What this covers
Lex Fridman talks with a number theorist, a theoretical computer scientist and a fluid-dynamics researcher, each reading a recent AI-derived result from his own field. Alex Kontorovich takes the zeta function: an AI proof that no zeros lie in the strip with real part between seven-eighths and one, a boundary that stood for about seventy years. He explains the Riemann hypothesis as an orchestra with every instrument at equal volume. He also argues that stochastic natural-language proofs need autoformalization in Lean on top of Mathlib, and he cites the disappearance of centaur chess as a warning.
Ankur Moitra supplies the computer-science side. He sets worst-case analysis against modeling real input structure, using cache paging as the example, where laptops do far better than theory predicts. He describes how a 3SUM result disturbed the field's picture of hardness. He also traces the Unique Games result to human groundwork by Dor Minzer and Irit Dinur. Steven Brunton walks through the forced three-dimensional vortex blowup for Navier-Stokes Equations, the quadratic nonlinearity behind it, and why Turbulence resists prediction. He closes with insects correcting flight faster than signals reach the brain.
The conversation turns contentious in a few places. Moitra's claim that "nature abhors computationally hard problems" is disputed, as is his account of subliminal learning, where a model fine-tuned on random numbers picks up an owl preference. Kontorovich's vision of a fully formalized, crowdsourced mathematics is also contested. Along the way, Perelman's years of community checking and the nineteenth-century rebuilding of infinity serve as historical comparisons.
The guests argue that AI's rapid solution of open problems expands what is reachable but does not make mathematics obsolete, because the field's value lies in understanding, narrative, and well-chosen questions, and the real strain shifts to verification, communication, and institutions.
- Formal verification (Lean/Mathlib) makes AI-produced proofs checkable, but checking proofs does not supply the understanding mathematicians seek
- AI-solved results changed which problems matter and how fields organize, even when the solutions were not conceptually new
- Engineering value depends on predicting real fluid systems, which the Navier-Stokes proof does not directly deliver
Human guidance over which questions matter remains central as AI produces mathematical breakthroughs.
- Brunton argues that mathematics is both created and discovered: humans create formalisms to understand the world, but those approximations are unreasonably effective, suggesting that simple generalizable descriptions underlie observations and are being uncovered.
“I do somehow fundamentally think that math is both created and discovered.”
- Brunton says research programs already systematically discover minimal physical laws from data, recapitulating known laws such as F equals ma and Navier-Stokes from observation, and he assumes similar discovery of minimal mathematical structures is plausible.
“there are absolutely research programs that are systematically, automatically discovering kind of minimal physical laws.”
- Brunton argues that good problem solving is measured by asking the right sub-questions, which are interesting, solvable, and interpretable, and that this skill has always been valued in colleagues and will become more valuable as AI solves problems.
“that is how I think you measure good problem-solving is asking the right sub-questions”
- Brunton argues that the gold standard in mathematics is finding a structure that is simple and generalizes, and that humans remain the ones asking questions, so even AI-produced breakthroughs depend on human guidance about which questions matter.
“one of the gold standards has always been how simple can your result be and how much does it generalize.”
- Brunton's version of the solved axiom is that anything a human mathematician contributes could also be contributed by an AI, so value must come from something beyond asking a great question or proving a great theorem, which still requires human expertise.
“even in that world, it should be something other than I asked a great question, I proved a great theorem”
- Brunton argues the AI results are a lift of fog over the existing convex hull, making machines interpolate and reveal the opaque interior of human knowledge, but he also thinks they are more extrapolatory than merely derogatory recombination.
“I do think it is more extrapolatory than that, though.”
Turbulence cannot be predicted from governing equations alone, requiring simulation, data, and testing.
- Brunton argues that the Navier-Stokes proof does not change day-to-day engineering: fluids remain as hard to simulate and predict as before, so the practical bottleneck is a trillion-dollar problem of fast prediction and inverse design, which is much harder than the million-dollar math problem.
“we solved the million-dollar math problem. There's a trillion-dollar engineering problem”
- Brunton describes how insects respond to gusts faster than neural signals travel to the brain and back, using strain-sensitive neurons on wings in stereotyped locations that fire during specific bending patterns, so local computation corrects flight from partial information without equations.
“there are strain-sensitive neurons embedded on the wings, maybe a few dozen to a few hundred of these in very stereotypical locations.”
- Brunton argues that a useful representation of turbulent fluids would be a simple coordinate system, analogous to temperature, pressure, and density for gases; finding it may combine brute-force search over data with deep mathematics, since physics breakthroughs such as Newton's laws were preceded by coordinate transformations.
“all of our great physics breakthroughs were predated by a coordinate transformation”
- Brunton argues that for a gas, a handful of macroscopic numbers such as temperature, pressure, and density suffice, but turbulent flows like wing flows, arterial flows, or river flows require every scale of vortices and every simulation variable to predict behavior, which makes turbulence fundamentally challenging.
“That simple description doesn't exist for turbulent fluids”
- Brunton, citing Feynman, argues that knowing a general equation like Navier-Stokes, Maxwell's, or Schrödinger's does not tell you which phenomena it can produce, so predicting sunsets or drag from the equation alone is far beyond current capability and requires simulation and testing.
“we currently don't know from the equation what phenomena is possible.”
Navier-Stokes blowup arises from nonlinear vortex stretching, not from black-hole-like physics.
- Brunton explains that the proof constructs a smooth external forcing that produces a specialized vortex in three dimensions, which stretches and thins like a figure skater pulling in arms, so its radius shrinks and vorticity becomes infinite while energy stays finite; the result applies only to the forced case in three dimensions.
“This can only happen in three dimensions. This is really important”
- Brunton says three-dimensional turbulence cascades energy from big vortices to ever smaller ones until heat diffuses, while two-dimensional weather flows show a reverse cascade where small vortices merge into larger ones, which is why hurricanes form.
“In two dimensions, like weather on Earth, you actually get a reverse cascade, which is why hurricanes form.”
- Brunton explains that the equation x-dot equals x-squared blows up in finite time, while x-dot equals x only grows exponentially; the convective term in Navier-Stokes is a quadratic nonlinearity of this kind, so without viscosity the fluid could potentially blow up, while viscosity may damp it.
“that quadratic nonlinearity, that quadratic kind of feedback is enough to cause this equation to blow up.”
- Brunton argues that blowup in Navier-Stokes is not analogous to black holes predicted by general relativity, because the fluid equations are only valid at macroscopic scales where averaging over many molecules applies; once you zoom to individual molecules, other physics takes over.
“these equations are an exceptionally good model of how fluids work at a certain scale”
- Brunton explains that Navier-Stokes is conservation of momentum for freely moving fluids, written for arbitrary small volumes and taken to the limit, yielding a partial differential equation for how the velocity field and pressure change in space and time.
“the Navier-Stokes equations is just a way of stating the conservation of momentum for fluids”
AI's rapid progress makes future-proof problem choice and careful, verifiable deployment essential.
- Brunton argues that the only problems worth working on now are those that will still matter in years, so researchers should avoid racing to solve problems that the next model generation will soon solve, and instead choose future-proof problems.
“you should only be working on future-proof problems.”
- Brunton proposes a matrix for deploying AI based on verifiability and stakes of error: AI is ideal for low-stakes, easily verified design optimization, such as car aerodynamic design checked in wind tunnels, while high-stakes unverified uses need a firewall with protocols and legal oversight.
“We need a firewall of what needs to actually be verified”
- Brunton argues that if another country had discovered teleportation or time travel, the community would reinvent it within about five years, so knowing a solution exists is strong motivation to reach it faster, even though the path itself also matters.
“knowing there's a solution also is an incredible motivation to get there much faster.”
- Brunton argues that intelligence is not morality: a superintelligent system has no implication of being moral, and its ability to rationalize can make bad conclusions persuasive, so the concern is that rationalization substitutes for common sense.
“super intelligence does not imply super morality”
Mathematics must shift from the lone genius model toward transparent collective organization.
- Mathlib's pull-request model lets maintainers and reviewers see exactly who contributed which simplification or reuse, making mathematical contribution transparent in a way journals do not, and formal verification enables this for the first time.
“instead of a paper, you have a PR, you have a pull request... you can see exactly who contributed what idea”
- Brunton argues that as in engineering fields with billion-dollar collaborative projects, mathematics must move from the lone genius model toward collective organization, and leadership becomes what a community needs during this organizational crisis.
“I’m thinking leadership is obviously the thing that a community needs when it’s having this kind of crisis of organization and individuality”
- Brunton notes that mathematicians tend to be independent, dislike being told what to do or study, and prefer solitude, which makes the shift toward collective organization a cultural challenge.
“this is a very independent group of people who don’t like being told what to do”
Durable fulfillment comes from intrinsic values, human connection, and deliberate unstructured attention.
- Brunton relays advice that a fulfilling career combines work that is important, interesting, and that you are good at, where two of the three is lucky and all three yields lasting happiness, so preparation should prepare one for the opportunity to arise.
“a fulfilling career is something where you can work on a problem that is important, that you find interesting, and that you’re good at”
- Brunton argues that many solutions to focused engineering problems come from peripheral attention to fields one does not work in, because a mental sidecar of observation feeds ideas into the main problem, so researchers need both focus and curiosity.
“Very often you find the solution when you turn over a rock and when it’s not something you’re thinking about”
- Brunton argues academics often pursue work for external validation, which can carry early careers but runs out of milestones, so a durable path requires intrinsic reasons and a sense of personal values.
“a lot of people who go the academic route, the reason they do things sometimes is really external locus of self”
- Brunton argues that as machines commodify intelligence, people will prefer doing things with other humans, such as in-person conversations and oral discussion, so family and community become the foundation of meaning rather than the intellectual puzzles.
“my hope is that we’re gonna realize we actually prefer to do things with humans”
- Brunton argues that boredom is important because curiosity and focus often arise from it, so people should deliberately allow themselves to be bored rather than filling every moment with stimulation.
“To actually be bored, be like curiosity and focus comes often from boredom. I would like to be bored more often.”