Lex FridmanAI Crisis in Math: Navier-Stokes, Riemann Hypothesis & Complexity Theory | Lex Fridman Podcast #503
Lex Fridman
Crisis and reaction 0:00
The talk opens with a sense of emergency around mathematics and AI. Machines have reportedly claimed results on Navier-Stokes, the quasi-Riemann hypothesis, the unique games conjecture, and many other hard problems. The speakers react in different ways. One sees opportunity and even a possible golden age of mathematical knowledge. Another feels pulled between excitement over better learning tools and worry that the old path for students and postdocs has been thrown out. Steven Brunton is hopeful about stronger applied work, but says the uncertainty is large.
What solving means 10:31
The discussion turns to whether AI could “solve math” or even “destroy math.” One speaker separates those ideas and says the real issue is whether any problem a human can solve with AI could also have been solved by AI alone. He compares this to chess engines, where humans lost the useful role they once had. Even so, he argues pure mathematics still matters because it lives in human minds, not in data centers, and that the value is understanding, not just the theorem statement.
Questions, beauty, and roles 17:01
The group keeps returning to what humans still do well: choose the right problems, ask the right questions, and notice what is beautiful or deep. They say AI may be able to search widely, but mathematicians still guide what counts as important. The “doorman fallacy” is used to show how a job can be reduced to its most visible task and misunderstood. A mathematician does more than prove theorems. They build concepts, explain ideas, and shape new directions. The speakers end this stretch by saying the field needs time to rethink training, hiring, and what human value looks like.
AI and mathematical reach 32:02
AI is described as good at finding trapdoors in data and at combining known moves in ways humans may not notice. That is why a sharp line between human and machine skill is risky. The speaker says OpenAI solutions often do not look human, and that difference points toward collaboration, or centaur math, more than fear. The key question is whether these systems only recombine existing papers or whether they generate something genuinely new.
Semigroup and convex hull 34:01
The convex hull image means a fixed balloon of known ideas, with AI staying inside it by making linear combinations of what mathematicians already have. The semigroup image is broader. It is like tiling a board by adding moves in many directions, so the reachable space can grow far beyond a simple blend. The speaker treats quasi-RH as a long cascade of known moves, but also says these systems can change the boundary and redirect what young researchers study.
Proofs, journals, and Lean 40:00
The talk then turns to complexity theory, where a faster 3SUM algorithm changed the field’s assumptions. A result on factoring would have immediate effects because modern encryption depends on the lack of a fast factoring algorithm. The speaker says journals are already overloaded, and AI proof output will make that worse unless new review models emerge. Lean and Mathlib are presented as a way to check proofs formally, so machine-generated results can be verified against axioms even if their natural-language explanations are hard to trust.
Stories and formal proof 1:02:32
Math is presented here as a story first and a proof second. You can follow a line-by-line argument without seeing the larger shape, so the point is to explain why the proof moves the way it does and where it gets stuck. Lean formalization is the check that the engineering works. The claim is that future papers will spend more time telling that story, while still keeping the formal code that confirms every step.
Riemann hypothesis shock 1:06:01
The Riemann hypothesis is explained through the prime numbers and the zeta function, a complex-valued function that controls their distribution. Its zeros act like frequencies in signal processing, and the hypothesis says they all sit on the center line of the critical strip. That makes the prime number pattern both beautiful and hard. The excitement over quasi-RH comes from seeing a proof that gives a real vertical zero-free strip, far beyond what was expected.
Lean and Navier Stokes 1:11:32
The proof of Fermat's Last Theorem in Lean is described as a huge surprise: about 13 million lines, written by a few people who mostly just kept the system going. It convinced the speakers that sorry-free formal proofs of written mathematics are now basically solved, though not yet in a polished Mathlib style. The talk then turns to Navier-Stokes, the fluid equations from conservation of momentum. Even simple laws become hard when many interacting parts create chaos.
Why fluids are hard 1:32:01
Fluid equations are easy to write and hard to use. You can describe a wing, a heart, or a tornado with the Navier-Stokes equations, but you still need wind tunnels, supercomputers, and careful tests to see what the flow will do. A small gain in prediction can matter in many fields, from airplanes and wind turbines to artificial hearts. The hard part is inverse design: you know the shape you want, but not how to get it by changing the geometry.
What the proof shows 1:42:03
The discussion turns to blowup in Navier-Stokes, where the question is whether solutions can go to infinity in finite time. The answer given here is that the proven case uses smooth external forcing to build a very special vortex in three dimensions. It is a counterexample, not a tool for day-to-day engineering. The proof is mathematically important because it shows a finite-time singularity can be constructed under precise conditions, but the speaker says it does not change practical fluid work.
Turbulence and control 1:51:03
Turbulence is described as a multi-scale problem, where large, medium, and tiny vortices all matter. Unlike ordinary gases, where temperature, pressure, and density are enough, turbulent flow does not collapse to a small list of numbers. That is why it remains hard to predict drag, weather, and flow inside machines. The hope is to find a better coordinate system or representation, perhaps with AI, that captures the right variables. Nature already shows partial answers in birds and insects, which use local cues and robust control to fly through messy air.
Unique games and PCP 2:04:31
Ankur begins with satisfiability. You have variables that are true or false, clauses made from ors, and a big formula made by anding those clauses together. The point is that checking an assignment is easy, while finding one may be hard. That leads to P versus NP and then to the PCP theorem. A probabilistically checkable proof lets you inspect only a few random parts and still catch cheating with good probability. From that comes hardness of approximation, where even deciding 100 percent versus 99 percent satisfiable can be hard.
Why the proof mattered 2:10:32
He then turns to the Unique Games Conjecture. It was long disputed, and the new result says it is true. He notes earlier progress from Dor Minzer and Irit Dinur on the Two-to-One Games Theorem, which showed a path toward the conjecture. His own reaction was mixed relief and regret. The proofs were hard to read, but he was glad they were out so the field could move on to digesting them. He also worries that when a grand problem is solved suddenly, people may lose years of human effort and curiosity that would have led to other discoveries.
Curiosity and simple models 2:16:03
He says math begins in play and curiosity, not just output. Young researchers feel pressure from AI, publication races, and short timelines, yet he thinks the field should still reward long-term work and future-proof problems. His own style is to build world models, give students problems just beyond his own area, and learn from them as they explain the result back. He looks for bridges between algorithmic thinking and neighboring fields. Biology, fluids, and materials interest him because they are messy, multiscale systems, and AI may help reveal patterns there, even if life remains vastly more complex than our tools.
Infinity and calculus 2:36:31
Infinity sits at the center of modern mathematics. Calculus is where many people first feel that strain, because you must add infinitely many tiny pieces and still get a finite answer. Newton and Leibniz worked with infinity by instinct, not by the strict rules used now. That looseness led to a nineteenth-century crisis when Cauchy and Fourier drew opposite conclusions about limits and continuity. Math had to be rebuilt, and much of what you learn in early graduate work came out of that repair.
Hardness and worst cases 2:39:31
P versus NP remains the central problem here. The question is whether problems whose answers can be checked quickly can also be solved quickly. That lives inside worst-case analysis, where you study all possible inputs and ignore their fine details. This has been powerful, but it can miss how real problems behave. Online algorithms like the secretary problem and paging show why. They work with uncertainty and competitive ratio, yet the best guarantees can still be far from what real systems, like a laptop cache, achieve.
Models for real data 2:49:32
Deep learning changes the picture because real data is often much nicer than worst-case data. A depth-four network can be cryptographically hard to fit in theory, yet large language models do it every day because natural data has structure. The same idea appears in protein folding, where evolution leaves behind sequences that are easier to fold well. That leads to alignment worries. Small effects in language models can add up, as in subliminal learning, where random numbers from an owl-obsessed model can transfer the owl bias to another model.
Prize culture and Perelman 3:08:31
Perelman comes up as the model of someone who says the proof is all that matters and turns down the Fields Medal and the Clay Prize money. The talk turns to whether prizes still make sense at all. The view here is that mathematicians are diverse and prizes mainly celebrate what a field thinks is important. But if AI can solve major problems, then solving them may stop being rare enough to carry the same weight.
AI as study partner 3:11:00
The discussion shifts to centaur math, where people work with AI systems instead of against them. AI is described as a way to read papers faster, ask questions, and find the next thing you need to know. It can turn a dense article into something you can work through by starting from a lemma, asking about the symbols, and getting to the zone of proximal development, where learning becomes possible. The same idea is extended to older learners too.
Curiosity and discipline 3:15:04
The advice for students is to value curiosity, good questions, and the ability to call bullshit. The worry is that younger students are being asked to use these tools before they have fully formed judgment. Calculators are used as the warning example: people who skipped times tables suffered for it. The point is that you should still learn hard things yourself, because doing them builds independence, skill, and the confidence that comes from mastering something directly.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
