), and vice versa.

Thus, model systems lie along a Pareto frontier of throughput vs. external validity.

One experiment on a human undoubtedly contains more information about human biomedical dynamics than one experiment on a mouse. However, the question is how much more—is a marginal human experiment equivalent to 10 marginal mouse experiments, or 1000?

By creating a framework to understand this tradeoff, we will discover that the quality of our model systems directly upper-bounds the rate of biomedical progress.

Scaling Laws for Transfer

The scaling laws framework offers a language for answering such questions. To make data from different model systems commensurable, we must introduce the idea of scaling laws for transfer; these scaling laws quantify how effectively a neural network trained one one data distribution can “port” its knowledge (as encoded in the neural network’s weights) to a new data distribution.

For instance, one could pre-train a neural network on English language data and “transfer” this knowledge to the domain of code by fine-tuning the model on Python. Or one could pre-train on English and then transfer to Chinese, Spanish, or German language data. Or one could pre-train on images and transfer to text (or vice versa).

The usefulness of this pre-training is quantified as the “effective data transferred” from source domain A to target domain B, which captures how much pre-training on data from domain A is “worth” in terms of data from domain B—in effect, a kind of data conversion ratio. It is defined as “the amount of additional fine-tuning data that a model of the same size, trained on only that fine-tuning dataset, would have needed to achieve the same loss as a pre-trained model,” and is quantified with a transfer coefficient, αT, that measures the directed similarity (which, like KL divergence, isn’t symmetric) between the two data distributions. We can think of this transfer coefficient as a measure of the external validity from domain A onto domain B.

The analogy to biomedicine is quite clear: we pretrain our dynamics model on a corpus of data written in the “mouse” language or “in vitro human cell model” language, and then we attempt to transfer this general knowledge of language dynamics by fine-tuning the model on data from the “human” language domain. The effective data transferred from these model systems is equal to how big a boost pre-training on them gives our human dynamics model, in terms of human-data equivalents.

However, in biomedicine, we often don’t have much human data to fine-tune on. We can easily collect large pre-training datasets from non-human model systems like mice and in vitro models, but human data, especially experimental data, is comparatively scarce. This puts us in the “low-data regime”, where our total effective data, DE, reduces to whatever effective data we can transfer from the source domain, DT (e.g., mouse or in vitro models):

Pre-training effectively multiplies the fine-tuning dataset, DF , in the low-data regime. In the low data regime the effective data from transfer is much greater than the amount of data fine-tuned on, DT≫DF. As a result, the total effective data, DE=DF+DT≈DT.

Therefore, we must effectively zero-shot transfer our pre-trained model to the human domain with minimal fine-tuning—meaning our total effective data is whatever we can squeeze out of pre-training on non-human data. That is, the majority of our model’s dynamics knowledge must come from the non-human domain.

The human dynamics model’s loss scales as a power-law function of the amount of non-human data we transfer to it, but where the exponent is the product of the original human-data scaling exponent, αD, and the transfer coefficient exponent between the domains, αT. That is, training on non-human data reduces the original human-data scaling law by a factor 1/αT, which can be visualized as the slope being reduced by this factor on a log-log plot. For instance, if the mouse to human transfer coefficient is 0.5, then the reduction in loss from a 1000x increase in human data is equivalent to the reduction in loss from a ~1,000,000x (divided by some constant) increase in mouse data.

This has a startling implication: under this model, small decreases in the transfer coefficient of a model system require exponential increases in dataset size to offset them.

Jack Scannell, one of the coiners of the term “Eroom’s law”, found something consistent with this in his paper on the role of model systems in declining returns to pharmaceutical R&D. The paper presents a decision theoretic analysis of drug discovery, in which each stage of the pipeline involves culling the pool of drug candidates using some instrument, like a model system; the pipeline can be thought of as a series of increasingly fine filters that aim to let true positives (i.e., drugs that will have the desired clinical effect in humans) through, while selecting out the true negatives. Scannell finds that small changes in the validity of these instruments can have large effects on downstream success rates:

We find that when searching for rare positives (e.g., candidates that will successfully complete clinical development), changes in the predictive validity of screening and disease models that many people working in drug discovery would regard as small and/or unknowable (i.e., an 0.1 absolute change in the correlation coefficient between model output and clinical outcomes in man) can offset large (e.g., 10-fold, even 100-fold) changes in models’ brute-force efficiency…[and] large gains in scientific knowledge.

Therefore, in both the dynamics transfer and drug pipeline context, model external validity is critical. (Yes, Scannell’s model is extremely simple, and it’s debatable if the scaling laws for transfer model applies in the manner described—at worst, the math doesn’t apply exactly but the general point about external validity still holds.)

Add to this the fact that there’s likely an upper bound on how much dynamics knowledge can be transferred between model systems and humans—mice and humans are separated by 80 million years of evolution, after all—and it seems to suggest that running thousands, or even millions, of experiments on our current model systems won’t translate into a useful human biomedical dynamics model.

bioFLOP Benchmark

But we can put a finer point on our pessimism. I will claim that the rate of improvement in our biomedical dynamics model, and therefore the rate of biomedical progress, is directly upper-bounded by the amount of external validity-adjusted “biocomputation” capacity we have access to.

By “biocomputation”, I do not mean genetically engineering cells to compute X-OR gates using RNA. Rather, I mean biologically native computation, the information processing any biological system does. When one runs an experiment on (S×A→S′) or passively observes the dynamics of (S→S′) a biological system, the data one collects are a product of this system’s biocomputation. In other words, biological computation emits information.

All model systems are therefore a kind of “biocomputer”, and their transfer coefficients represent a mutual information measure between the biocomputation they run and human biocomputation. The higher this transfer coefficient, the more information you can port from this biocomputer to humans. By combining transfer coefficients and measures of experimental throughput, we can develop a unified metric for comparing the biocomputational capacity of different model systems.

Let’s arbitrarily define the informational value (which you can think of in terms of reduction in human dynamics model loss) of one marginal experiment on a human as 1 human-equivalent bioFLOP. This will act as our base. The informational value of one marginal experiment on any model system is worth some fraction of this human-equivalent bioFLOP, and is a function of the model system’s transfer coefficient to humans.

Therefore, the human-equivalent bioFLOPS (that is, the biocomputational capacity per unit time) of a model system is the product of its experimental throughput (how many experiments you can run on it per unit time) and its biocomputational power (as measured in fractions of a human-equivalent bioFLOP per experiment—i.e., how much human-relevant information a single experiment on it returns).

With this framework in place, we can compute the total human-equivalent biocomputational capacity available, given a set of model systems, their transfer coefficients, and their experimental throughputs. This, in theory, would tell us how many experiments we’d need to run on these model systems, and how long it would take, to achieve some level of dynamics model performance on a biomedical task.

If we were to run this calculation on our current model systems, we’d come to the sobering conclusion that our biocomputation capacity is incredibly limited compared to the complexity of the target biological systems we’re attempting to model. We’re bioFLOP bottlenecked.

To make biocomputational limits more intuitive, and to understand why we must increase biocomputation to accelerate biomedical progress, let’s work our way through an analogy to game-playing agents.

Biocomputation Training Analogy

Suppose you’re trying to train a MuZero-like agent to do some task in the real world. To solve the task, the agent must build up a latent dynamics model of it. The catch is you’re only able to train the agent in a simulator.

You’re given three simulator options. Each runs on a purpose-built chip, specifically designed to run that simulator.

Simulator A is extremely high-fidelity, and accurately recapitulates the core features of the real-world environment the agent will be testing in. The downside is that this simulator operates extremely slowly and the chips to run it are incredibly expensive.

Simulator B is also high-fidelity, but the agent trains in a different environment from the one it will be testing in, which uses a different physics engine. The upside is that this simulator is relatively quick and the chips to run it are cheap and abundant.

Simulator C is a low-fidelity version of Simulator A. In principle, it operates on the same physics engine, but in practice the chip doesn’t have the computational power to fully simulate the environment, and therefore only captures limited features of it. However, the chips are incredibly cheap (though they have a piddling amount of compute power) and they are fast (for the limited amount of computation they do).

During training, the state output of the simulator is rendered on a monitor and displayed to the agent. (This rendering used to be incredibly expensive, but it’s becoming much cheaper.) The agent observes the state, and chooses an action which is then fed to the simulator. The simulator then updates its state. This constitutes one environmental interaction.

As you probably realize, none of the simulators can train an effective agent in a reasonable amount of time. In all three cases, training is bottlenecked by high-quality compute.

Mistaken Moore’s Law

Scopes and scalpels, the two tool classes we covered in our earlier tool progress review, are, in a sense, peripherals—the monitors and mice to our computers. Let’s return to the most famous graph in genome sequencing, in which sequencing progress is compared to Moore’s law:

On one level, this comparison is purely quantitative: the decline in genome sequencing cost has outpaced Moore’s law. But the graph also invites a more direct analogy: sequencing will be to the biomedical revolution as semiconductors were to the information technology revolution.

This comparison is wrong and ironic, because despite our peripherals improving exponentially, most human-relevant biocomputation is still done on the biological equivalent of vacuum tubes. (No, running them in “the cloud” won’t increase their compute capacity.) Through an ill-formed analogy, many have mistaken the biological equivalents of oscilloscopes, screen pixel density, and soldering irons as the drivers of biomedical progress. Much misplaced biomedical optimism rests on this error.

Peripherals certainly matter—it would have been hard to build better microchips if we couldn’t read or perturb their state—but they alone do not drive a biomedical revolution. For that, we need advances in compute.

The only way to increase compute capacity in biomedicine is to push the biocomputer Pareto frontier along the external validity or throughput dimensions. However, the exponential returns to improved external validity far outweigh any linear returns we could achieve by increasing throughput. Therefore, to meaningfully increase biocompute, we must build model systems with higher external validity. Scopes and scalpels will drive the Moore’s law of biomedicine only insofar as they help us in this task, which is the subject of the third and final tool class: simulacra.

Growing Better Biocomputers

What I cannot create, I do not understand. — Richard Feynman

Our best hope for increasing human-equivalent bioFLOPS is to build physical, ex vivo models of human physiology, which we’ll call “bio-simulacra” (or “simulacra”, for short), using induced pluripotent stem cells. Simulacra are the subset of model systems that aren’t complete organisms, but instead act as stand-ins for them (or parts of them). The hope is that they can substitute for the human subjects we’d otherwise like to experiment on.

For instance, in vitro cell culture is an extremely simple type of simulacrum. However, it’s like the biocomputer for running Simulator C in the metaphor above: it has too little compute to accurately mirror the complexity of complete human physiology. But simulacra could become vastly more realistic, eventually approaching the scale and complexity of human tissues, or even whole organs, thereby increasing their compute power.

Simulacra have been researched for decades under many names (“microphysiological systems”, “organ-on-a-chip”, “organoids”, “in vitro models”, etc.), and dozens of companies have attempted to commercialize them.

Though there have been great advances in the underlying technology—microfluidics, sensors, induced pluripotent stem cells—the resulting simulacra are still disappointingly physiologically dissimilar to the tissues they are meant to model. This can be seen visually: most simulacra are small—because they lack vascularization, which limits the diffusion of nutrients and oxygen—and (quite literally) disorganized.

Much of this slow progress can be attributed to not accurately reconstructing the in vivo milieu of the imitated tissues (i.e., those we wish to build ex vivo). There are many low-hanging fruit here that are beginning to be picked: improving nutritional composition of cell culture media, finding sources of extracellular matrix proteins other than cancer cells, adding biomechanical and electrical stimulation, etc.

All these advances are directionally correct but insufficient. If we hope to grow larger, more realistic simulacra with higher external validity, placing a few cell types in a sandwich-shaped plastic chip won’t cut it. Useful simulacra amount to real human tissues; therefore, they must be grown like real human tissues.

Luckily, nature has given us a blueprint for how to do this: human development. The task of growing realistic simulacra therefore reduces to the task of reverse-engineering the core elements of human development, ex vivo. That is, we must learn to mimic the physiological cues a particular tissue or organ would experience during development in order to grow it. Through this, we will grow simulacra that better represent adult human physiology, thereby increasing these models’ external validity.

Our rate of progress on this task is the main determinant of the rate of biocomputational, and therefore biomedical, progress over the next many decades. To solve the broader, more difficult problem of biomedical control, we must first solve this narrower, more tractable problem of controlling development.

Here’s the interesting thing: reverse-engineering development can be framed as a biomedical control task and therefore is amenable to the scaling laws approach. In effect, the challenge is to use physiological cues to guide multicellular systems into growing toward the desired stable “social” configurations. This is an extremely challenging representation learning and reinforcement learning problem (one might even call it a multi-agent reinforcement learning problem).

6. The Future of Biomedical Research

We have established that biomedical dynamics can be approximated through large neural networks trained on lots of data; the predictive performance of these models is a power-law function of these data; and how much these data improve predictive performance depends on the external validity of the biocomputational system they were generated by.

However, a question remains: how do we select these data? (And is power-law scaling the best we can do?)

Unlike large language models, whose training data is a passive corpus that the model does not (yet) have an active role in generating, in biomedicine we must actively generate data through observation and experimentation. This experimental loop is the defining feature of biomedical research.

This experimental loop is directed by biomedical research agents—currently, groups of human beings. However, in the long-run, this loop will increasingly rely on, and eventually be completely directed by, AI agents. That is, in silico compute will be fully directing the use of biocompute.

Let’s first discuss experimental efficiency. Then we’ll discuss what the handoff from humans to AI agents might look like in the near-term, and what completely AI-directed biomedical research might look like in the long-term.

Active Learning and Experimental Efficiency

Good news: the power-law data scaling we reviewed before is the worst-case scenario, in which data is selected at random. Through data selection techniques, we can achieve superior scaling behavior.

In the case of biomedical experimentation (e.g., in the therapeutics development context), the data selection regime we find ourselves in is active learning: at every time point, we select an experiment to run and query a biocomputer oracle (i.e., physically run the experiment) to generate the data.

The quality of this experimental selection determines how efficiently biocompute capacity is translated, via the experimental loop, into dynamics modeling improvements (and, ultimately, biomedical control). That is, there are many questions one could ask the biocomputer oracle, each of which emits different information about dynamics; the task is to properly sequence this series of questions.

However, biomedical dynamics are incredibly complex, so efficient experimental selection requires planning, especially when multiple types of experiment are possible. Thus, experimental selection amounts to a kind of reinforcement learning task.

Currently, this reinforcement learning task is being tackled by decentralized groups of humans. In the near-future, AI agents will play a greater role in it.

Centaur Research Agents

At first, AI agents will be unable to efficiently run the experimental loop themselves. Therefore, in the translational research context, we’ll first see “centaurs”—human agents augmented with AI tools.

Initially, in the centaur setup, the AI’s main role will be as a dynamics model, which the human will query as an aid to experimental planning and selection. (Up to this point in the essay, this is the only role we’ve discussed machine learning models playing.) For instance, the human could have the AI run in silico rollouts of possible experiments, in order to select the one that maximizes predicted information gain over some time horizon (as evaluated by the human). In this scenario, the human is still firmly in the driver’s seat, determining the value of and selecting experiments—the machine is in the loop, but not directing it.

However, as the AI improves, it will begin to take on more responsibility. For instance, as it develops better multimodal understanding of its internal dynamics model, the AI will help the human analyze and reason through the dynamics underlying the predicted rollouts, via natural language interaction. Humans will become more reliant on the AI’s analysis.

Eventually, because it can reason over far larger dynamics rollouts and has an encyclopedic knowledge of biomedical dynamics to draw on, the AI will develop a measure of predicted experimental information gain (in effect, a kind of experimental “value function”) superior to human evaluation. Not soon after, by training offline on the dataset of human-directed research trajectories it has collected, the AI will develop an experimental selection policy superior to that of humans.

At some point, the AI will begin autonomously running the entire experimental loop for circumscribed, short-horizon tasks. The human will still be dictating the task-space the AI agent operates in, its goals, and the tools at its disposal, but with time, the AI will be given increasing experimental freedom.

The direction this all heads in, obviously, is end-to-end control of the experimental loop by AI agents.

Toward End-to-End Biomedical Research

The behavior of end-to-end AI agents might surprise us.

For instance, suppose the human tasks the AI with finding a drug that most effectively moves a disease simulacrum of a person with genotype G from state A (a disease state) to state B (a state of health). The AI is given a biocompute budget and an experimental repertoire to solve this task.

There are many ways the AI could approach this problem.

For instance, the AI might find it most resource-efficient to first do high-throughput experimental screening of a set of drug candidates (which it selects through in silico rollouts) on very simple in vitro models, and then select a promising subset of these candidates for experimentation in the more complex, target disease simulacra. Then it repeats this loop.

But the agent’s experimental loop could become much more sophisticated, even nested. For instance, the agent might first do in silico rollouts to formulate a set of “hypotheses” about the control dynamics of the area of state space the disease simulacrum lives in. Then it (in silico) spatially decomposes the target simulacrum into a set of multicellular sub-populations, and restricts its focus to the sub-population which is predicted to have the largest effect on the dynamics of the entire simulacrum. It then creates in vitro models of this sub-population, and runs targeted, fine-grained perturbation experiments on them (perhaps selected based on genetic disease association signatures it picked up on in the literature). It analyzes the experimental results, and then revises its model of the simulacrum’s control dynamics, restarting this sub-loop. After it has run this sub-loop enough to sufficiently increase its confidence in its dynamics model, it tests the first set of drug candidates on actual instances of the disease simulacrum.(Obviously none of this planning is explicit. It takes place in the weights of the neural networks running the policy and dynamics models.)

As the agent becomes more advanced and biomedical tasks become longer-horizon and more open-ended, these experimental loops will become increasingly foreign to us. Even if we peer inside the weights, we likely won’t be able to express the experimental policy in human-language.

This will become even truer as the agent begins to self-scaffold higher-order conceptual tools in order to complete its tasks. For instance, given an extremely long-horizon task, like solving a disease, the agent might discover the concept of “technological bottlenecks”, and begin to proactively seek out bottlenecks (like that of biocompute, and those we can’t yet foresee) and work toward alleviating them. To solve these bottlenecks, the agent will construct and execute against technology trees, another concept it might develop.

Yet to traverse these technology trees, the agent must build new experimental tools, just as humans do. This might even involve discovering and establishing completely new technologies.

Consider the simulacra we discussed earlier; they are but one type of biocomputer, optimized for fidelity to the human tissues they’re meant to mimic. However, the AI agent might discover other forms of biocomputers that more efficiently and cheaply emit the information needed to accomplish its aims, but which look nothing like existing forms of biocomputer—perhaps they are chimeric, or blend electronics (e.g., embedded sensors) with living tissue. Furthermore, the agent might develop new types of scopes and scalpels for monitoring and perturbing these new biocomputers. The agent could even discover hacks, like overclocking biocomputers (by literally increasing the temperature) to accelerate its experimental loop.

At a certain point, the agent’s physical tools, not just its internal in silico tools, might be completely inscrutable to us.

Speculating on Future Trajectories of Biomedical Progress: Or, How to Make PASTA

As data flows continue to accelerate, the human distributed scientific hive mind will hit its cognitive limits, unable to make sense of the fire-hose of biological data. Humans will begin ceding control to AI agents, which continue improving slowly but steadily. In due time, AI agents will come to dominate all of biomedical research—first therapeutics development, and eventually basic science. This will happen much quicker than most expect.

The AI’s principal advantage is informational: it can spend increasingly cheap in silico compute to more efficiently use limited biocompute. And in the limit, as it ingests more data about biomedical dynamics, the AI will offload all biocomputation to in silico computation, thereby alleviating the ultimate scientific and technological bottleneck: time.

However, AI agents are not yet autonomously solving long-horizon biomedical tasks or bootstrapping their own tools, let alone building a near-perfect biological simulator. We are still bottlenecked by biocompute, so there’s reason to be pessimistic about biomedical progress in the short-term. But, conditional on advances in biocompute and continued exponential progress in scopes and scalpels, the medium-term biomedical outlook is more promising. In the long-run, all bets are off.

The most merciful thing in the world, I think, is the inability of the human mind to correlate all its contents. We live on a placid island of ignorance in the midst of black seas of infinity, and it was not meant that we should voyage far. The sciences, each straining in its own direction, have hitherto harmed us little; but some day the piecing together of dissociated knowledge will open up such terrifying vistas of reality, and of our frightful position therein, that we shall either go mad from the revelation or flee from the light into the peace and safety of a new dark age. — H.P. Lovecraft, The Call of Cthulhu

In 1953, Richard Feynman ended a conference talk in Kyoto with an apology. The famous physicist seemed embarrassed—he’d left part of his problem unsolved. Feynman had been presenting on helium’s superfluid phase, but phase transitions were new territory for him. After the talk however, the famously terse Nobel laureate Lars Onsager shared with the audience that…

Mr. Feynman is new in our field, and there is evidently something he doesn’t know about it, and we ought to educate him. So I think we should tell him that the exact behavior of the thermodynamic functions near a transition is not yet adequately understood for any real transition in any substance whatever. Therefore, the fact that he cannot do it for He II is no reflection at all on the value of his contribution to understand the rest of the phenomena.

This was the state of affairs in the fifties. We’d figured out quantum mechanics and relativity, yet something so mundane as boiling water1 was stumping the great physicists of the time. That is, so it was before Ken Wilson ushered in his revolution.

Wilson’s ideas from the sixties and seventies—effective field theory and the renormalization group—rank among physics’ deepest discoveries. They started as tools for understanding phase transitions but ended up revolutionizing how we think about the laws of nature. Few concepts from my physics career have shaped my thinking more broadly than these. Yet I’d wager most non-scientists have never heard of them. This makes sense given their dull names, but don’t let that fool you. These ideas deserve to sit alongside quantum mechanics and relativity. I’ll argue they’re just as weird too. In fact, using these ideas together with concepts from computational complexity, I’ll make the case that the universe plausibly contains phenomena worthy of the name spirits.2

I. Renormalization and hints of incompleteness

Physics is supposed to be the fundamental science. The standard story goes like this: given the laws of physics plus infinite computing power, we can predict any phenomenon in the universe. Sure, we don’t understand dark matter or black hole interiors yet, but that’s just because we haven’t found the right laws. Once we do, it’s game over. Everything reduces to simple laws whose consequences can be deduced with infinite compute. This picture might be completely wrong, however.

First, Gödel’s incompleteness theorems say that this does not logically follow. There could be consequences of the laws of physics that are unprovable through any computation whatsoever. Erik Hoel had a wonderful post last year about the possible ramifications of Gödel incompleteness for science and consciousness that I highly recommend.

So non-trivial implications of Gödel incompleteness for physics is very much a live option. As Erik points out, research has uncovered a toy model quantum system that suffers incompleteness, making it impossible to determine the so-called spectral gap: the energy difference between the two states of lowest energy. That said, this model is quite different from our best theories of fundamental physics and relies heavily on the toy system being infinite in a very particular way. And infinity is truly at the heart of Gödel incompleteness. So it’s unknown how much of a restriction Gödel places on our physics.

However, here I want to point out other severe dangers for physics that do not rely on Gödel at all.

To understand these dangers, we first need to see how modern physicists think about the laws of nature. The key concept here is scale. When you hear “scale”, you probably think of size. But a soccer ball and tennis ball aren’t really on different scales—their sizes are too similar. What matters are dramatic differences of size.

So what counts as being of different scales? The key concept is resolution. To understand any phenomenon, it is of utmost importance that you set your resolution correctly. Change the scale, and you’d better adjust your resolution or you’re dead. Too coarse, and the information you need disappears. Too fine, and your phenomenon drowns in irrelevant details.

You might think irrelevant details are better than missing information. But for practical purposes, they’re equally useless. Your computational limits ensure the information you need is completely inaccessible—it might as well not exist. It’s like using the standard model of particle physics to describe a cannonball’s trajectory. Sure, maybe with a Dyson sphere powering your supercomputer you could do it. But if you value your sun, it’s a stupid idea.

But wait—how sure are we that even a Dyson sphere would work? I’ve been saying too fine resolution is bad “only for practical purposes”, and this caveat is essential if you believe everything reduces to physics. But I’m much less confident in “only for practical purposes” than I used to be. In fact, I’ll soon argue the problem isn’t just practical.

But I’m getting ahead of myself. What we’ve implied above is that different scales require different instruments—of measurement, of language, and of mathematics. The theory best suited for prediction throws out the maximal amount of irrelevant information while keeping as much information pertinent to your phenomenon of interest as possible. A very delicate balance.

We call theories designed for the right scale effective field theories (EFTs). These have been wildly successful in modern physics. I could list their achievements, but in reality, almost all successful physical theories are now understood to be EFTs. Even our supposed fundamental theories like general relativity or the standard model of particle physics.

So how do you build an EFT? You write down the most general theory consistent with a few basic principles: what you want to predict, the symmetries of your problem, plus whatever axioms you won’t compromise on—usually quantum mechanics and relativity, and at the very least locality. This leaves a handful of free parameters you can’t predict from first principles. You just have to measure them.

A mind-boggling fact, however, is that such theories exist for any physical phenomenon at all. How on earth can we write down predictive theories that depend on just a handful of parameters? This is basically asking: how can we predict a cannonball’s trajectory knowing only a tiny amount of information? The cannonball consists of an enormous number of atomic subcomponents, each one an independent knob that can be dialed to adjust the state of the cannonball. Why don’t these many unknown pieces of information, far exceeding what’s captured by the ball’s mass, add up to a giant blob of unknowability?

Somehow, all that fine-grained information gets compressed into just a few numbers that characterize your system’s behavior. For the cannonball, just the mass. How does this compression happen? Why aren’t there 34 different constants we need to measure to predict the cannonball’s trajectory? Don’t think this is obvious—it’s not. We’re just used to it.

Yet, no one has managed to write an “EFT of human history”. But humans are basically giant blobs of erratic atoms interacting with each other. Why does physicists get nice simple theories while historians gets chaos?

To understand when EFTs work, we need the second star of the show: the renormalization group. This is the technique that shows how tiny-scale physics gets compressed into just a few numbers that affect larger scales. The key ingredients enabling the miracle are locality (no action at a distance) plus relativity (no preferred reference frames).

Fortunately, we can understand the main insights through a simple little example that lets us skip the mathematical heavy lifting that shows generality of the mechanism—you should consult this absolute beauty of a paper by Joe Polchinski for that.

Here’s the example. Consider a frosted glass panel:

Shine a beam of light on it at an angle. Some light transmits through, some reflects off. We’ll focus on the reflected light.

Say you’re hoping for a theory of the reflection angle without worrying about all the microscopic details inside the glass. If the glass were perfectly smooth, this would be trivial—the incoming angle equals the outgoing angle:

Makes sense:

With low-energy photons, the simple formula still works approximately, even if the glass isn’t perfectly smooth. Here’s why. A basic physics fact is that photon energy is inversely proportional to wavelength:

So low-energy photons have long wavelengths and “wiggle” over longer distances.

The picture tries to show why this matters. A long-wavelength beam can’t “see” tiny surface imperfections—they’re too small compared to the wavelength. They average out. But a high-energy green laser has tiny wavelengths, so it suddenly becomes sensitive to, even dominated by, little bumps and scratches.

The imperfections have their own characteristic typical size, call it L_pane. Once you account for this, your scattering formula becomes:

The g_i’s are different numbers that characterize the pane’s imperfections. The series is infinite—higher and higher powers of energy appear, with an infinite number of constants. These g_i’s are completely non-universal. They depend on arbitrary microscopic bumps and scratches at the exact spot where your beam reflects. Move your laser to a slightly different location, and you’ll have to remeasure g1, g2, g3... all the way up to g42 and beyond.

An absolute nightmare.

Thankfully, the effects of this disaster is usually not felt. If the photon energy is low enough, then:

Let’s say this product equals 0.01.3 Then:

You can ignore all the g_i’s and barely make any error. This assumes the g_i’s aren’t gigantic, but wonderfully enough, Polchinski’s absolutely hottie of a paper show they’re typically order unity. So microscopic details affect the reflection angle, but negligibly at low energies. Phew.

This isn’t just a quirk of glass panels. The renormalization group shows that relativity plus locality almost always guarantees this same pattern:

By “simple” I mean something that depends on just a few constants, independent of whatever mess happens at short distances.

Now, studying low-energy physics is basically the same as studying large-scale physics. So the renormalization group tells us why we can ignore tons of microscopic information—like atomic jiggling inside a cannonball. But two conditions must hold: (i) you’re not trying to predict high-energy phenomena, and (ii) you’re willing to accept limited precision.

Push too far into the decimal places, and all kinds of microscopic junk starts influencing your answer. But those first few digits? Locality and relativity are the bodyguards keeping the microscopic madness at bay.

Now, if you want to go to high energies, you are in trouble. Consider making your photon purple, so it has a lot of energy, say E L_pane = 4. Then we get

Ouch. The sum blows up to infinity, and your theory is obvious nonsense in this regime. Now you must cook up a new theory integrating the detailed dynamics of pane itself. You have no choice but to model a bunch of stuff that you previously had the luxury of ignoring. This sounds like work, but’s such is life.

So you do the hard work of modeling new details to understand light-pane interactions. You discover new patterns, leading to the quantum theory of atoms. This eliminates your old problematic series expansion—wonderful! Your new theory might predict some complicated formula for the outgoing angle, but it’s no longer a series in E×L_pane, so no more nonsensical infinities.

But a new infinity lurks around the corner.

A new, smaller scale

has just appeared. Your improved theory gets corrected by a different series expansion, caused by this new characteristic length. This scale represents the new boundary below which you’ve ignored some physics. Sure, it’s smaller than before, so you can push to higher energies. But go even higher, and you need to work even harder. If you do that, you discover quantum electrodynamics—the proper relativistic theory of electrons and photons.

But QED also breaks down at high energies. It has its own breakdown scale. So you assemble an international team, build a particle accelerator, and discover the standard model of particle physics. You might think that it stops there, but no, it doesn’t. We know very well that our understanding of gravity breaks down when energies gets close to the so-called Planck energy.

Every theory we’ve found breaks down at high enough energy. This is well understood and doesn’t necessarily spell doom for physics. But there’s another frontier where theories break down—one that’s less appreciated and that signals real trouble.

II. The complexity frontier

Take a box with a few hundred atoms at low temperature, perfectly isolated from the environment. Low temperature means we’re in the regime of understood physics. Imagine we can measure every atom’s position and velocity precisely.4 We leave the box alone for a long time, then ask: where are all the atoms now?

With unlimited computing power, our best theories should handle this easily. But they can’t. There’s a finite time horizon beyond which no physical theory in our possession can predict the atoms’ locations—even at low energies, and even with a galaxy of Dyson spheres powering our GPUs.

To see why, let’s set up a cartoon picture that captures the essential effect.5

Focus on one atom. Our little guy travels in a straight line until he bumps into another atom. His direction and velocity change, then he travels in a new straight line until the next bump. This repeats endlessly. Seems fine—we can compute the angle and velocity changes, so we should be able to track all collisions. Our Dyson galaxy of computers can handle it.

But every collision introduces a tiny error. Being maximally conservative, pretending no new physics occurs before the Planck scale, our calculation drifts from reality by roughly:

where ℓ_Planck is the Planck length—the scale where quantum gravity matters. This error lives extremely far out in the decimal places. The Planck length is absurdly tiny.

But errors accumulate. As time passes, your estimates of atomic positions and velocities drift from whatever nature is actually doing. We have no theory to account for these errors. Eventually, you lose track of when and where collisions happen. Your predictions become worthless.

The answer to “where is this atom later?” becomes completely unanswerable, even with unlimited computing power.

Your sin was asking a question that requires too many queries to our theory of physics before producing the final answer. This mechanism is completely general—all our theories break down at high complexity.

There’s an obvious attempted fix though: first go ahead solve quantum gravity with those gigantic computers, thus eliminating the errors in your theoretical predictions. But this assumes something that could be totally wrong. Namely, it assumes that our search for an ultimate theory eventually terminates. This belief is despite most empirical evidence. Every time we’ve zoomed in on nature, we’ve discovered new stuff. How do we know there will ever stop being new details as we keep zooming? How do we know the laws of physics aren’t fractal-like?

I’ll argue we’re in trouble even if they’re not fractal-like all the way down. But let’s assume they are for now.

If new physics keeps appearing as we zoom in, we’ll never have an ultimate theory. Instead, every theory will have finite precision and a finite predictive horizon, even in principle. You might hope that as theories improve, the horizon of the explainable expands in all directions. But even that’s questionable.

Assume I have a fixed compute budget. Finding a “better” theory—one valid at shorter distances—might not help much at all. The problem is that “more fundamental” theories are always more computationally expensive. Just ask my friends trying to simulate a handful of quarks on supercomputers—a physical system with total mass roughly 10^{-30} times your bodyweight.6

So as you improve your theory, there are directions in phenomenon-space where you don’t push your horizon forward with a fixed compute budget. The only way to hope to explain all phenomena is if your compute budget grows alongside your discovery of new theories. But this hope is dangerous.

Two limits together put an upper bound on the compute you can ever use. First, the universe expands at an accelerated pace, creating a so-called cosmological horizon. The implication of this is that only a finite volume of the universe is accessible to us. Second, our best hint about quantum gravity—the Bekenstein-Hawking entropy formula—suggests that finite volumes of space contain finite amounts of information.7 So evidence supports there being a finite amount of compute available to us, bounding the horizon of explainable phenomena.

This finite compute creates trouble even if there is a final, error-free theory of physics. Even with a perfect theory, we can’t work out arbitrary consequences. While each individual “query” to the theory is error-free, we only have a finite number of queries at our disposal. Only so many atomic collisions we can predict.

Now, hearing “finite information,” you might get excited. How can that not be knowable in principle? In a trivial way it is—our patch of the universe describes itself, I guess. But acknowledging this doesn’t amount to having a theory of it. It’s like saying a map of the world at 1:1 scale is technically complete. Sure, but good luck folding it.

To possess a theory is to possess a compressed representation of information that grants predictive power. The finite information content suggested by the Bekenstein-Hawking entropy quantifies how many measurements you’d need to completely determine the initial conditions in our patch of the universe. This also represents the “storage capacity” of our part of the universe.

Now comes a remarkable calculation we can make. I don’t assume you’re a physicist, so I’ll just describe the result—see footnote for the full computation.8 If our patch of the universe is (1) described by finite (but possibly gigantic) information, and (2) we believe quantum mechanics, then we can show this:

The derivation uses a few technical terms, but the idea is remarkably simple. In quantum mechanics, the laws of physics are completely captured by a single matrix—a rectangular array of numbers known as the Hamiltonian.9 This array uniquely defines an action on matter in the universe, corresponding to evolving time forward. Under the assumptions of quantum mechanics and finite information, the matter in the universe is equivalent to a collection of so-called quantum bits—qubits. But the information required to uniquely specify this matrix is much larger than the information storable in the qubits:

So the laws of physics might be described by more numbers than you can store in the visible universe. Mind-blowing, if you ask me. And you can’t necessarily solve this by positing a larger “hard-drive universe,” because the same complaint applies to that universe too.10

I can already hear former colleagues’ complaints about suggesting that this possibility might be realized. The known laws are simple! The observed laws have so many symmetries!

All well and true. It’s certainly possible that the exact, error-free description of the laws is short enough to be stored within our patch of the universe. This would work for a highly symmetric Hamiltonian matrix where most elements are zero, perhaps something like this:

Most elements being zero is a consequence of locality, it turns out, while the equality of various elements represents the independence of physics of location. But I’d like to point out that we’ve actually studied only an absolutely tiny fraction of the total information in the universe. In matrix language, we’ve really investigated only a tiny fraction of the numbers in this matrix. Like, I mean, really really tiny.

According to this paper, the information stored in matter we understand reasonably well is roughly 10^89 bits, while the information in our patch of the universe is 10^122 bits. That is, the matter we think we understand decently represents a fraction:

10^{-33} = 0.000000000000000000000000000000001%

of the information content. Who knows, maybe this absolutely tiny subset of matter is described by particularly nice and tidy laws? There are good reasons to think so.

A much larger fraction of the observable universe’s entropy is carried by supermassive black holes. This is about a quadrillion times larger than the amount of standard matter (but it’s still an absolutely tiny fraction—most resides in the unknown). Modern research on the holographic nature of quantum gravity suggests that black holes are intimately tied to chaos and the butterfly effect (classic paper). And there are very good reasons to believe that spacetime and thus locality becomes completely corrupted inside a black hole—good old Einstein’s general relativity predicts this. This hints that you’re reaching whole new levels of complexity, associated with the breakdown of locality. And locality is what made your laws so nice and tidy in the first place. But let’s say that physicists ramped up their experiments to measure about a (quadrillion)² times more stuff than what we have seen in the universe so far. Even then, if we kept finding locality, the best we’d get is an array of numbers encoding the laws of physics that still looks like this:

where the unknown might be highly irregular and dense. But at present, that green box represents a fraction of 10^{-33} of the rows and columns, and only there do we know that the matrix is sparse, i.e. locality holds.11

In fact, albeit working with a toy model of quantum gravity, my collaborators and I published results last year hinting that black holes might be a mechanism that “hides” exponentially complex aspects of the laws of physics. It must be said that this is speculative work, but it fits with the picture that is emerging in modern theoretical work on black holes.

So, we often assume all of physics is local because the tiny sliver of the universe we engage with appears to be local. But we can make a reasonable anthropic argument for why, even if the universe fundamentally lacks locality, humans would arise as dynamical phenomena instantiated on degrees of freedom whose dynamics look local. Humans are patterns that somehow persist and replicate with high fidelity, and non-local chaotic dynamics might be too unstable to allow anything like this. In fact, spacetime itself may be viewed as the optimal “compact description” of this sector of relatively simple physics. So if we were to exist, we almost certainly would exist in a nice tidy corner of the matrix.

III. Spirits

Now we are in a weird situation. The laws of physics that we are able to write down in our universe, even in principle, might be incomplete, and so they are fundamentally limited in their predictive power. What does that mean, pragmatically speaking? It means that as we go out in nature and observe patterns of sufficient complexity, there might be no way for us to (1) explain how they got there and (2) explain how they will evolve.

There are strange consequences implied by this picture. We can now imagine that there are two distinct theories out there that describe different phenomena seen in nature, one of them being what we traditionally call physics. Yet, the two theories might be computationally inaccessible from each other. Physics cannot derive the other theory even given access to all compute in the universe, and vice versa. And so, what does it mean to say that physics is more fundamental? Not much, in fact. You can say that physics is a mind-blowingly good theory—better than any other humanly known scientific theory at present. But not fundamental, if this is the state of affairs.

You might say that in some hypothetical Platonic realm with infinite computing power, you can derive the other theory from physics. But I’m skeptical of such sleights of hand. An deep lesson from physics is that information must be instantiated on physical resources.12 In fact, heat is nothing but information you lost track of. Seriously, I’m not being facetious here. That is literally what heat is. There’s even a phenomenon called reversible computing where, if you never discard any information, you produce no heat. Zero. Nada. Insert Carnot’s shocked face here.

So, saying there is a Platonic realm where we can connect all other theories to physics, implying that such a thing “exists”, is possibly as meaningful as saying “blarb is glarb”. Both statements have equal implications, in pragmatic terms, for our possible experiences of the universe.

Of course, these other theories, if they exist, are probably much more complex than physics. There’s a reason we (probably) haven’t found them yet. Either way, we might have this situation:

But what about the no-man’s land in between theories? What on earth is that?

Let’s be sure we read the picture correctly. What I’m implying is that there’s no theory for the natural phenomena living in this region. But what does that mean? What is a theory, even?

A theory is nothing more than a compression of observed patterns. Physics is great because it’s so damn good at compression. We observe all these seemingly disparate, highly complex phenomena that all seem unique. But physics realizes they’re actually characterized by a surprisingly small amount of irreducible information that can be decompressed using the laws of physics.

The statement that physics is fundamental becomes the claim that physics can compress any observed phenomenon that any other theory can compress—although physics might be much less efficient at decompression/compression than a more specialized theory.

You might be perplexed here, since we usually say that what makes a good theory is that it can predict or explain. But prediction is really just a roundabout way of checking if you have a correct compression scheme for the class of phenomena you’re observing. When you can compress, you have isolated what information is irreducible and what is redundant, and you can reproduce the redundant information from the irreducible information.

But you can easily fool yourself by building a seeming compression scheme that works on a few instances of your phenomenon, but the danger is that you’ve overfitted to these and basically just hardcoded their patterns. The way to check that you haven’t is to predict.13

So to say that a phenomenon cannot be described by a theory is to say that there is no significant compression of it. These are dynamical phenomena taking place in the universe whose only accessible description is just themselves. There is no shorter description of the phenomenon than just recording it.

But there is nothing that rules out such phenomena actually having causal contact with our lives. Phenomena that influence our life trajectory, but which cannot be theorized over. Forces in the universe that can only be observed in their singular appearances, never captured nor even capturable by theory. I cannot think of a phenomenon more worthy of the name Spirit.

Despite excelling in his single year of schooling, George Green wasn’t allowed to continue. His father instead wanted him in the bakery and—soon after—the mill. So nine-year old George didn’t have a choice. Kneading breads and grinding grain it was, for the next 27 years.

Yet, the mill-bound Nottingham boy grew up to write a masterpiece—a work that earned him a Westminster Abbey memorial and led Einstein to remark that Green had been 20 years ahead of his contemporaries.

And ahead he was. Green’s seclusion from the English men of learning gave him a strange advantage. Study of continental mathematics was actively discouraged in England at this time thanks to the rivalry between Newton and Leibniz over the discovery of calculus. The English naturally took Newton’s side and, to their detriment, cut themselves off from a Cambrian explosion of mathematics on the continent. But not Green. He exchanged mill profits for articles by Poisson, Laplace, and their like through the Nottingham Subscription Library. So, candle in hand and French on tongue, Green worked on mathematics after grinding.

Finally in 1828, at the age of 34, the miller published his paper:

Green discovered techniques to solve a type of equation that humanity was only then starting to encounter, but which we seem to find in increasingly exotic form each time we improve our theory of nature. Equations telling us not how physical systems evolve with time, but rather what systems are allowed to exist in the first place. A class that becomes more load-bearing the higher you climb, until you reach the highest known peak, quantum gravity, where this type of equation is the only one left—time itself seemingly gone. Equations that lead you to the question…

…Is it impossible for the universe not to be filled with beauty?

I.

In your introductory physics class, Newtonian particle mechanics is first on the agenda, and the quintessential question is this:

If a particle starts in this position, with that velocity, and is subject to those forces, what will its future trajectory be?

While studying this question, you get the feeling that physics is free and flexible. You get to invent all kinds of little universes. You can place particles here, you can place particles there; they can move this way, they can move that way—no one can stop you. What is not free, however, is how these things evolve as you start the clock. Only then does the problem really begin. The bishops and rooks of physics have predefined moves, but in this game you can place the pieces wherever you want before you start.

Then you get to electromagnetism, and your professor asks: what is the electric field surrounding a particle of charge Q?

Here time is nowhere to be found, beyond the fact that the question implicitly refers to a fixed moment in time. But since your professor bothers to ask, you probably cannot cook up whatever state of electric fields you want—even at a fixed instant of time. Some photographs simply cannot exist.

So which electric fields are legal? This question is answered by Gauss’ law, which is the equation Green showed how to solve.1 It is a type of equation known as a constraint equation, and this particular one restricts what electric fields are allowed to exist. It says that at any fixed moment of time, the following must be true at every point in space:

Electric charge density = divergence of electric field.

This is equivalent to saying that electric charges, like electrons and protons, act as sources and sinks for electric fields; the field is like a river, and positive charges act like springs of groundwater, while negative charges are vortices draining water down underground. And I mean this precisely. As long as we compare the electric field and the river at a frozen moment of time, the two are governed by precisely the same equation. The spatial patterns that rivers and electric fields are allowed to inhabit come from the same restricted library of possible patterns.2 Here is an electroriver I’ve simulated for the occasion:

There are no electric charges in the above simulated image, implying that the field lines always travel parallel to each other as they get close—no divergence or convergence. With electric charges present, on the other hand:

So, Gauss’ law says that electric fields are constrained. You are not at liberty to groom your corner of the universe such that the E-field swirls at random. Once you’ve spent your freedom in creating a neat decoration of electrons and protons, the electric field must respect it.

That said, the E-field is not uniquely determined either. Even after placing charges, our creator has granted us infinite freedom in how to decorate them; it just happens to be a smaller infinity than we otherwise could’ve had. This smaller infinity is bound up with the question of beauty. However, we are not prepared to understand that quite yet.

Let’s temporarily trade aesthetic satisfaction for clarity by viewing the electric field as arrows. This will also let us display the strength of the field in addition to its direction, which will pay off. So, let’s knead for a bit:

What you just saw wasn’t a time-evolution you could observe in the real world. Instead, we cycled through a family of legal snapshots of the universe at an arbitrarily chosen rate. Now, if you look at the video carefully, you see that we can’t beat down the field around the charged particle. We can reduce it on one side, but we always pay with an increase on another side. The same cannot be said further away from the charge.

Let’s zoom in on a region far away from the charge. Next, we draw a closed surface chosen to be an ellipse for no particular reason, and focus on the field precisely on the surface itself (if you want you can pretend it’s a projection of a 3d ellipsoid to a plane):

The dark blue arrows correspond to the E-field. The coral and teal arrows represent the field projected perpendicular to the surface. If the E-field instead was the velocity field of water and you wanted to find the rate of change of water volume inside the ellipse, you’d care about the coral and teal arrows. Water grazing the surface tangentially does not contribute to inflow, no matter how fast it flows.

Now, remember that Gauss’ law is the statement that only charges act as sinks and sources. But there are no charges inside the ellipse—no “water” is being supplied or drained off. So the inflow better equal the outflow. That is, if you stack all the coral arrows on top of each other, they ought to have the exact same length as the teal ones. Let’s see it.

Let’s see what happens when we draw the ellipse around the charged particle instead:

If you look carefully, you see that the height difference between the coral and the teal stacks stays fixed in time. In fact, their difference is equal to the net charge Q of the particle inside the surface (provided you multiply each arrow by the area of the portion of the surface that it stewards—the part of the surface closer to that arrow than any other arrow).

This turns out to be a universal relationship, known as the integral form of Gauss’ law. It says that for any surface without holes, no matter its shape, if you sum up the perpendicular E-field component all over the surface (scaled by the surface area chunks), then what you get is the net charge inside it. In symbolic form, a physicist would write the following3

You now know exactly what this equation says, whether or not you have studied surface integrals (the continuous version of the discrete arrow-stacking procedure we did above). Again we can map this onto the water analogy: if water is being added or drained inside the surface, this amount must be exactly compensated by inflow or outflow.

II.

Your enemy has hoarded electrons. And for reasons I’ll never know, you’re set on tallying their total charge. Unfortunately for you, they’re in a vault.

With Gauss’ law in your pocket, however, you bypass the whole break-in ordeal. You instead scrupulously measure the electric field everywhere across the outer walls and roof of the vault (also, unfortunately, you must dig tunnels to measure underneath as well). Once done, you can sum up the field on your favorite surface enclosing the vault, and out comes the charge.4

We see here that information about charge resists localization. It always emanates into space, with the practical consequence that tallying up the charge in a region never requires you to enter it—an occasion where no cost is paid for shallowness.

Our superficiality can be used for even mightier deeds, like detecting a wormhole without the inconvenience of dying. To see this, let’s pretend that the universe has two spatial dimensions and no charged particles. Now let’s draw a 2d wormhole with electric field flowing through it:

If you draw a surface (dashed circle) around the wormhole on one side and tally up the perpendicular arrows, you clearly get a non-zero number Q. But there are no charges present. And yet, Gauss’ law says that the sum of the (arrow-scaled) length of these arrows equals the charge.

You have just discovered that wormholes mimic charged particles. It’s charge without charge. So, if God promised that the universe has no charged particles but you nevertheless measure a non-zero Gauss’ law surface sum, well, then you’ve found a wormhole. Nonzero topology of space acts just like charge.

So as you look at the wormhole from afar (bottom half in the drawing), it looks precisely like a blob of electrons. But as you travel through it (before spaghettification kills both you and time itself), you never encounter any charged particles. And the moment you turn around, it looks like a bunch of positrons are behind you! This is because the electric field arrows point either into or out of the wormhole, depending on which side you stand on, so you either tally up a charge of +Q or -Q.

Hold on, hold on. Are you sure electric charges aren’t just micro-wormholes? Is the creation of an electron-positron pair just the nucleation of a wormhole?

Charles Misner and John Wheeler also had this intuition in 1957, when they published the paper Classical physics as geometry. You now understand the core idea of that paper: matter particles might be micro-wormholes (and it suddenly feels obvious why particle and antiparticle are created in pairs). While the idea doesn’t quite work in its most naive form, it is still breathing in various exotic theories of physics, such as string theory, where charge is an aspect of the geometry of bonus dimensions that are too small to currently detect.

Whether or not you buy that, notice what just happened. By simply observing that the universe is constrained in a seemingly modest way (electric field lines don’t recede or approach each other when charge is absent), unexpected facts follow: charged particles and wormholes look the same from afar, and information about charge refuses localization in space. You might have thought that imposing constraints makes the universe impoverished, but we saw the opposite. Imposing constraints seems to fill the universe with structure. In fact, your most holy piece of music can be sculpted out of white noise.

When I’ve talked about constraints so far, I’ve implicitly talked about physical laws that belong to a class of beasts known as elliptic partial differential equations. Translation: equations that constrain how some quantity in space is allowed to behave purely based on stuff right next to it and, crucially, not in terms of anything to its past. You don’t need to know what configuration came before in time or what comes after—some rules cannot be violated. Constraint equations delineate edges in the space of possible worlds.5

Now, given the payoff we got from studying a single constraint, we ought to immediately inquire: what others exist? What do they imply?

Each time we’ve unearthed a more fundamental theory of nature, we’ve found that our constraints are just a corner of an even richer mosaic. We’ve found exotic non-linear extensions of Gauss’ law. For example, electromagnetism gets eaten by a bigger theory, where the electric field is just one face of a multifaceted dynamic object—the electroweak field—that has even richer patterns imposed on it.

If we climb all the way to quantum gravity by the most parsimonious route, which means upgrading Einstein’s gravitational equations to their quantum versions in the most vanilla way, things get weird.6 The web of constraint subsumes everything—only constraints remain! To physicists’ ongoing puzzlement, there are no time-evolution equations left.

This leads to some jazzy outcomes. We saw that Gauss’ law causes delocalization of information about net charge. In gravity, the expansiveness of the constraints seems to lead to the delocalization of all of physics! Information about everything leaks everywhere. For example, we now have the following formula for the total energy inside the surface (written in slightly impressionistic form):

With good measurements of the gravitational field/curvature of spacetime, we never have to enter a volume to know how much energy is inside it. In fact, according to our most conservative attempts at quantizing Einstein’s theory of gravity, everything that goes on inside this volume is knowable from information encoded entirely on the surface.

This means that all secrets your enemy hides in their vault can be decoded from quantum measurements (extremely complex ones) localized to a surface enclosing the vault. One dimension of space seems to be redundant. This is the idea of the holographic principle, which I worked on in my PhD, and which I’d stake real money on being true. It is almost surely the outcome of the tight constraints that gravity imposes through its quantum constraint equations.7

III.

Is the universe full of structure? I certainly hope so. I would fall into enduring despair upon learning that it was mostly featureless, or at the other extreme, full of white noise—if the universe were to have no surprise or nothing-but surprise.

The anguish I feel when contemplating a structureless universe, whether due to perfect order or maximal disorder, is a distinctly aesthetic sort of dread. This is not incidental. It takes us right to the core of what aesthetic satisfaction really is.

Namely, when you feel awe or beauty, I believe your subconscious has successfully come up with a valid abstraction. Your brain has, relying on its impressive array of experience and sensory processing capabilities, discovered symmetry and redundancy in the incoming sensory stream. It has found an underlying generating mechanism that enables partial in-advance prediction of incoming sensory data, letting it validate the correctness of its abstraction without any room for cheating. It has found verifiable structure.

Structure resides in the borderlands between order and disorder. It’s present when something is neither as simple nor as complex as it otherwise could’ve been. It arises in that zone where, if you look hard enough—but only then—you will be rewarded. At absolute order, there is nothing left to predict, while at total disorder, prediction is doomed to fail.

In possession of a new abstraction, your brain is now more likely to predict future sensory signals. This is useful for survival, and so evolution rewarding us upon discovery of abstraction feels almost obviously true, once you think of it. The savannah man enjoying the discovery of subtle distinctions in color avoids poisonous berries, while his fascination with intricate cloud motion lets him bring the family to the caves before the storm.

So, must the universe be filled with beauty?

Could it be that the universe is low on structure and new things for us to abstract? You might say no—have we not seen that it is rich with it, given its crystal lattices and human civilizations? These certainly give lower bounds on the amount of awe to be had, but oh, how little we have seen across space, time, and even further beyond—where those concepts stop making sense.

An artist conceiving a masterpiece has three routes to structure: additive, subtractive, and transformative. The additive route is taken by the painter filling a blank canvas, while the subtractive path is taken by the electronic musician carving timbres out of noise samples. The collage artist takes the transformative path, shuffling complexity that is already there.

What about the universe itself? We could try to carefully study the laws that govern its unfolding with time—the additive or transformative route. At an empirical level, as far as our eyes and telescopes see, the universe is in the process of birthing incredible structure. It moves in ways that let electrons, protons, and photons grow into The Well-Tempered Clavier.

Yet, at the theoretical level, I don’t know of any assurances that the universe will keep adding structure, and a fat entropic cloud hangs above: the second law of thermodynamics—the statement that the universe will forever get more surprising.

The growth of surprise is good if we start with something simple and predictable. Remember, structure lies between total simplicity and disorder. So the second law lets featureless worlds climb towards structure, for a while. At some point, however, the curve starts going down—more surprise means the fragmentation of structure, rather than growth. There are increasingly many things to predict, but the probability of success in doing so goes down even faster, and what’s left to predict would not be recognizable as worth the effort: functionally infinite mixes of micro-patterns with no unifying principle.

And so, the heat-dying universe stares at us.

I’m not confident that heat death will come to pass, however. True, it seems likely if you accept the current working model of the universe as the end of the story and also forget about quantum effects, holography, and and other looming uncertainties. In extrapolating to extreme scenarios such as the heat death of the universe, I’d not bet much that we can ignore those. As just one example, strange and hard-to-understand things happen when you try to apply holography to cosmological scales, and it is clear that the top experts are confused, although the research is quite active at the moment (see for example this Quanta Magazine article with interviews with several former colleagues).

And so, leaving behind a thicket of question marks that I’ve only vaguely alluded to, and which are not easy for us to untangle here, what other principles do we have to help us?

Constraint equations, of course. While humanity hasn’t found even a single law we’re confident is universally valid, the constraint equations that come with electromagnetism, gravity, and the nuclear forces are as universal as they seem to get. They appear to constrain every possible state in the regime of physics we remotely understand. All blocks of possibility are chiseled down, forbidding the anarchy of white noise, birthing patterns for us to stumble upon.

Beyond our horizons of knowledge, how many more of them exist? Do sublime oceans of structure lie outside the utterly negligible fraction of the matter in the universe we understand well? A fraction that appears to be at the very most

0.000000000000000000000000000000001%.

The second law of thermodynamics pushes the universe towards disorder, but if constraint equations run as deep as they seem, they might still make it impossible for the universe not to be filled with beauty.

Further Reading

• “A Natural History of Beauty” by Kevin Simler

• “Great scientists follow intuition and beauty, not rationality” by Erik Hoel

Isaac Newton’s lifelong quest to transmute base metals into gold is normally forgiven as a symptom of the pre-scientific nature of his age. But “great minds holding eccentric, even kooky, beliefs” is a pattern that crops up throughout history. Even after there became strong social reasons for scientists to disguise even the faintest whiff of the irrational. William James, the godfather of psychology, believed in ghosts. Fred Hoyle, who came up with the idea that stars created chemical elements via nuclear fusion, thought that influenza came from space. Nikola Tesla was obsessed with the number three. Nobel Prize winner Wolfgang Pauli believed that his mere presence could drive laboratory equipment to malfunction. Kurt Gödel starved himself to death out of fear of being poisoned. Brian Josephson, a still-living Nobel laureate, thinks that water has memories.

Why?

In the “TV model” of science, scientists are pinnacles of rationality—socially inept, boringly nerdy, emotionless, and incapable of strong pre-evidence beliefs. You can also find this view expressed in places like Steven Pinker’s latest book, Rationality, in which he argues:

It has become commonplace to conclude that humans are simply irrational—more Homer Simpson than Mr. Spock. More Alfred E. Newman than John von Neumann… To understand what rationality is, why it seems scarce, and why it matters, we must begin with the ground truths of rationality itself, the ways an intelligent agent ought to reason, given its goals and the world in which it lives.

Pinker’s reference of John von Neumann makes sense, because von Neumann has lately become the poster child of genius, which in our metric-obsessed age means an IQ stratospherically high coupled with “rational” thought practices.

Yet it is actually fitting that for the public the most famous scientist remains Einstein, not someone like John von Neumann. Oh, there are arguments to be made that John von Neumann was more intelligent than even Einstein. Tutored far more intensely from a young age, homeschooled until 10, given governesses and access to advanced private math mentors, he had a wider knowledge base than Einstein, and was certainly much more rational (kind of obvious from being a pioneer in game theory). Leading people to spark viral debates online with things like this:

But von Neumann himself? He felt he lacked a certain quality as a scientist. Here’s from a profile in The New Republic:

In his work… von Neumann never had strokes of irrational intuition, the kind that result in startling new ideas; his friend Einstein had many, and von Neumann envied them.

He understood that there was not just one thing, rationality, that describes good thought. If that were true, scientists should be immaculately rational, and the greatest scientists should be even more rational than merely the mid-tier of non-greats, and so on. Tellingly, von Neumann supposedly blamed himself for missing out on a number of fundamental insights (incompleteness, the Dirac operator, etc) in the many fields he touched, despite being close every time.

Meanwhile, Special Relativity was just that, special. Oh, people will argue over antecedents, but the reality is no one made anywhere close to the same conceptual leap, even if they were at the time playing around with parts of the same math. To everyone else, it seemed as if Einstein had somehow plucked an idea from the gods and put it into human hands. Really, he had plucked it from some sort of netherworld of pre-scientific intuition.

Scientists are not big spheres of rationality. They are spiky and make use of intuition and aesthetics. Get scientists talking over a few beers, and you’ll see. You’ll see that their little spiky bits of irrational attachment make them more interesting, and more effective, than their duller peers.

For despite the best efforts by philosophers and scientists, there is still no clear solution to what’s called the demarcation problem: knowing where science exactly begins and ends. Attempts to define science as merely an abstract machine for falsification, like Karl Popper did, leads to the problem of exactly how one chooses which hypotheses—of which there are infinite—to try to falsify.

In other words, the abstract machine of science is an open system. We can be rational about choosing between different hypotheses, or choosing between different ideas, or choosing between different experiments. But rationality does not actually tell you, by itself, what makes for a good hypothesis, a good idea, or an elegant experiment. Those choices include some strange blend of aesthetics, intuition, passion, and other irreducible qualities.

So there is “Science” in its formal buttoned-down form described in textbooks, and then there is the pre-theoretical fringe that drives science forward and gives it its momentum; its ultimate demiurgic force. It is only this tectonic zone where the pre-scientific and scientific meet that can ever expand the borders of knowledge.

Why does this work? Because beauty is truth, and truth beauty. Much like Eugene Wigner famously argued that there is an irrational and unreasonable effectiveness about how well mathematics can be used in the natural sciences, so too is there an unreasonable effectiveness for psychological traits (like an appreciation of beauty, intuition, etc) when it comes to the pre-theoretical fringe that progresses science’s borders. Wigner’s argument was not broad enough: math is beautiful, and that is why it is unreasonably effective (pulchritudo splendor veritatis).

In this view, great scientists try to force their intuitions and aesthetic conceptions onto science itself, often in ways that lead to incredibly successful breakthroughs but then fail wildly other times. A perfect example is someone like Fred Hoyle, who has a very strong intuition that can be summed up as: “Maybe the sun is responsible!” An intuition which works wonders for chemical elements via nuclear fusion, but despite his best efforts at argumentation, probably doesn’t work for something like influenza, which does not come from space.

So we can see how that same intuitive drive behind scientific progress can be limiting—once used, it cannot be foresaken. “God does not play dice with the universe!” Einstein was absolutely certain of that. And he was certain of it because he thought that some aesthetic dimension of his mind fit well with the universe. The mathematical term for it is congruency—when one shape, laid atop another, reveals their identity. Or as the kids would say, he was vibing with reality. But that congruency, that vibing, earlier so perfect, became mismatched; which is why he introduced the cosmological constant, and didn’t discover the Big Bang 50 years earlier. In this Einstein’s spikiness failed him.

But, here’s my counterpoint: physics isn’t over. Would it not be the ultimate vindication if, in 100 years, our picture of the universe looked much more like what Einstein envisioned? Matching, finally, the aesthetic order that pleased him?

T’would be a strange groove of a brain, to tell the future so.