This might be a silly thought, but what if we weren’t able to understand what the LLM discovered, because of our limited way of perceiving and understanding this information? We’d need a way to understand what the information represents, and it might not represent anything we can understand. Would the machine be able to teach us what it discovers in a way that would make it make sense? Does that make sense?
Am I looking at this the wrong way? It feels very “over my head” but at the same time very much the way I’ve felt about “the laws of physics” and mathematics for a while. This was a very interesting read. Thank you.
That is a very valid concern, and from a human perspective, it would seem frustrating to have laws that can only be used, not understood. I turned your question into a text where I distinguish between explanations we can understand and predictions we can use, and how this relates to science in general.
Are natural laws meant for us to understand the universe?
When I began my master's in physics, I had a single, very human goal: to understand the universe. Yet, as I’ve grown older, I’ve come to realize that the universe is likely to be richer and more complex than any explanation from a human vantage point can pin down. We can try to understand and predict what seems significant to us. Humans have always sought to explain the world around them, through myth, religion, and philosophy. What truly separates science and natural laws from these other frameworks is not explanatory comfort, but predictive power. As long as we follow the recipe of a natural law, we get a valid prediction.
Historically, there has always been resistance when our predictive "recipes" expand past the limits of current human intuition. Negative numbers, rational numbers, and complex numbers were all fiercely rejected before finally being accepted because they were necessary to perform certain operations effectively. We see this in physics, too. As our predictions expanded into domains far removed from our evolutionary environment (humongous distances, quantum scales, or massive accelerations) reality just got strange.
We can do the experiments, formulate the laws, and make the predictions. We can even learn to find them beautiful and develop a mathematical intuition for them. But it is fundamentally harder to understand them the way we understand Newton's laws, which map cleanly onto the physical interactions our brains evolved to handle. There is even a fierce debate today over whether following the mathematical implications of our laws outside their verified, testable scope (like multiverses or infinite parallel worlds) should even be called science. For many physicists dealing with quantum mechanics, the prevailing attitude is simply, as physicist David Mermin coined: "Shut up and calculate." In other words: follow the recipe, make your predictions, and reap the rewards, even if the underlying reality defies human comprehension.
Furthermore, we often use laws in explanatory ways that turn out to be approximate truths. Antoine Lavoisier famously concluded that in a hermetically sealed chemical reaction, the total mass never changes. This was a crucial, beautiful stepping stone toward understanding atoms. But is it perfectly true? No. We know from E = mc2 that binding energy contributes to mass. When you burn a sugar molecule, the release of energy corresponds to an infinitesimally tiny decrease in total mass. We just couldn't measure it because the molecular binding forces are so minuscule compared to the masses involved.
We see this with Einstein, too. He postulated the principle of relativity: that the laws of physics are the same for any observer, and there is no universal reference frame. For a human, this assumption was crucial for formulating incredibly predictive laws. But zoom out to the actual universe, and we find the Cosmic Microwave Background (CMB) radiation permeating everywhere. Not only does the CMB essentially provide a cosmological "rest frame", meaning if you move fast enough, the radiation blueshifts in front of you, but it also possesses distinct structure.
On the order of parts per million, the CMB is mottled with tiny hot and cold spots. Because of this, your exact place in the universe matters. Your velocity matters. Even your orientation matters. If your equipment were sensitive enough, there is almost no conceivable experiment where your location, orientation, and speed wouldn't be theoretically deducible from how you interact with the CMB. So, in practice, our "universal" laws are brilliant, highly useful compressions that have a scope of validity, dependent on the complex, symmetry-breaking reality of the environment they operate in.
All of this brings me to Jodie's excellent question: Will we understand what an AI finds? Maybe. But maybe it will simply provide us with alien "recipes" that we can follow for making flawless predictions, even if we cannot intuitively grasp the concepts they represent. A recipe we do not understand will undoubtedly be harder for us to handle, but it could still be profoundly useful.
Pragmatically, what makes science so uniquely powerful is not its ability to give us intuitive explanations, but its capacity to predict the future. Still, at heart, I remain that romantic who wants to understand the universe. If we ever run this AI experiment, I hope it yields data that deepens our fundamental understanding. In truth, I suspect we will get a little bit of both: new layers of profound comprehension, mixed with mathematics that forever might lie just "over our heads."
I'm confused. A Law requires no understanding of math or physics at all. Science attempts to explain natural phenomena, so that they are understood.
Science often gets stuff wrong, that's why people have to keep doing it, to make it better. Science also does a better job of explaining than do religion and philosophy.
super, super interesting stuff; ill be sure to follow up on his work
Why worry about description? Seems like descriptions (in English?) are also pretty contingent on human cognition.
the first genuinely novel take I've seen in a long time. thank you!
This might be a silly thought, but what if we weren’t able to understand what the LLM discovered, because of our limited way of perceiving and understanding this information? We’d need a way to understand what the information represents, and it might not represent anything we can understand. Would the machine be able to teach us what it discovers in a way that would make it make sense? Does that make sense?
Am I looking at this the wrong way? It feels very “over my head” but at the same time very much the way I’ve felt about “the laws of physics” and mathematics for a while. This was a very interesting read. Thank you.
That is a very valid concern, and from a human perspective, it would seem frustrating to have laws that can only be used, not understood. I turned your question into a text where I distinguish between explanations we can understand and predictions we can use, and how this relates to science in general.
Are natural laws meant for us to understand the universe?
When I began my master's in physics, I had a single, very human goal: to understand the universe. Yet, as I’ve grown older, I’ve come to realize that the universe is likely to be richer and more complex than any explanation from a human vantage point can pin down. We can try to understand and predict what seems significant to us. Humans have always sought to explain the world around them, through myth, religion, and philosophy. What truly separates science and natural laws from these other frameworks is not explanatory comfort, but predictive power. As long as we follow the recipe of a natural law, we get a valid prediction.
Historically, there has always been resistance when our predictive "recipes" expand past the limits of current human intuition. Negative numbers, rational numbers, and complex numbers were all fiercely rejected before finally being accepted because they were necessary to perform certain operations effectively. We see this in physics, too. As our predictions expanded into domains far removed from our evolutionary environment (humongous distances, quantum scales, or massive accelerations) reality just got strange.
We can do the experiments, formulate the laws, and make the predictions. We can even learn to find them beautiful and develop a mathematical intuition for them. But it is fundamentally harder to understand them the way we understand Newton's laws, which map cleanly onto the physical interactions our brains evolved to handle. There is even a fierce debate today over whether following the mathematical implications of our laws outside their verified, testable scope (like multiverses or infinite parallel worlds) should even be called science. For many physicists dealing with quantum mechanics, the prevailing attitude is simply, as physicist David Mermin coined: "Shut up and calculate." In other words: follow the recipe, make your predictions, and reap the rewards, even if the underlying reality defies human comprehension.
Furthermore, we often use laws in explanatory ways that turn out to be approximate truths. Antoine Lavoisier famously concluded that in a hermetically sealed chemical reaction, the total mass never changes. This was a crucial, beautiful stepping stone toward understanding atoms. But is it perfectly true? No. We know from E = mc2 that binding energy contributes to mass. When you burn a sugar molecule, the release of energy corresponds to an infinitesimally tiny decrease in total mass. We just couldn't measure it because the molecular binding forces are so minuscule compared to the masses involved.
We see this with Einstein, too. He postulated the principle of relativity: that the laws of physics are the same for any observer, and there is no universal reference frame. For a human, this assumption was crucial for formulating incredibly predictive laws. But zoom out to the actual universe, and we find the Cosmic Microwave Background (CMB) radiation permeating everywhere. Not only does the CMB essentially provide a cosmological "rest frame", meaning if you move fast enough, the radiation blueshifts in front of you, but it also possesses distinct structure.
On the order of parts per million, the CMB is mottled with tiny hot and cold spots. Because of this, your exact place in the universe matters. Your velocity matters. Even your orientation matters. If your equipment were sensitive enough, there is almost no conceivable experiment where your location, orientation, and speed wouldn't be theoretically deducible from how you interact with the CMB. So, in practice, our "universal" laws are brilliant, highly useful compressions that have a scope of validity, dependent on the complex, symmetry-breaking reality of the environment they operate in.
All of this brings me to Jodie's excellent question: Will we understand what an AI finds? Maybe. But maybe it will simply provide us with alien "recipes" that we can follow for making flawless predictions, even if we cannot intuitively grasp the concepts they represent. A recipe we do not understand will undoubtedly be harder for us to handle, but it could still be profoundly useful.
Pragmatically, what makes science so uniquely powerful is not its ability to give us intuitive explanations, but its capacity to predict the future. Still, at heart, I remain that romantic who wants to understand the universe. If we ever run this AI experiment, I hope it yields data that deepens our fundamental understanding. In truth, I suspect we will get a little bit of both: new layers of profound comprehension, mixed with mathematics that forever might lie just "over our heads."
I'm confused. A Law requires no understanding of math or physics at all. Science attempts to explain natural phenomena, so that they are understood.
Science often gets stuff wrong, that's why people have to keep doing it, to make it better. Science also does a better job of explaining than do religion and philosophy.