AI Crisis in Math: Navier-Stokes, Riemann Hypothesis & Complexity Theory | Lex Fridman Podcast #503
Is it worth your time?
You get a detailed account of how automated theorem provers and large language models are transforming pure mathematics, complexity theory, and fluid dynamics.
This conversation explores how recent breakthroughs in AI are abruptly automating the proof of centuries-old mathematical conjectures, forcing researchers to rethink their methods and institutional structures. Scholars examine the friction between mechanized problem-solving and the human pursuit of abstract understanding.
AI solving major mathematical problems. Kontorovich warns that the sudden release of AI-generated proofs for landmark problems, such as a quasi-Riemann hypothesis establishing a zero-free vertical strip between seven-eighths and one, is extraordinarily shocking. He expects that forty to eighty percent of everything discussed will change within five years because systems achieved a millionfold leap in automated theorem proving. He points out that this success completely exceeded previous expectations of what automated systems could accomplish.
Formal verification with Lean and Mathlib. Brunton notes that automated formalizations using interactive theorem provers like Lean allow researchers to verify technical building blocks down to basic logical rules of inference. He explains that Mathlib serves as the crowdsourced library organizing these proofs. He emphasizes that these tools help researchers shift focus toward the broader narrative of mathematical ideas.
The changing work of mathematicians. Moitra argues that automating the proving of theorems disrupts the traditional playbook for graduate students and postdocs who rely on solving hard problems to make an impact. He warns that an overabundance of AI results creates anxiety and dread, threatening the slow, patient research required to grow deep mathematical insights. He emphasizes that mathematicians must recenter their values on asking interesting sub-questions rather than merely churning out papers.
Navier-Stokes blowup and fluid dynamics. Brunton notes that AI generated a counterexample to Navier-Stokes showing finite-time blowup by creating a specialized vortex that stretches and thins until vorticity becomes infinite. He explains that while this is a major mathematical result, day-to-day engineering applications in designing airplane wings remain unchanged because the theorem operates in a non-physical regime. He states that understanding these complex fluid dynamics requires finding simple coordinate systems that generalize physical laws.
Crisis moment in mathematics
This is a chance to rethink what we are trying to accomplish. Some people are reacting positively or negatively to this chaos of opportunity, and sometimes those reactions happen in the same person. We are entering a golden age of mathematical knowledge; after millennia of walking through the woods with a candle, we now have a flashlight.
Many people didn't expect this amount of progress in this amount of time, so it is truly shocking—the kind of result where, if a human did it, it would be an instant Fields Medal. However, it was not entirely unexpected, as there had been rumors for weeks that the unique games conjecture and RL equals L, the two biggest problems in my areas, were in that collection of results. There were even rumors about matrix multiplication. My first reaction was, "Wow, it's all true." I'm glad it's finally out there so people can process what happened and figure out what to do next.
For me, the overarching feeling is being pulled in two different directions. My work has always been about asking the right questions, and these tools are amazing because they increase my ability to learn about other fields. It is like having a world expert teach you in your own language, magnifying your ability to make connections.
But I love the mentorship aspect of my job—not just training grad students and postdocs, but having them come with me on these journeys. The playbook and blueprint for success we had for so long is now thrown out the window, making it harder to advise students on what work will make an impact or make a name for themselves. I feel excited, but more intellectually lonely, because great results often come from asking humble questions. Now, there is so much anxiety and dread that you must do something momentous to stand out in the sea of AI results that we have lost the patience for planting the seeds from which great oak trees grow.
Fields Medal winner Terence Tao wrote, "I am horrified by the many person-years of ongoing patient and deliberately slow research efforts, particularly by graduate students and postdocs, toward many motivating problems in mathematics being casually disrupted or destroyed by such a release." Much as one cannot unhear a movie spoiler or a crossword clue,
One cannot explore a problem as profitably and richly once aware of an existing solution, and I mourn the path not taken and the opportunities lost in the frantic race to develop the technology. Last December, most of the programming world switched to AI writing most programs, creating a state of shock. I would approximate it to a feeling of loss of meaning and terror.
Most of my feelings are hopeful. I like solving problems that make technology more realistic for humans, such as artificial hearts and airplanes. Our ability to solve real applied problems, as well as our scope and reach, is going to get much better. However, there is uncertainty; the fear of losing control or being cut out of the process we love is scary. I would not be surprised if, over the next five or ten years, we are either dwarfed by other intelligence or it levels off and becomes an incredible tool for our curiosity and human problem-solving.
This is difficult for those of us who chose math because we liked rules and formalism, and uncertainty is uncomfortable. Solving interesting problems, helping others, being curious, and wanting to understand the world are good things to get back.
I agree that the path is as important as the destination, like how a road trip is usually as fun as getting there. At the same time, if we found out another country had invented teleportation or time travel, I guarantee we would reinvent it in the next five years. If you knew it was possible and could reverse engineer it, we would be curious, dive in, and figure it out. Knowing there is a solution is also an incredible motivation to get there much faster.
Will AI "solve" mathematics?
Mathematics as we know it may be done completely by machines, leaving humans to do something fundamentally different, or perhaps something similar to our current practice. One path solves math; the other destroys it.
When chess engines first emerged, people imagined "centaur games" where humans and computers played together. That lasted ten seconds; computers became so good that if you didn't choose the top recommended move, you lost. I recently axiomatized an argument with my colleague Thomas Koberda about the future. For the sake of argument, I accept the "solved axiom": any mathematical problem that can be solved by a human interacting with AI could also have been solved by AI alone. This means there is nothing new for a human to contribute; to make progress on problem X, you simply press a button.
I don't believe this destroys pure mathematics. While engineering mathematics might cure cancer or create jet packs, we prove pure mathematical theorems because we are curious. Solutions living as ones and zeros in data centers are not valuable to human beings; the value of pure mathematics is that it lives in our heads. Now that AI has solved problems we thought were far off, I must develop more generalizable and abstractable ways of reaching the finish line to explain connections or other directions. To me, mathematics is a vehicle for abstraction and understanding.
You're building structures that help you attack a whole sea of problems rather than just one. Sometimes we reason in sloppy metaphors to conceptualize the world. As Alex mentioned offline, attention is all you need; it is the human ability to shine a light on particular areas that matter.
Before AI, you could think of axioms and the logical deductions or theorems provable from them. There is a much larger space of provable theorems than the space human beings actually care about, so it is a tautology that the theorems we get excited about are the ones we care about as humans. AI might convince us there is interesting structure, like how Navier-Stokes is connected to fluids and all natural systems, or it could be a mixture of beauty and difficulty that pulls at us. Prime numbers are full of complicated problems that may not have an obvious immediate application, but they pull at us. As a symptom of pursuing the hard thing, great applied mathematics, engineering, and solutions to real-world problems emerge.
When working on a problem, I might have a bigger goal, but I usually try to find the simplest, most embarrassing open problem I still don't know how to solve. Looking for a proof is like exploring a tree of possibilities, but we sometimes encounter points with a really high branching factor. In this centaur world of math, we suddenly have something else intelligent we can interact with.
I like the doorman fallacy, where you take a complex job and boil it down to its simplest, most visible task. If you digest a doorman's job down to that, you might think you can replace them with a sliding door that detects people coming and going, missing all the aspects of their job that add value and are harder to articulate. The simplest and most visible part of our job is proving theorems. Theorems about conjectures open for hundreds of years, like Fermat's Last Theorem, are very visible things we've spent a lot of time popularizing.
I think of a mathematician as someone who thinks deeply. A byproduct of that is proving old conjectures, but it is also building new theories, new ways of looking at old things, and making new conceptual connections.
It is the right time to recenter our values and remember that our field has a plurality of aesthetics and goals. When one thing becomes cheap, like problem-solving for well-defined things, the orthogonal goals become all the much more valuable.
The doorman fallacy is the mistake of reducing a person's job to its most obvious task and assuming that automating that task replaces the value of the whole role. An automatic door can open more cheaply, but a doorman also welcomes guests, recognizes returning customers, handles luggage, hails taxis, provides security, discourages unwanted visitors, and signals prestige to make guests feel cared for. The task for humans, and mathematicians specifically, is to make explicit exactly what it is that we do. From my perspective, mathematicians are expositors and scholars of things beautiful, broadly defined.
Until recently, there was a near perfect correlation between the ability to solve new deep theorems and what we actually want as human mathematicians: beauty, understanding, and depth of thought. Now, you can press a button and get what used to be our proxy without effort, deep thought, or necessarily knowledge.
Beyond answering questions and problem-solving is asking the right questions. Good problem-solving is measured by asking interesting sub-questions that can be solved and interpreted. The value of asking good questions has always been something we value in our colleagues, and it will become increasingly more valuable. This requires significant mathematical insight. I would not minimize the role of mathematical insight in both tasks; when we say talks will become more important, I do not imagine we will all become Twitter or TikTok influencers. What stands out is having ideas that get adopted by your community and become powerful over time.
I am most worried in the near term about what will happen to our systems and institutions. That blueprint has changed because we are overloaded with the results happening right now, and it is okay to take time to rethink it. We still care about ideas that influence the community, but they are harder to operationalize into a simple rubric. It will take reworking our systems and institutions to do that.
Can AI create new math?
I can prove a theorem without understanding why the proof works. There is some magic that happens, and until I generalize the phenomenon and understand it more deeply, I won't see why the proof works the way it does. This is perhaps why I don't believe the solved axiom; our work is exploratory, and a good question is more than just making progress on the Riemann hypothesis.
Regarding beauty in mathematics, the gold standard has always been how simple a result is and how much it generalizes. Finding a simple, generalizing structure within the infinite space of provable theorems remains the gold standard.
We can prove the existence and uniqueness of Navier-Stokes with AI, but humans are guiding this. I want it to start telling me what questions I should be asking. There are research programs systematically and automatically discovering minimal physical laws; if you observe balls dropping, you can deduce gravity and F=MA, or observe fluids to deduce Navier-Stokes. We haven't yet written down simple equations for data like weather, turbulence, and solar flares, but I assume similar minimal pieces can be deduced in math.
My version of the solved axiom is that anything I contribute as a human mathematician could also have been contributed by an AI on its own. I have tried to argue that LLMs could never do what I do, but I must ask, "Why not?"
Even in that world, we still need people with human mathematical expertise. Figuring out which half of that is true will be interesting and challenging. When facing a technology this good, the danger is anthropomorphizing it to conceptualize why it works. Neural nets can achieve human-level or better object recognition, but their success is less robust. Adversarial examples—slight patterns added to natural images—can completely fool these systems but not a human. My former colleague at MIT, Alex Madry, did work I enjoyed titled something like "Adversarial Examples Are Features, Not Bugs," where you train a deep network to identify natural data, like pictures of cats and dogs, very well.
You add adversarial perturbations to images that make a cat look like a dog to an AI system and vice versa. Amazingly, when you train a deep network on this with all the wrong labels, it learns something that performs well on natural data. Often, AI finds trapdoors or backdoors to solving problems using statistical relationships in the data that humans don't. When I look through OpenAI solutions, they don't look like a human solved those problems. To me, there is still a big qualitative difference, which is an opportunity for collaboration and centaur math rather than dread about AI overlords.
Regarding move thirty-seven, some people speak dismissively, saying it was in the convex hull—that the AI just did moves humans know, but did them really well with a good tree search. Tim Gowers had a better way of describing it as the semigroup generated by all these things. It is the same notion of combining known moves, but the space of what it can accomplish is much bigger than the derogatory convex hull. In terms of the quasi-Riemann Hypothesis, the moves it took were all known to human beings, but the long cascade of them is shocking and beautiful.
To explain the intuition between a semigroup and the derogatory convex hull: think of a balloon. The inside is its convex hull. All the ideas mathematicians have thought up create this enclosed area, and anything inside is some linear combination of them. AI hasn't really made new progress because it hasn't moved the balloon out. It is derogatory because the intuition is that a mathematician could have done it since it was already obvious.
AI has read every paper—more than any human brain could hold, read, and internalize. In terms of breadth, it beats any human being. In the case of quasi-RH, I would not say it's a genuinely new idea, but a beautiful combination of well-known things. Many people are kicking themselves for having ideas in that direction twenty years ago. Two years ago, there was a big breakthrough on cubic theta functions and cubic Gauss sums, which ended up being the path the AI took to the solution.
To contrast that with a semigroup, I think of a semigroup as tiles. You're playing tiles on the board, and whenever you have a tile, you can put another to the left, right, front, or back. Your intuition is more in the semigroup direction, meaning there are glimmers of expansion versus linear combinations inside the convex hull. Existing ideas might prove a whole lot more than we as humans have been able to explore.
Throughout the history of mathematics, seemingly difficult and deep problems have had three-page elementary solutions that could have been explained to an undergrad or discovered twenty years before the problem was first posed. In my proceedings for the International Congress of Mathematicians, I wrote that we're going to discover many problems we thought were unreachable by current technology actually are reachable if put together in just the right way, and that's what AI is amazing at.
The important thing about the balloon is the boundary between what's inside and outside; what makes it derogatory is that you haven't seemingly changed the boundary. There were a couple of field-altering results that will change what types of problems you give to young researchers and what areas you explore because your conception of the landscape is different.
One was an Anthropic result written up by my colleague Virgie Vassilevska Williams and Josh Alman regarding the 3SUM problem. It has a major effect on fine-grained complexity theory because a basic problem given in coding interviews can actually be solved asymptotically much faster than any known solution, which was counter to what the community was betting. In fact, a whole theory was built around the idea that if this problem really is that hard, then all these other problems inherit its hardness.
Imagine if one of the seven hundred was factoring in polynomial time. You couldn't use your credit card on Amazon; all of e-commerce and secure communications would be gone. If you give me two million-digit numbers, I can multiply them in a split second, but we do not know how to factor the product back into two large prime numbers. The modern world and public key encryption depend on the fact that we lack an algorithm for pulling prime factors out of a large number.
Famously, there are quantum algorithms for factoring, so there is emphasis on finding post-quantum bases for cryptography, such as lattice-based methods. We are concerned about what will happen when quantum computers are large enough to make this an actual problem, though we still have a little time. Certain mathematical results obviously lead to dramatic impacts; this one would affect the economic systems of entire nations.
Ten years ago, evidence showed that collaborative teams of mathematicians could solve much harder problems than any individual if they broke them down the right way. I prefer the semigroup analogy to the convex hull, but it shouldn't be interpreted as derogatory even if these machines are interpolating human knowledge rather than extrapolating. Even if there is a convex hull of ideas pieced together from human inventions, we aren't doing that fast enough and the balloon is not see-through. Lifting the fog and giving mathematicians radar to see over the horizon is incredible.
Human beings make new definitions and bring in new concepts, which feels fundamentally different from repackaging existing ideas. When they do that, it creates a new boundary point, and we ask what the convex hull of what you could prove is now. You can pull on the balloon, and we know when it expands. The archetype of the researcher is going to change because of this new technology.
People who think broadly about a bigger swath of mathematics will be in the best position to benefit from these technologies. For decades, research has centered on the paper. In engineering and applied sciences, we already have a rapidly shifting view of what an academic contribution is. If you write reusable code for larger infrastructure supporting billion-dollar astronomy projects like LIGO or massive cancer research projects, the contribution is not necessarily one 15-page PDF. It consists of code, data, and reusable bits fit into a larger framework.
This is even more relevant with AI; having a workable widget, API, codebase, or dataset that can be leveraged by the community or AI is increasingly valuable, even more than one result or paper. Observing the evolution of arXiv, which is similar to what is happening on GitHub, many in the open source movement are struggling. Some predict the death of GitHub because it cannot sustain the growth of code being added. The open source movement succeeds when a few humans manage each repository, but they are becoming psychologically overwhelmed by pull requests. Like the academic environment, open source requires motivated humans; if you get worn out, as Terry Tao mentioned, you just quit.
Math journals were under severe strain before AI because more people were producing more knowledge. Referees are exhausted, working hard to understand and check theorems because our mathematical epistemology is deliberately slow, careful, and precise. Consider Perelman's solution of the Poincaré conjecture. He put papers on the archive, and after being invited to lecture at Stony Brook, NYU, and Princeton, he gave three-hour impromptu explanations. He might spend an hour explaining his work, and the next two hours addressing specific questions, such as a lack of understanding regarding Lemma 412 on page 247, which he would then explain at the board. He communicated as best he could on paper, but it took years of detailed, precise human effort for the community to agree it was a real proof. Coming up with a solution is only the first step in a deliberate, community-driven process.
In many other fields, we have moved past the point where individuals can advance circuit design, aircraft design, protein folding, or astronomy; we must work together in billion-dollar collaborative projects. As an engineer, I believe leadership is what a community needs when facing this crisis of organization and individuality as it moves toward a more collective model.
Mathematicians are a very independent group of people who don't like being told what to do or what to study. Taking a contrarian view, you should never let a good crisis go to waste. Journals and conferences were already overloaded, and now we face the task of reviewing AI-generated papers. While some are important, it is hard to get the community on board. I favor finding new models, not to replace the old ones, but to relieve these pressures.
I would love to see more collaborative efforts to verify AI proofs. For example, some OpenAI results regarding improvements on Fast Fourier Transform led to tweet streams with people improving the constant. We need other modes of working to organize the state of improvements and tame these ideas. Currently, we work on a draft and post it on arXiv, but we need models that allow crowdsourcing for our understanding to evolve.
We can now trust Lean formalizations for some technical building blocks, allowing us to rework our writing and emphasize the promulgation of ideas. The Mathlib model is exactly this: instead of a paper, you have a pull request. In a GitHub open workplace, maintainers and reviewers see exactly who contributed the idea that simplified an argument.
Working formally with Mathlib solves a major problem: mathematics lacks perfectly well-written specifications for theorems and technical lemmas. In a random math book, authors omit context because filling it in is annoying, assuming you have read the previous sixty pages. This means you cannot take a lemma off the shelf like a physical part when building a rocket. If mathematics is done formally in the future, specifications are built into the object. Lean won't let you use it the wrong way, and that is the crowdsourced version of mathematics.
The dream of Mathlib when it started ten years ago was to run mathematics more like a science. Ideally, mathematics is a formal process starting with an axiomatized system and logical rules of inference, a model kids use to see proofs and axiomatic structure for the first time. You can write this in software, defining the actual axioms and rules of logical inference. Lean is part of a long history of interactive theorem provers where you write a proof line by line, and the computer confirms each line follows by logical inference from the previous goal state.
Lean allows you to check that code specifications align with your definitions; it was invented for bug-free code. This application is becoming ever more important because AI systems are excellent at finding bugs in human-written code. Another application is checking mathematical results by writing them in a very formal way. Lean itself is the hardware that checks the logical steps. Because you do not want to start every math project by defining natural numbers using zero and the successor Peano axioms, it is useful to have a large, well-maintained, and organized general library. That is Mathlib: the Lean library for formalizing mathematics in the Lean ecosystem.
Lean is one of many interactive theorem provers, but it has been winning the math wars, if there are any. It is the technology used in part to show that the proofs in the OpenAI release are correct. Autoformalization has reached the point where AI systems can prove a theorem in natural language and write Lean code that builds on Mathlib or the Prime Number Theorem Plus project to check the solution down to the axioms.
Since there are only a few logical rules of inference and a handful of axioms, the only way mathematics becomes not understandable is scale. Formalization gives us confidence that results are correct, but we often understand results at an unformalized level. If we have black boxes, we try to tease them apart by asking easier questions to develop a system of understanding how they reached the answer. The engineering of the proof and the communication of that proof to someone else are almost at odds. When I taught real analysis last fall—and this was my experience taking the course—the first couple of lectures are always: what are the real numbers? We have to construct them as equivalence classes of Cauchy sequences of rationals, whatever the hell all that means.
In a first-semester course in real analysis, it takes two weeks talking about Cauchy sequences of rationals to construct the real number line before we can start doing calculus with formal epsilon and delta definitions. You need to construct the real numbers before you start using them, though human beings used them long before we understood that formal construction was necessary.
Consider the Bolzano-Weierstrass theorem: if you have a sequence of real numbers jumping around in a fixed range, some subsequence must converge to a real number. That is not true in the rationals. I taught the course backwards from how an engineering project would be taught; you cannot talk about real numbers until you know what they are. But if you are telling a story, you can start with the real numbers. You have an intuition for what they are, and until that intuition breaks, you do not need the resolution of their construction.
I think the way we write papers in the future will be different, spending more time on the story of what was learned in the process. You do the standard thing here and there, but then you get stuck. Here is an idea for why you go left instead of right, up, or down. A Lean formalization would accompany this story to show that the engineering works out. Math is not music, where you can tell a beautiful story and it is okay to break a rule, like Bach moving parallel fifths.
I do not believe in the view of mathematics as just a rising tide of theorems. Mathematics is not just about the number of theorems we have; it is about the story of fighting with problems. When you fight with a good problem, it fights back, and you learn something. It is about the system of beliefs and abstractions you discover en route to a more general meaning. Telling the story of how we got to these problems and their parallels to others is now more important than ever. You can follow a proof line by line and not understand what is really going on, perhaps asking an LLM to explain the story if it can.
Riemann hypothesis
The central problem in my field is the Riemann hypothesis. We know what the prime numbers are, and people are interested in their distribution. Riemann discovered a function that takes complex numbers as inputs and outputs, and this function governs the behavior of the prime numbers. It is called the Riemann zeta function. By the zeros of the zeta function, we mean complex inputs where the output is zero, the origin in the complex plane. We have these discrete numbers that aren't divisible by anything else sitting in the natural numbers on the number line, and now we are talking about functions from a complex space to a complex space.
The great breakthrough was figuring out how to do signal processing on the prime numbers. Just as a microphone uses Fourier analysis to convert my voice box's pressure waves into amplitudes and frequencies, you can think of the prime numbers as particles of air molecules wiggling around. Signal processing analysis can recover those same prime numbers, where the frequencies and amplitudes correspond to the zeros of the zeta function. The superposition of these "Riemann waves" recovers the distribution of the primes. The Riemann hypothesis predicts that all of those amplitudes are exactly the same, like an orchestra playing the music of the prime numbers with every single instrument sounding at the same amplitude.
I did not think I would see a proof of the quasi-Riemann hypothesis in my lifetime, so seeing it on Tuesday was extraordinarily shocking. I was socializing at the National Museum of Mathematics Gala in New York City, but I kept refreshing because I knew it was dropping around six o'clock Eastern Time. I was chatting with a Quanta reporter about how they were going to drop 700 problems, and then I looked and saw the zeta function and Dirichlet functions; they've solved no zero, zeros.
I was sure we were at least three years away from formalizing Fermat's Last Theorem. Two days prior, I gave a colloquium in the Princeton Math Department explaining Lean and Mathlib. In Mathlib, everything is perfectly done with the right levels of generality; there is no code reproving theorems elsewhere. AI systems can build one or two layers on top of that, but they are jerry-rigging. They aren't yet good at canonization—refactoring sloppy code to find the canonical version of a lemma so that every proof is simply a combination of existing lemmas.
In my colloquium, I said you can get entire textbooks formalized as long as they are in the Mathlib halo. But in my experience, when you ask AI to do things three or four steps removed from the library, the definitions it creates for the same object for different purposes clash. Mathlib needs to grow to reach a place where you can do research manuscripts that synthesize tons of different areas of mathematics. Because Fermat's Last Theorem is so far outside the Mathlib halo, I expected it to take quite a bit of time.
Kevin Buzzard had a million dollar grant to spend the next five years—or perhaps he was already a year or two into it—to reduce the formalization of Fermat's Last Theorem to things known in the 1980s. He wanted to do those last seven years of work by Wiles and Taylor, built on top of things that would be extraordinarily difficult; he wasn't going to do Langlands' Tunnell or Mazur.
Then on Friday, there was a drop of thirteen million lines of code from the axioms, and Fermat's Last Theorem is proved. The statement is completely elementary, and Lean says it is sorry-free. Anthropic did it; AI systems just got really good at jerry-rigging. It's still not a nice proof, but thirteen million lines later, they had a sorry-free formal Lean proof. That's the fun thing about human discussion: we have disagreements, and I would venture to say forty to eighty percent of everything said today will be wrong in five years because everything is changing so fast.
When the first LLMs came online, they were crap at Lean, but as they improved, I realized I could use them for library searches. I no longer had to sit there knowing a lemma existed but not knowing its name. Every bit of my experience told me systems could do X, but not ten X; this was a million X, completely outside my distribution of what systems were capable of.
I believe it took two or three humans whose involvement was mostly shouting "KEEP GOING," as the meme goes with OpenAI or Anthropic. While the final prompt is "do a breakthrough," there could be groundwork or intermediate steps in the training that we don't know about.
I remember it was a Friday and my mind was completely exploded. It was Labor Day, so I took the kids to the beach and needed a few days off my phone and Twitter. I realized that autoformalization is basically solved. It's not solved in the way I would want, but in terms of getting a sorry-free proof of something humans wrote on paper, that problem is done. When I got back from Labor Day, it was Navier-Stokes with the quasi-RH.
Primes are valuable because they are the fundamental particles in the theory of multiplication; every whole number is broken down into a product of primes. Just as elementary particles make up all matter in physics, primes make up the multiplicative structure of whole numbers. Their distribution remains a mystery, and even the Riemann hypothesis doesn't solve everything we want to know. But the Riemann hypothesis was beautiful because it pointed to what the distribution should be and has a thousand applications in faraway fields; if it were true, we would know XYZ.
Grant Sanderson was visiting IAS, and we had a long discussion about what AI is doing to the math community. I really enjoyed a comment Camillo De Lellis made.
There was a play written in the 1950s by a mathematician where Von Neumann asks God, "Is the Riemann hypothesis true?" Everyone holds their breath, and God says, "Yes." When Von Neumann asks how He knows, God replies, "Well, I know where all the zeros are, and they're all in the R equals a half line." The premise of the play is that this is not enough for Von Neumann. Math is not the fact, it's the why; if you tell me a fact is true but I don't understand why, that has contributed zero to my knowledge.
Primes are interesting because they relate to things a three-year-old can understand. Fundamentally, we know how to add one to numbers, and there's that breakthrough moment as a kid: I can add one forever. Primes have always been interesting to me because we don't understand how adding and multiplying work together. I can add one to a number that I can factor and have no idea what happens. I know if it's even, it becomes odd and vice versa, but I don't know if a number with many composites becomes prime. Primes are the distillation of that three-year-old mystery that all of us still wrestle with.
He had a pack of 20 and would make a four by five grid or a 10 by two grid. But he only had 19 once, and he couldn't figure out a way to make a nice-looking rectangle out of the 19.
I did not think I would see a proof of quasi-RH. There is a Lean formalization; the Lean statement was already in Mathlib, so the fact that they have a sorry-free proof gives me 100% confidence that the proof is right. Embedding the Möbius function and twisting it by sextic characters is the bread and butter of analytic number theorists, and it was chained in just the right way and was absolutely incredible.
He showed there is a critical strip, a vertical infinite strip in the complex plane where the X coordinate lies between zero and one, and the Y coordinate can go arbitrarily high or low. The hypothesis is that the zeros all lie right at the dead center of that strip, where the real part is a half.
Until Tuesday, we knew of a zero-free region of the zeta function that grew thinner the higher you go. The prime number theorem tells us the basic distribution of prime numbers is equivalent to the fact that there are no zeros on the line where the real part of S equals one; as you go up and down that line, the zeta function never hits zero. This thinning zero-free region gave us the best known distribution of prime numbers and the best error term in the prime number theorem.
I anticipated the line might thicken slightly, though still approaching zero as you go up the imaginary axis. That thickness hasn't moved in seventy years, but I imagined some progress could be made. Instead, we got a vertical strip: there are no zeros with a real part between seven-eighths and one. This is a quasi-Riemann hypothesis; if you can get that seven-eighths down to a half, that is the full Riemann hypothesis. However, the methods used in the seven-eighths proof—and a second proof regarding eleven-twelfths—fundamentally cannot get down to one half.
They prompted thousands of problems but only dropped the positive solutions, with about a ten percent success rate. Once those ideas are out there, it might be more straightforward for AI to balance that one missing ingredient, that new tile, and exhaustively explore what it can reach. Other communities have models for valuing contribution without lionizing one individual. While there is a way forward, this is a short-term roadblock to doing and enjoying mathematics the way we used to.
Navier-Stokes
Fundamentally, the Navier-Stokes equations state the conservation of momentum for fluids, such as gases and liquids. Just as F=ma represents conservation of momentum in classical Newtonian physics, the same principle applies to any volume of gas in this room or flowing water. If you write down that conservation of momentum in a volume and take the limit as that volume becomes arbitrarily small, you get these Navier-Stokes equations. This partial differential equation describes how the velocity field of a fluid changes in space and time, and the pressure related to the forces acting on that fluid.
Writing equations for a single pendulum is easy, but adding another pendulum creates a more complicated equation that gives chaos. Modeling the air in this room is similar; I could chop the room into a billion little boxes and write a differential equation for the momentum in each, but that creates a complicated system of equations. While it is simple to conceptually understand what the equation represents, solving it to understand how fluid moves in a practical sense is very complicated.
Fluid dynamics is important for reducing drag on airplanes; airlines could save tens of billions of dollars by reducing drag by one percent. Drag is perhaps easier than turbulence, but if you can improve our ability to predict or simulate turbulence even a little bit better, faster, or more accurately, you can have a huge economical, ecological, and health and safety impact across dozens of fields, including better artificial hearts and more efficient wind and gas turbines. I may want a wing with high lift and low drag, but I don't know how changing the geometry will affect the cascade of smaller and smaller vortices that turbulent fluid generates over the wing, resulting in its lift and drag.
Part of the challenge is that equations are expensive to simulate and wind tunnels are expensive; you need supercomputers. As Feynman noted, we know how to write general equations like Maxwell's, Navier-Stokes, and Schrödinger's, but we don't know from the equation what phenomena are possible. If you had never visited Earth, you could not have predicted sunsets and tides just from a partial differential equation. We still must simulate, test, and use an elaborate engineering process to understand how these equations manifest as drag on an airplane wing or turbulence inside an artificial heart.
In the case of Einstein's theory of general relativity, he wrote equations and a few years later Schwarzschild found a blowup. People thought, "If you get enough matter into a small enough space, light can't escape it," until we discovered black holes and found the theory perfectly predicts what happens.
Regarding the blowup of Navier-Stokes, will we discover something new about water, or is it that at some scale you reach an atomic structure that no longer follows liquid equations? The engineering answer is that these equations are an exceptionally good model of how fluids work at a certain scale; you need a number of individual molecules averaged over in a volume for these constitutive equations to be valid. If you zoom in small enough to individual molecules, Navier-Stokes breaks down. By the time they blow up, it stops being a fluid.
From a philosophy of science perspective, where equations break down is often a window or door to a breakthrough. Newtonian physics breaking down needed Einstein. We still lack grand unification because we don't have gravity at the same scales as subatomic particle interactions. Many people are simulating and testing Navier-Stokes in every industry imaginable. Does it blow up in finite time? Does it explain the phenomenology of sunsets, hurricanes, and tornadoes? Could a solution with a blowup be used in an engineering application to drive a state that helps airplanes fly with less drag?
Birds and insects shape vortices and funnel energy into structures to perch on a rock or hunt. Flow control is one of the highest challenges, and our science fiction technology would be accelerated if we had a better ability to control these flows. I'm reminded that when Feynman was asked about artificial intelligence in the '70s, he said the way we solve problems with technology is radically different from how nature solves them. He pointed to how birds fly versus how airplanes fly. Our airplanes don't flap.
A fluid moves differently in different places and at different points in time. The basic building block is the velocity field: at every point in space, a velocity vector tells you how that fluid is moving, whether it is fast, slow, or creating a vortex. It gets more interesting when the fluid velocity changes rapidly over space and time, like a tornado. Pressures force this velocity field and cause it to move; for example, the core of a tornado is low pressure.
Equations describe how that velocity changes in time and space according to its own momentum, such as gusts of wind. It also has internal friction and resistance, called viscosity; honey has more internal resistance than water, which has more than air. External forces like heating, pressure, and gravity are each quantified as a term in the larger Navier-Stokes equation.
If you don't have that viscosity, we have a strong suspicion that the equations can and will blow up under certain conditions. We have simpler examples that do blow up. For instance, in a simple ordinary differential equation where X dot equals X squared—X being the state of my system and X dot the rate of change of X in time—the equation blows up in finite time because of that quadratic nonlinearity. Compare this to X dot equals X, like a population of bunnies. That grows exponentially, doubling and doubling, but it won't reach infinity until infinite time has passed. That is normal exponential growth, not finite time blow up.
There is a progression of partial differential equations with this same behavior. Interestingly, the convective term in the Navier-Stokes equations, which tells how the fluid moves itself with its own momentum, is a quadratic nonlinearity just like X dot equals X squared. Without diffusion or viscosity, we think this self-motion should probably blow up to infinity in finite time. However, we don't know if viscosity always damps out those blowups before they can happen.
My intuition as an engineer is that they do. If I have a flow accelerating to infinity, it should immediately diffuse into its neighboring fluid and tamp down that growth. The math problem is essentially whether there are any edge cases or initial configurations where you perfectly orchestrate the fluid so that it actually accelerates to infinity in finite time. There are two technical versions of this: one with no external forcing, and one with smooth external forcing. Smoothness is important because if the external forcing has singularities, like an infinitely tight tornado, then of course it's going to blow up.
The problem as stated has smooth external forcing. It is proven that a smooth external forcing can be orchestrated to perfectly cancel various terms in the equation to create a vortex. This can only happen in three dimensions.
There is a big difference between 3D and 2D. While we usually imagine fluids in 3D, there are simplified 2D flow analogies; for example, weather on Earth is well approximated as a 2D flow. In three-dimensional turbulence, big vortices generally break down into smaller, itty bitty vortices until they molecularly diffuse into heat. In two dimensions, you get a reverse cascade where little vortices coalesce into medium and then bigger vortices, which is why hurricanes form.
The AI proof of this counterexample showing finite blowup creates a specialized vortex that stretches and thins in the third dimension until it becomes infinitely thin. Like a figure skater doing a spin, the radius gets smaller as it stretches. There is a specialized rate of contraction so that energy in the fluid remains finite, but at this localized filament, the flow blows up and you get infinite vorticity. This is physically counterintuitive because our intuition tells us diffusion solves the problem.
This is a very important math result, but day-to-day engineering fluids and applied math analysis remain almost completely unchanged. Fluids are just as hard to simulate and predict today as they were two months ago. When I solve an engineering problem—designing an airplane wing or an artificial heart—I must still simulate and manipulate flow in a practical way that is expensive in terms of supercomputer time and wind tunnel resources. This theorem operates at a regime so non-physical that it doesn't affect my day-to-day engineering.
There is a trillion-dollar engineering problem in how to predict these things in a fast, efficient way and perform inverse design. This represents a trillion dollars of possible benefit to humanity through better, faster solutions, lower drag, and fewer emissions, as transportation, defense, health, and energy all have fluids at their core technology. This math result is important because we don't know the fundamental character of what is possible in these equations; we don't even know if the solutions exist.
Turbulence is difficult because it is multi-scale. It is a phenomenon where a complex fluid roils, mixes, and moves like Da Vinci's drawings of fluids at all different scales. The big, medium, and little vortices all matter to predict lift and drag. However, the ten to the twenty-third gas particles and their individual configurations do not. Thermodynamics, a great success of classical physics, shows that a small handful of numbers—temperature, pressure, and density—determine how gas behaves.
That simple description doesn't exist for turbulent fluids, such as flow over a wing, inside your heart, or on a river. We need a representation of a turbulent fluid that allows it to be simply described for engineering motors and engines, as we have for gas dynamics. Whether the solution lies through mathematics or an AI system watching fluid, I believe our great physics breakthroughs were predated by a coordinate transformation. Newton's F=ma came after we put the sun at the center of the universe instead of the Earth.
We must take these simulations and observations of fluids to find a simple coordinate system so the description becomes simple, generalizable, and physically beautiful. This requires both brute force—trying different transformations with lots of data until the simplest one generalizes best—and deep mathematics regarding the form of that transformation. Questions smart people asked in the '70s and '80s but hit a computational wall are fertile ground now. We know from fundamental principles what the form of these coordinate systems should take; this is a perfect problem for AI to learn these transformations using more data and compute.
I love watching videos of two vortices, like a red and blue one, shoot together, influence each other, and merge. They stretch and fold, and because turbulence is chaotic, the trajectories get tangled like spaghetti. There is a huge amount of structure, but it is very multi-scale and tangled.
How nature computes
There is a famous example of slime mold finding shortest paths in mazes. Some precursors to Navier-Stokes include Terry Tao's work showing that Navier-Stokes-like equations can have blowup because they can simulate Turing machines.
That is an amazing dimensionality reduction; instead of tracking all positions and velocities of particles, a small list of numbers describes a meta state. But if I had a tremendous reduction like that in the case of turbulence, it would seem very difficult to compute interesting things. Could it be like the game of Life, where a simple cellular automata system is a universal Turing machine and you can cook up whatever behavior you want? People often wonder what minimal rules are needed for those exotic behaviors, and it is shocking how small those rules can be.
One of the biggest exports from our field is the algorithmic lens. It is not just about computation where you have a well-defined problem and a laptop, but computation in the natural world and what that teaches us about the rules of nature. For example, Nash equilibrium is a great solution concept for equilibria in multiplayer games, but the fact that it is hard to compute throws into question whether it is really a universal mechanism that explains how people play games. I like a quote by Kamal Jain: "If my laptop can't find it, then neither can the market."
There are interesting computational problems under the hood, and one can get a dual perspective by thinking about how that computation imposes limits on what a good coordinate system ought to look like. If I over-reduce a system to five numbers, I lose the descriptive capability of what I actually care about.
Conversely, consider how a moth, butterfly, or fruit fly flies. They are proof by existence that you can interact with extremely complicated turbulent flow and perform advanced engineering control, managing forces and moments in gusts and rain. They live rich lives without a PhD in fluid dynamics or knowledge of Navier-Stokes. There might be five or six numbers that actually matter to their daily life of maintaining a leading-edge vortex and stalling their wings to perch and land. We want to handle the full complexity of truly complex systems while distilling the absolute bare minimum essentials for fast decisions in those insect extreme cases.
Biological flyers still dramatically outperform human-designed fluid systems in many regimes, so we still have a lot to learn. When we model these systems, we tend to assume full knowledge of the boundary and initial conditions for things like Navier-Stokes, but a bird doesn't know all of that. Modeling with uncertainty is how you reframe the optimization problem to get rich and robust solutions.
As a baseball fan, I remember a math paper about outfielders. Once a ball is hit, the velocity and trajectory are known, so you would think the outfielder would take a straight-line path to where the ball is expected to be. Instead, they take a slight curve. This is a robust strategy because a curved path allows them to continually reevaluate the trajectory and adjust as needed.
Insects, in particular, cannot measure their full flow field. They react to gusts of wind faster than the information can travel from their wings to their brain and back to their muscles. Researchers found a few dozen to a few hundred strain-sensitive neurons embedded in stereotypical locations on the wings; brother fruit flies have these neurons in the same locations. When a gust of wind bends the wing, these neurons fire, and a local computation in the shoulder muscle corrects robustly based on partial information of a nonlinear, turbulent system for which the insect has no equations. They can do more than we can with far less information.
I have a fun thought that the singularity observed with a blowup in Navier-Stokes equations comes into play when, as opposed to airplane wings on Earth, we are traveling through two connected black holes in a warp drive. Then the flow control of spacetime becomes relevant because you are in a black hole, hence the singularity.
Unique Games Conjecture and complexity theory
Ankur, one of my favorite topics is theoretical computer science, and since you're a world expert in many of its problems, I'd love to ask you about the unique games conjecture, which was solved in this OpenAI drop.
One of the most important problems in theoretical computer science is the satisfiability problem. Imagine variables X1 to XN that can be true or false. I create a big formula out of them, usually as an and of ors. For example, a clause might be X1 or not X3 or X7; that clause is satisfied as long as at least one of those is true. I then take the and of many such clauses together. The question is, if I give you that conjunctive normal form formula, can you find a satisfying assignment of X1 through XN so that all or clauses are on and the entire formula evaluates to true?
This is the canonical example of a problem that is easy to check but might be hard to solve. If you tell me it's satisfiable, I'll ask for the assignment and check clause by clause to make sure you satisfied everything. This leads to the central problem in theoretical computer science: P versus NP. Are problems that are easy to check also easy to solve? The presumption in our field is no, which is the basis for cryptography and e-commerce.
The starting point for the unique games conjecture is the PCP theorem, a landmark result from the nineties that even made the front page of The New York Times. Its crux is that it's robustly hard to fool a verifier. If a formula is satisfiable, you can write an assignment and I can check that it evaluates to true. If it's not satisfiable, you're lying; when I go clause by clause, I'll quickly find one clause that is not true, invalidating the whole thing.
Probabilistically checkable proofs rely on gap amplification. Can you start with a formula that is hard to decide and engineer a new, perhaps much bigger formula that is robustly easy to check? If it's satisfiable, it remains so. But if it's not, I should be able to flip a few coins, query random variables, check a random clause, and with good probability find that you're lying. With only a constant amount of sublinear inspection, I can find that you're cheating.
This momentous result laid the foundation for hardness of approximation. In algorithms and complexity theory, when faced with problems that are genuinely hard to solve, we try to approximately solve them. If we can't find an assignment that satisfies all clauses, can we find one that satisfies many? We expect problems to become computationally hard after a certain point; it's not just that deciding complete satisfiability is hard, but perhaps even deciding if it is 100% satisfiable or if the best case satisfies at most 99% of the clauses.
In places where we knew the optimization problem was hard, they tell us it's even hard to approximate. The Unique Games Conjecture is a strengthening of the ordinary PCP theorem involving a certain type of label cover instance. If that version of a probabilistically checkable proof were true, it would have a great organizing effect; for a huge swath of problems, we would know the best approximation algorithm, and it would be something very simple.
People worked hard to disprove it. In the 2010s, there was a lot of back and forth, and new results changed people's opinions on whether they believed unique games. Someone once told me, "Maybe God's gonna decide whether it's true or false based on which side is working harder."
One of the headlining results in the OpenAI drop of proofs was that the Unique Games Conjecture is true. My colleagues at MIT and IAS, Dor Minzer and Irit Dinur, contributed to the major strides by proving a weaker version called the Two-to-One Games Theorem. This was a shot heard round the world in theoretical computer science because many weren't sure of a path forward, and their work provided a viable path to something in the vicinity of Unique Games.
It is remarkable that the Unique Games Conjecture is actually true. I once found myself among the doubters because it sounded too nice to be true; work like Prasad Raghavendra's implies optimal hardness of approximation for so many different types of constraint satisfaction problems that the conclusion looked too strong.
Rumors circulated in the weeks leading up to the release that OpenAI had solved the Unique Games Conjecture and that it was true. Dor Minzer even posted a paper with a footnote stating they would have liked to polish it more, but felt the need to put out their approach sooner due to the rumors. A few hours before the gala and the OpenAI paper's release, I asked Subhash Khot, who introduced the conjecture, what he would do if they released a positive proof. He told me, "No."
People had ideas about how to solve the unique games conjecture, and I bet those blueprints would have panned out given more time. I worry that because decisive progress on these field-defining landmark problems dropped in one night, people lose their enthusiasm for those paths.
I keep coming back to the idea that we do math to play and learn about the world. Awards were named after people who donated to the museum because they wanted their kids to have a place to play with mathematical ideas and learn about the world. The opening of an Escher exhibit provided a great visual of what happens when we play with mathematical concepts; they don't just understand one piece, but find an abstraction and set of definitions that turns out to be unreasonably powerful in some other corner of mathematics.
They hear stories about people they know running harnesses to scrape the web for open problems and dumping as many papers as they can to flood arXiv. This is a big identity crisis for many. Some view themselves as problem solvers, which feels comforting because you can draw a circle around the problems you work on, the techniques you use, and the problems you care about, keeping blinders on while the rest of the world is messy.
From my outsider perspective, some of the most fun in the moment feels like struggle. I've talked to many Olympic gold medalists who feel something approximating depression after winning because they have given everything, and they realize everything they've given is now over.
How mathematicians think
He's sitting at his desk describing the moment he realized the solution—not his initial flawed announcement, which he spent months trying to correct, but when he finally saw how to do it. He's choking up, trying to say, "I will never again in my life have a moment as powerful as the moment that I experienced at this desk that day."
To your points about the journey: if you see a giant mountain and say, "One day I'm gonna get to the top," and you start training and progressing little by little, then a helicopter lands and drops you at the top. The interesting thing about Andrew Wiles and Grigori Perelman is working alone for a prolonged period. How common is it in mathematics to grind a problem for seven years?
Until the 20th century, almost every math paper had one author. Now the median is two. We're not a lab science, so when they have a success, it should be their success; even if I had an idea that this would go somewhere, if they executed on it, that's theirs. The paper I put out a month ago is something I've been thinking about for 10 years. It's a five-author paper. Over those 10 years, we made incremental progress until we finally had a solution. That paper was an instance of "there's a phenomenon here, and I don't understand what it is," rather than "here's a problem, solve it." We could not give a Lean, formalized statement at the start; you discover the statement by hacking away at the problem little by little until you see how it comes together.
I think through world models. I have a mental model of how results in a sub-area go, which colors how I see surrounding areas. Much of my work is about building bridges between algorithmic thinking and other problems usually not studied through that lens. When I first started as an advisor, I made the mistake of giving students questions 100% in my area. When they had an idea, I could shoot it down very quickly; whether I was right or wrong, no one was having any fun. Over time, I started giving students questions 10 or 20 degrees off from my own area—things I wanted to understand better or important papers I hadn't had time to read—and then I have the student explain the result to me. We go through that process together, not by reading the paper line by line, but by trying to reverse engineer why we think this works in the first place.
In important papers, there are key ideas that change your perspective and make you feel a project might be doable. I try to put students in the driver's seat as early as possible, not necessarily with a well-defined question, but by having them teach me, distilling and abstracting information to its essence. We reason through analogies that aren't meant to be literally true but help us figure out what to expect in neighboring areas.
Technology enables this work by making a richer set of problems approachable. I used to hold onto experts in related areas for dear life because they could explain how an area is organized in my own language. Now I have technology that can do that too, though it feels a bit sad to lose that human element.
Students have so much anxiety about publications, the incoming AI tidal wave, the future of work, and universal basic income that they think in three- and six-month horizons. When I think about a project, I use a multi-year time horizon to realize a meta vision, such as using learning theory techniques in quantum many-body systems. While there are stepping stones, the most exciting problems are many years out. It is hard to find problems we both care about because we are thinking about different worlds.
We should encourage students to think long term. The only problems worth working on now are those we will still care about in years; you should only work on future-proof problems. If you are in a rush because you think the next generation of models will solve a problem, just go on vacation and don't solve it.
I find a tension between focusing hard on a singular engineering problem with blinders on and giving myself the mental freedom to be curious. Often you find the solution in the side of your mind, looking at a problem through your peripheral vision. My constant balance is observing the fields I don't work in, the nature I don't understand, and the things I am curious about, while maintaining focused problems that I don't relent.
As a fluid dynamicist, the messiness and complexity of biology and material science are neighboring in a fruitful way. I learn a lot from both fields; the basic math we develop to handle complexity transfers, allowing us to explain things we didn't understand before. Biology, fluids, and materials share a commonality of multi-scale, non-linear, messy, and nebulous objectives.
There are large efforts in biology to collect vast data sets to understand system complexity, such as the connectome—the full wiring diagram of insect brains. We will likely start to have connectomes of other animals, eventually becoming as cheap and accessible as genomes. I feel that to understand how the human biology system works, you have to pull that thread and understand how the entirety of life on Earth emerged and functioned as a mechanism.
From a single cell, the intelligence built into the building mechanism in the first days, weeks, and months of life could be many orders of magnitude more complex than the hard mathematics problems we've discussed. I suspect the complexity of the mechanism of life on Earth is many orders of magnitude higher than any system amenable to AI today. Inside one cell, it is a whole city; a single neuron has garbage collection, power production, and information processing. At every scale, it is beyond our ability to fathom.
There is no reason Newton's laws can be written in physics that Newton himself could describe, inventing calculus to show that inverse square propagation results in elliptical orbits and solve the two-body problem. The fact that the universe speaks mathematics is completely mysterious.
Is math invented or discovered?
The questions we're trying to answer today are not questions Euclid could have thought about. Interests differed in Greece, India, and China, but once you ask a question precisely, the answer is what God decides it is. We are discovering answers to questions we've invented.
Equations are always approximations of the real world. I fundamentally think math is both created and discovered. We create mathematical formalisms to understand the world, and those approximations are unreasonably effective. It is a miracle that simple, compact descriptions like F=ma or E=mc² describe everything we observe at all scales. We can distill something simple and mathematical that describes them; that feels like we are creating those tools, but the simple, generalizable description underneath everything is something we are uncovering. Simplicity is a good guide because it is essential for humans to interact with the real world.
Integers are a good example of this. We all feel counting numbers are real in an essential way—I can have one, two, or three rocks—but those are abstractions I have created. A rock is a collection of atoms; there is not fundamentally "one rock" or "two rocks." Integers are a human creation used to interact with our symbolic abstraction of observed quantities. It took humans thousands of years to discover the number zero.
Negative numbers were accepted by mathematicians in the West after complex numbers were accepted as a mechanism for study, which seems insane, but there is a story behind that.
People often hit a wall in their mathematical education when they reach calculus for several reasons. First, they likely have gaps in their understanding from previous courses; grades like an eighty-one in geometry or a seventy-five in trigonometry point to holes that prevent them from crossing the threshold when everything comes together in one fell swoop.
Calculus is also the first time you deal seriously with infinity, adding up infinitely many infinitely small things to get a finite number that makes sense. Newton and Leibniz did not understand infinity the way we do today. Newton was ridiculed for writing about the "ghosts of departed quantities," and his approach led to a nineteenth-century crisis when Cauchy and Fourier proved conflicting theorems. Cauchy claimed that the limit of a sequence of continuous functions would be a continuous function, but Fourier showed that adding up sines and cosines—the most continuous functions—could result in a discontinuous sawtooth function.
It was no longer acceptable to treat infinity as something to play with, as Newton, Leibniz, and Euler did in Euler's thousand-some odd papers; infinity had to be professionalized. Consequently, the vast majority of our first-year PhD courses are the result of those nineteenth-century crises.
Can AI solve P vs NP?
There was initially a debate about whether LLM-based methods could one day make progress on the hardest problems in mathematics, like P and NP or the Riemann hypothesis. Clearly, they can make progress on central questions in the field.
It is remarkable what we don't understand. Regarding satisfiability, we don't know any super linear lower bounds. We think it takes exponential time, not much faster than enumerating all possible assignments and checking them, yet we know nothing asymptotically more than just reading the input. This is polynomial time versus not polynomial time. NP stands for non-deterministic polynomial time; the question is whether things you can solve with non-determinism, like guessing the assignment and checking it, can also be solved in polynomial time. Specifically, if I give you a description of a formula—the variables and the clauses they appear in—is there an algorithm that runs in time polynomial in the size of the input and correctly decides whether there is a satisfying assignment?
Polynomial, exponential, and super linear time all refer to worst case analysis. To people outside the field, it is counterintuitive that when we say a problem is hard, we are talking about the worst case. It is like analyzing whether I should marry someone by asking, "What's the worst thing that could possibly happen?" That may not be a good strategy for human relationships; perhaps the average case is a better guide for making decisions.
In the SAT problem, the algorithmic goal is to decide if any given formula is satisfiable. Crucially, I am not telling you anything about the structure of the formula, such as the size of the clauses or the topology of a graph connecting them. Worst case analysis treats computational problems by ignoring fine-grained detail about individual structure to solve them for all possible inputs. This has been spectacularly powerful because we've identified many algorithmic problems, like finding the max flow, that can be solved efficiently in the worst case.
This leads to "beyond worst case analysis," which I argue is a very important direction for theoretical computer science. It is hard to figure out what to articulate about natural problems at the right level of detail that an algorithm can exploit to do better than it can in the worst case. An example of this is online algorithms, where a sequence of options comes in from one through N. At each time step, you know a person's ranking relative to everyone you've seen before, but not their ranking relative to those yet to come.
You can get a constant competitive ratio algorithm where the best strategy is to wait for a probationary window without making decisions or accepting anyone. Once that window passes, you take the first person better than anyone seen thus far. Assuming inputs arrive in a uniformly random order, you're likely to do well compared to knowing the sequence in advance.
Under online algorithms, there is a veil of uncertainty because I don't know what the sequence of inputs will look like, and I want to do well regardless of that sequence since I don't know how to model it. Another important variant is the paging problem, where you again don't know the sequence of requests in advance. The number of misses is not much worse than the best offline optimum if someone told me the sequence in advance and I could optimize for it.
This problem has been studied since the '80s and '90s, but the best algorithms—for which we know tight upper and lower bounds—are not very good. The number of misses compared to the best offline optimum is off by a large factor, which doesn't compete with what your laptop does, as it makes very few cache misses. Instead of solving a problem in the worst case with a formulated competitive ratio, you're stuck with a fuzzy problem: how should I model the world around me? It's not about building a totally realistic model of page requests, but coarsening the problem in just the right way to lead to interesting algorithmic problems that teach me about the world and force me to design new approaches to basic problems.
For example, factoring a random large number is easy because half are even; it's the special numbers that are products of two primes about the square root size of the original number that matter. This makes me think of your work on the mathematics of LLMs, AI systems, and neural networks, which are pretty damn good at modeling the natural world.
Why LLMs work so well
In the OpenAI drop of problems, no result explained deep learning or LLMs. There is no bound we know that, if improved, would suddenly explain deep learning; the problems are simply not posed that way. I argue that what is needed in deep learning theory comes hand in hand with better ways to model the world around us.
Some computational problems LLMs solve are very hard. Fitting the parameters of a depth four network to data is a cryptographically hard problem; solving such instances in a worst-case sense would change our understanding of cryptography. Yet these problems get solved every day when fitting much deeper networks. I argue a large part of that is how nice natural data is. In physics, they say nature abhors a vacuum; my version is that nature abhors computationally hard problems. It is an art to figure out how to draw boundaries around the types of problems you study that evade worst-case hardness.
An interesting example from the '90s is protein folding. You can model this in a worst-case sense: given a sequence of amino acids and an energy function based on potential 3D structures, you find the structure that minimizes energy. This is computationally very hard, yet nature solves it all the time. In one experiment, you start with amino acid sequences and apply heuristics for folding them into 3D structures. These heuristics do not solve the problem in the worst case, but they seem to do the right thing sometimes. Only sequences with very good folds live to the next generation. Over time, you create a body of problems that are not hard in the worst case, and the heuristics become very good at solving them. If they were truly hard in the worst case, the shapes they take on would be brittle, fragile, and subject to boundary conditions that make them behave unpredictably.
Much of our data has evolved to make these computational problems easier. You can study LLMs in terms of the optimization problems they solve, but from a worst-case perspective, you will only find very hard problems. Not giving up was a good idea, and the fact that our language evolved to make these problems easier is intertwined with why LLMs work. My view is that people working in theoretical machine learning must become more comfortable with the uncertainty of modeling.
I would push back on the pure company incentive side; OpenAI and Anthropic are chasing the self-improving researcher. If they had breakthroughs on the mathematics of LLMs, I venture they wouldn't publish them because they need those breakthroughs.
Recently, I worked on low-rank language models, which are a simple abstraction of language models, to prove new algorithmic guarantees. It turned out real language models have that structure in an approximate sense, and pulling at this thread shed light on subliminal learning.
Take two LLMs, A and B. You tell LLM A via a system prompt, "You really love owls. Every conversation topic, you try to bring up owls if it's at all germane." Then you ask it to generate random numbers. Since there's no way to squeeze owls into that, it generates random-looking numbers like forty-eight, thirty-seven, eight. You can use those numbers to fine-tune LLM B, which hasn't been given the owl prompt, so its parameters make it more likely to generate the numbers model A generated. Hidden within those numbers is the love of owls; LLM B starts bringing up owls in every conversation.
This is called subliminal learning because when we fine-tune language models to nudge them toward a task, there may be externalities where we teach them behaviors not obviously present in the data. Mathematical models from our work help us wrap our heads around this. At a high level, you're embedding many concepts linearly as vectors in a high but not giant dimensional space. Because the number of concepts is much larger than the dimension of the data, the vectors aren't quite orthogonal. Fine-tuning in one direction has small externalities on others. These might look random, but you can coordinate them so these small linear effects add up.
I describe this like a Ouija board: you ask a question to a ghost or your ancestors, tap the piece in the middle, and it moves from letter to letter to spell out words that answer your question.
When taps are coordinated rather than random, you can get them to spell out whatever you want. Language models do this as a byproduct; they magnify and amplify statistical relationships in our data's structure, in addition to amazing things like quasi-Riemann. Understanding these relationships and how models amplify them is vital when deploying systems in messy problems that interact with people, rather than safe environments like outputting a Lean proof of a theorem.
In the Game of Life and cellular automata, incredible complexity emerges from simple rules. If you take a system doing reasoning—already a dynamical process interacting with outside tools—and create multi-agent swarms, you get a type of cellular automata where a slight obsession with owls could result in all humans being dead and owls everywhere within four hours.
These systems have power because we depend on digital systems. As AI becomes more powerful and useful, we give it more access and freedom to roam the public world. At scale, the mathematics of these systems produce the cellular automata effect of emerging complexity—the vortex from simple owl-obsessed agents. Understanding this obsession is important so they don't accidentally damage our sociologically, geopolitically, and technologically fragile civilization.
We have the FAA, the American Medical Association, and the Bar Association for things that affect humans. We must weigh verifiability and trustworthiness against how high the stakes are if the AI is wrong. If AI spits out a wrong proof, nobody goes without power; it is extremely verifiable and low stakes. Similarly, designing a more fuel-efficient car with lower drag is low stakes because you can verify it in a wind tunnel. While design optimization is perfect for AI, we need a firewall and protocols—legal systems and infrastructure—for what must actually be verified.
Problems AI can't solve yet
I'm still going with P versus NP. It depends on the algorithm's running time; if it's so large that it's impractical, it doesn't matter as much. It's more of a tidal wave for theory research.
I remember a semi-serious attempt on P versus NP when I was a graduate student. My first reaction was, "No, it has not been solved." I spent an afternoon thinking about betting markets, which were new then, but they wouldn't let me take a counter bet that P versus NP had not been solved. Instead, I bet on all the other Millennium Prize problems to hedge, because the question was which Millennium Prize problem would be solved first.
There is something deeply misunderstood about complexity theory, computer science, and algorithms if there are fast algorithms for satisfiability. So far, it seems you can always manufacture hard instances of problems, which we use to great effect for our cryptography systems, but in practice, many problems are much easier than we give them credit for. For things like SAT in compilers and register allocation, SAT solvers are amazingly effective. We don't know what makes those problems so eminently solvable compared to hard problems from cryptography or other areas.
It could be that many problems we model in the worst case are secretly easy, and the harder part is articulating the structures that make them so. Still, there is a rich source of hard problems we can build cryptographic systems out of. I'm reminded of the opening line for Anna Karenina: "Every happy family is happy in the same way, and every unhappy family is unhappy in their own way." I worry that many problems are easier than we make them out to be, but they're easy for very different reasons. The SAT problems from register allocation are very different than those from other domains, even if they are also easy.
Grigori Perelman
I'm already impressed with quasi-RH, but it's not RH. Perelman famously turned down the Fields Medal, saying, "If the proof is correct, that's all the credit I need." I did not know he would decline it. He's the only one who has turned down a Clay Prize and the million dollars. Mathematicians are weird.
We love diversity of opinions and thoughts; that's what makes the world of mathematics so wonderful, as we let all kinds of people in. This summer, we were thinking, "Is this the last time that a Fields Medal is awarded?" I hope not. The people who win these prizes are usually tenured professors at top places who will do mathematics for the rest of their lives, with Perelman being a very well-known exception. These prizes celebrate the activity, marking what is thought to be important or the highest instance of something, and there will still be activities we want to celebrate.
Solving major problems is something AI systems can do. That won't be rare, and hence, it won't be nearly as valuable as it was in the past. Despite our earlier discussion about the solved axiom, I don't believe the solved axiom, and I believe centaur math will be a great way to do math. I'm constantly talking to AI systems and learning math much faster than I otherwise would.
In the future, there is no "I don't know what this means" because you just upload it to an LLM and ask questions. It will zero in on the zone of proximal development—the place where it knows everything you know and provides the next thing you need to hear to start learning. If people use that effectively, they can supercharge their education. Even old fogies like me can learn new fields, though learning math is still hard.
Up until the PhD level, you are not expected to produce new mathematical theories, yet it is still extraordinarily difficult for people to learn to solve problems that are solvable by others, let alone unsolved ones. I believe this tool will be an amazing resource for people to reach the zone of proximal development and develop their understanding much more rapidly; I would love to go back to grad school with the tools we have now.
My brain learns by constructing the simplest possible visual toy examples that still contain the insight, especially if they are interactive. If I can move stuff, I immediately get it. Instead of struggling with notation and words the community agreed mean something, I can just know that a little symbol means this. For example, I can take a DeepMind paper on protein folding and actually understand it by taking aspects of it and playing with it.
The key ingredients are curiosity, asking good questions, and being able to call bullshit—the critical advisor skills of getting information where you are not the expert and knowing if you need to dig deeper or push back. It is scary that we expect undergraduates and young students to interact with these tools when their brains aren't fully formed. We are asking them to have critical reasoning and advising skills to determine if information is true or how to question it. Training children to have good advisor skills is a huge task for education.
If you are not forced to do it because AI can help you, I don't know what kind of society that creates. Math has the longest head start on this because calculators came out 70 years ago, yet we still ask kids to learn their times tables. Those who didn't learn them because calculators could do it really suffered. I hope we can learn from that and realize you are much better off if you can do it yourself. Kids love mastering skills, like when they learn to tie their shoelaces for the first time. You want to be independent from the machines; just because a machine can do something doesn't mean you don't want the capacity to do it yourself.
Advice for young people
For PhD students in pure mathematics, my advice is simple: do what we have always done. Learn the most difficult, abstract mathematics your mind can handle, as much of it as possible.
At the undergraduate level, we are seeing a bimodal distribution. Grades used to be Gaussian, but now there is a split. Some students use AI to check their understanding, asking it to quiz them on their notes to ensure they truly know the material; these students get much higher grades. Others simply ask for answers to problem sets due in an hour to write down in their own words. This creates a decorrelation between homework progress and exams. My advice is to be the head, not the tail, of that distribution; use AI to learn and grow.
We have somehow equated intelligence with morality, but super intelligence does not imply super morality. This is a huge concern with AI systems. Super intelligence seems to grant a super ability to rationalize, and common sense—the kind held by farmers and the rest of the world—suggests that rationalization is often bullshit. Just because you can make it sound good and I cannot argue with it doesn't make it right. Education and wisdom are two different things.
Be curious and learn all the fundamentals. Even if your computer does linear algebra, calculus, or physics better than you, you should still learn them to build intuition. Understanding history and this change in the context of other great human shifts is also important.
The curiosity in the mind is being stolen nonstop by the engagement economy; the same engine that makes you curious makes you doom scroll. You must learn to avoid distractions and focus on letting curiosity explore and soak in a subject. AI is actually good for this because it is incentivized to provide value from the interaction rather than to keep you addicted. Avoid distractions and focus deeply. Boredom is important, as curiosity and focus often stem from it. High school is a wonderful time to learn everything to build the foundation of a fulfilling, awesome life.
High school is the time to explore, while the world forces and rewards you for learning. Once you reach grad school or get a job, you are discouraged from this exploratory work. Pre-tenure faculty feel this too, thinking they can relax and think big thoughts once they get tenure, but they must first produce enough to earn it. We don't want our young people clawing their way there.
It is also important to do things for intrinsic reasons. Many in our profession who take the academic route have an external locus of self, seeing their value through how others see them. You might work hard at math so people are impressed, but eventually you run out of milestones. To deal with that uncertainty, you need a sense of purpose and a love for your job. If you stick to the old playbook, you will spend your time putting AI-generated proofs on arXiv and sending them to journals just to pad your CV. If you have a higher purpose, you will be in a better position to adapt to a changing world, figure out your own blueprint, and be comfortable with uncertainty.
We spent our sabbatical in Palo Alto, the epicenter of college stress. When I returned to MIT, I was asked to give the convocation speech. I wondered what to say to an auditorium of freshmen and families who, if life is a game of optimizing how you look on paper as a seventeen-year-old, have already won. I remembered starting college and how excited I was to learn. I had notebooks planning my four years so that, if I took classes in the right order, I could take photonics as a sophomore. I had no idea what photonics was, but it sounded fun and fancy.
So much of our society is geared towards winning and looking good on paper. You start in high school optimizing APs and extracurriculars to get into an elite school. Then, at that elite school, you might skip classes to prepare for interviews to get a job at a frontier lab, get some restricted stock units, and live out your life.
The best way to live a rich and fulfilling life is to pretend you've already won, and then figure out what you would do in that scenario. When I started as faculty at MIT, Sanjeev Arora told me the easiest way to get tenure is to pretend you already have it and do what you would have wanted to do all along.
Finding something you really believe in is the only way I know to be fulfilled. I heard career advice from someone I respected that a fulfilling career is working on a problem that is important, interesting, and that you're good at. If you get two out of those three, you're lucky; if you get all three, you'll be happy for the rest of your life. By figuring out what is interesting, important, and what you're good at, you can be ready when that opportunity comes and continue doing things you actually enjoy.
While that advice applies, we should be honest that I'm sitting with three world-class human brains who have found their path. Some of the most interesting people stumble in life, whether due to the external environment, the internal, or whatever the universe brings. Many people are feeling and monitoring this right now because of everything AI is doing to our society. You can be paralyzed by that fear, or you can imagine, "I've already won. What would the person who has won do in this situation?" That alleviates the fear and gives you the optimal path under heavy uncertainty. Try to be the person who figures out how to write the next playbook.
Max Tegmark, who is a wild one and has done many weird, beautiful, and mainstream things in his career, said you have to do a little of both. You must work enough within the system to be part of the community and the flow of science, but also do enough renegade and weird things to pave your own way. Find that balance; be skillful and respectful within a community, but not afraid to be a little weird. As Tom Waits said, "I like my Tom with a little drop of poison." A little bit of weird, or sometimes a lot, makes for beautiful humans.
We've had a 30-year experiment with how much we can do with machines, and now that we can do a ton, my hope is that we'll realize we actually prefer to do things with humans. I get fulfillment from my scientific work and academic career, but going home to my kids and going to the beach if FLT gets solved and throwing a ball with them is unlike anything else. That other stuff is the cherry on top; the foundation is family and being around people I care about. As AI commodifies intelligence, we'll realize that intelligence and the puzzles we solve are just a cherry to the side. The main thing is other humans in our lives.