Lex Fridman Podcast
#503 – AI Crisis in Math: Navier-Stokes, Riemann Hypothesis & Complexity Theory
This debrief was written and voiced by AI from the episode's audio.
🎧 AI Debrief
🎙️ Full Episode
Key Moments
Alex Kontorovich reframes AI's solving of major open problems not as a crisis but as a chaos of opportunity, emphasizing the shift from working by candlelight to having a flashlight in the woods of mathematics.
Ankur Moitra articulates the core tension: AI tools magnify his ability to learn and teach, but the traditional pathway for graduate students to build careers through problem-solving is now thrown out the window, leaving him intellectually lonely.
Ankur introduces the 'doorman fallacy' to explain why AI solving theorems doesn't destroy mathematics. Mathematicians don't just prove theorems; they ask deep questions, build new theories, and make conceptual connections.
Alex and Tim Gowers explain how AI combining known moves (semigroup) is actually much more powerful than just staying within the convex hull of existing ideas. The metaphor clarifies what AI is doing and what we don't yet know.
Alex recalls the exact moment he saw the announcement that quasi-Riemann hypothesis had been proved. He was at a gala, holding whiskey, and couldn't process what his eyes were seeing. It was outside his model of what was possible.
Alex describes how AI formalized Fermat's Last Theorem in Lean in days, generating 13 million lines of code with minimal human guidance. It shattered his expectations about how long formalizing even a partial result would take.
Steven explains the Navier-Stokes blow-up proof. AI constructed a hyper-specialized vortex that stretches and thins itself while maintaining finite energy, creating infinite vorticity at a point. It's physically unintuitive and mathematically stunning.
Steven draws a sharp line. The AI proof of Navier-Stokes is mathematically landmark but practically useless. Real engineering problems—designing better aircraft wings, artificial hearts—remain unsolved and worth trillions.
Ankur recalls the weeks of rumors before the proof dropped, his meeting with Subhash Khot (who posed the conjecture) hours before release, and Khot's striking response: 'No, I won't read it.' The conjecture Khot dedicated his life to was solved by AI, and he had no interest in the proof.
When asked what makes a good mathematician, Ankur emphasizes curiosity over raw intelligence, intrinsic motivation over external metrics, and purposeful work over CV optimization. This becomes the through-line for all the advice given.
Ankur describes how language models embed concepts as vectors, and when you fine-tune on one direction, small externalities leak into other directions. You can coordinate these leaks to make AI say whatever you want, like a Ouija board. This matters for alignment and understanding how AI actually works.
Ankur gives the most concrete psychological advice: pretend you already have tenure, then figure out what you'd work on. This removes the anxiety spiral and points toward intrinsically motivated research. He credits Sanjeev Arora with giving him this advice when he started as faculty.
Search Episode
Summary
In this wide-ranging conversation, three leading mathematicians—Ankur Moitra (MIT), Steven Brunton (University of Washington), and Alex Kontorovich (Rutgers)—discuss a watershed moment in mathematics and AI. Over recent weeks, AI systems claimed to have solved longstanding mathematical problems including the Navier-Stokes equations, the quasi-Riemann hypothesis, and the unique games conjecture. The episode explores both the excitement and anxiety this breakthrough generates. While the mathematicians emphasize that AI is solving well-posed problems rather than generating fundamentally new mathematical ideas, they grapple with profound questions: What does it mean to be a mathematician when machines can now prove theorems faster than humans? Will human mathematical insight become obsolete, or will it evolve toward asking better questions and seeking deeper understanding? The discussion ranges from complexity theory and fluid dynamics to the philosophy of mathematical beauty, the role of human struggle in discovery, and practical advice for students facing an uncertain future. Across all three experts runs a common thread: mathematics is about more than solving problems—it's about understanding, connection, and the human quest for meaning.
Key Takeaways
AI has demonstrated the ability to solve major open problems, but solving a problem is not the same as understanding why it works or discovering new mathematical concepts. The value of mathematics lies in the journey and the insights gained, not just the destination.
highThe 'doorman fallacy'—reducing a complex role to its most obvious task—applies to mathematics. Proving theorems is the most visible part of a mathematician's job, but mathematicians also ask the right questions, build new theories, and make conceptual connections that AI has not yet demonstrated.
highHuman and AI mathematics will likely remain complementary (centaur math) rather than one replacing the other. AI excels at combining existing ideas across vast literature; humans excel at asking which ideas matter and why.
highThe mathematical community must reimagine institutions, hiring practices, and how we value contributions. Problem-solving alone may no longer be a reliable path to professional success, pushing mathematicians toward pursuing intrinsic motivations rather than external validation.
highStudents and early-career mathematicians should focus on learning fundamental mathematics deeply, asking intrinsically meaningful questions, and building diverse skills rather than optimizing for short-term problem-solving metrics or AI-generated paper output.
highThe role of struggle, curiosity, and boredom in human learning and discovery remains irreplaceable. The 'solved axiom'—the idea that anything a human can do, an AI could also do alone—is not yet true and may never fully be true, leaving room for human mathematical creativity.
mediumAI breakthroughs in mathematics will accelerate further progress in applied fields like fluid dynamics, but foundational engineering and design problems remain computationally hard in ways pure mathematics may not be, preserving important roles for human expertise.
mediumNotable Quotes
“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 one is aware of an existing solution.”
— Terry Tao (quoted by Alex)
Captures the emotional weight of AI's sudden breakthroughs for the mathematical community, particularly the loss of the struggle and discovery process that has historically driven mathematical research.
“My work has always been about asking the right questions. To that extent, these tools are amazing for me because they increase my ability to learn about other fields. But the part that I really love the most is the mentorship aspect, and the blueprint for success is now thrown out the window.”
— Ankur Moitra
Illustrates the torn feelings many mathematicians experience: excitement about AI as a tool paired with anxiety about mentoring the next generation and defining what 'success' means.
“I would say most of my feelings are pretty hopeful. I think I'm an engineer. I like solving problems that make technology more realistic for humans. Our ability to solve real applied problems is going to get much, much better. But there is a lot of uncertainty.”
— Steven Brunton
Represents an engineering perspective that AI progress is fundamentally positive for applied work, even as uncertainty about the future remains.
“The biggest, most overarching feeling is really feeling like you're getting pulled in two very different directions. On one side, these tools magnify my ability to understand the world. On the other side, I feel intellectually lonely because the playbook for training the next generation is gone.”
— Ankur Moitra
Encapsulates the core tension of the moment: intellectual possibility versus institutional and emotional disruption.
“The value of pure mathematics is that it lives in the heads of human beings. So just to add to this, I mean, I'm glad you came right out with the question that's on everyone's mind, so thank you for that.”
— Alex Kontorovich
Asserts that human understanding—not solutions in data centers—defines the purpose of mathematics, a key premise underlying the optimistic case for the future of the field.
“Pretend you already have tenure. What would you do? Because a lot of the people we're giving advice to have really so many different opportunities. Finding something that you really believe in is the only way I know to be fulfilled.”
— Ankur Moitra
Provides concrete, psychologically grounded advice for students and early-career researchers facing anxiety and uncertainty about their future.
Predictions
The Riemann hypothesis will not be solved by AI systems within the next 5-10 years without significant new mathematical insights beyond those evident in the quasi-RH proof.
— Alex Kontorovich · Confidence: medium · 5-10 years
P versus NP will remain unsolved and is unlikely to be resolved by AI within the coming decade; it is fundamentally harder than problems like quasi-RH because it requires new understanding of computational complexity boundaries.
— Ankur Moitra · Confidence: high · 10+ years
Applied mathematics and engineering problems (turbulence, inverse design, flow control) will remain computationally hard for AI despite progress on pure mathematics problems, because these problems lack the well-posed structure of formal theorems.
— Steven Brunton · Confidence: high · ongoing
The mathematical community will experience short-term disruption to hiring, tenure, and publication metrics, but will eventually adapt by valuing different contributions (mentorship, asking good questions, beauty, clarity) rather than problem-solving alone.
— Ankur Moitra · Confidence: medium · 5-10 years
AI will continue to make breakthroughs on formalized mathematics (especially in Lean proof verification and mathematical libraries), but breakthroughs on understanding deep physical principles (why Navier-Stokes works, the structure of turbulence) will lag.
— Steven Brunton · Confidence: medium · 5 years
The intrinsic motivation and curiosity of some mathematicians will be preserved and even enhanced by AI tools, while others will experience demoralizing effects; a bimodal distribution of outcomes will emerge.
— All three · Confidence: medium · ongoing
Action Items
PhD students and early-career researchers should focus on learning deep mathematical fundamentals and developing strong intuition rather than racing to publish papers or solve well-known problems.
AI can now solve well-posed problems efficiently, so competitive advantage lies in asking better questions, understanding connections across fields, and thinking deeply about what matters.
Graduate advisors should explicitly teach students how to find intrinsically meaningful research questions and help them navigate uncertainty, rather than following the old playbook of 'solve a hard problem and you'll succeed.'
The institutional landscape for mathematics is changing rapidly; students need psychological and intellectual resilience in addition to technical skills.
Mathematicians and institutions should experiment with new models of contribution beyond the single-authored paper, including collaborative Lean formalizations, open-source mathematical libraries, and teaching resources.
Mathlib and similar projects demonstrate that curated mathematical knowledge repositories can be more valuable than individual papers; the community should invest in these new forms of work.
Undergraduate and graduate educators should be intentional about preventing students from using AI to avoid struggle, while teaching AI as a tool for verification, learning, and exploring adjacent fields.
Early evidence suggests a bimodal distribution where students using AI for checking understanding excel while those seeking shortcuts lag; educators need clear pedagogical guidelines.
Research community and funders should support work on understanding AI systems' mathematical reasoning (subliminal learning, Ouija board effects) and alignment in high-stakes applications, rather than assuming AI will solve all mathematical problems.
Theoretically hard problems like P vs NP and empirically hard problems like turbulence control remain, and their solutions likely depend on understanding the mathematics of AI itself.
Young mathematicians should maintain curiosity and intrinsic motivation by asking: 'If I already had tenure, what would I work on?' rather than optimizing for external metrics or fearing AI displacement.
Career satisfaction and meaningful contribution depend on alignment between personal values and work, not on beating AI at problem-solving.