DiscussionTechnical

AI Crisis in Math: Navier-Stokes, Riemann Hypothesis & Complexity Theory | Lex Fridman Podcast #503

Lex Fridman

Three leading mathematicians discuss the unprecedented impact of AI systems solving major unsolved problems in mathematics, including the Riemann hypothesis (quasi-RH), Navier-Stokes existence, and the Unique Games Conjecture. The conversation explores both the exciting possibilities and existential concerns for the mathematics community, examining what it means for human mathematical practice when machines can solve problems faster than humans can formulate them.

Summary

In this emergency episode of the Lex Fridman Podcast, three world-class mathematicians—Alex Kontorovich (number theorist at Rutgers), Ankur Moitra (theoretical computer scientist at MIT), and Steven Brunton (fluid dynamics expert at University of Washington)—gather to process the shock of AI systems claiming to have solved some of the hardest problems in mathematics. Over recent weeks, OpenAI and Anthropic released proofs of the quasi-Riemann hypothesis, improvements on Navier-Stokes, the Unique Games Conjecture, and hundreds of other mathematical results.

The panelists discuss their initial emotional reactions, ranging from awe and terror to cautious optimism. Alex describes seeing the quasi-RH proof drop at a mathematics gala and being unable to compute what he was seeing. Ankur discusses the pull he feels between excitement at new tools and worry about losing the struggle that makes mathematical work meaningful. Steven expresses hope about engineering applications while acknowledging uncertainty about whether humans will be dwarfed by AI intelligence or whether AI becomes merely a powerful tool.

A central theme is distinguishing between different types of progress. The speakers discuss the 'solved axiom'—whether anything a human mathematician can do, an AI system could also do independently—and debate whether recent AI results represent genuinely new mathematical ideas or clever combinations of existing knowledge. Alex argues the solutions lie in the 'semigroup' of human knowledge rather than the 'convex hull,' meaning AI can reach much further by combining known techniques, but questions whether this represents true mathematical innovation. The conversation explores how problems that seemed unreachable have fallen, and how this changes the landscape of what future problems seem feasible.

The panelists examine specific breakthroughs: the quasi-Riemann hypothesis and why primes matter; Navier-Stokes and the billion-dollar engineering challenge of understanding fluid dynamics; and the Unique Games Conjecture in complexity theory. For each, they explain both the mathematical significance and what the AI proofs revealed about problem structure.

A major concern is the impact on graduate students and early-career mathematicians. Terence Tao's worry about destroying 'person-years of ongoing patient research' is discussed at length. The group worries about whether PhDs will lose motivation to pursue hard problems when solutions might appear overnight, and whether the field will fragment into those using AI to generate papers versus those seeking deeper understanding.

The conversation addresses broader questions about what mathematics is for. They argue that mathematics isn't just about proving theorems but about understanding, building generalizable theories, asking important questions, and communicating ideas. They discuss the 'doorman fallacy'—reducing mathematics to its most visible task (proving theorems) and missing everything else mathematicians do. The group suggests mathematics in the age of AI should emphasize beauty, generalization, and the human experience of discovery.

The panelists discuss formalization in Lean and Mathlib, explaining how AI can now produce formally verified proofs (like the 13 million-line proof of Fermat's Last Theorem) but these still require human work to be made elegant and canonical. They debate whether mathematics could be done entirely formally or whether informal reasoning that builds intuition remains essential.

The conversation turns to theoretical computer science and the mathematics of why deep learning works. Ankur explains how problems that are hard in the worst case turn out to be easy in practice because natural data has structure that evades worst-case hardness. He discusses the 'Ouija board' effect where LLMs magnify small correlations in training data through coordinated small effects.

The group discusses P versus NP, reflecting on whether it might actually be true that P equals NP (with impractically slow algorithms), and whether understanding which problems are easy in practice matters more than worst-case analysis. They explore how AI systems might revolutionize understanding of optimization problems by finding shortcuts humans missed.

A significant portion addresses what mathematics means in a philosophical sense—whether it's invented or discovered. The speakers suggest a pragmatic answer: humans invent the questions we ask based on our current state of knowledge, but once precisely formulated, we discover the answers. They reflect on how the universe's 'unreasonable effectiveness' in being described by mathematics is mysterious.

The panelists offer advice to anxious students, PhD candidates, and early-career researchers. Key themes include: stay curious and learn fundamentals deeply; do things for intrinsic reasons rather than external validation; work on future-proof problems; learn from history and context; develop critical thinking skills; embrace boredom as a source of creativity; and imagine you've already won before deciding what to pursue. They emphasize that finding intersections of important problems, interesting questions, and things you're good at is rare but crucial.

The discussion closes with reflections on community and human connection. Despite the power of mathematics and machines, the panelists emphasize that family, relationships, and being around people you care about provide the true foundation of a meaningful life. They suggest that as AI commodifies intellectual achievement, humans may rediscover the value of direct human connection and collaboration.

Key Insights

  • Terence Tao argues that discovering solutions to major open problems via AI destroys the exploratory value of human effort on those problems, as 'one cannot explore a problem as profitably and richly once one is aware of an existing solution,' mourning 'the path not taken and opportunities lost.'
  • Alex distinguishes between AI solutions representing work in the 'convex hull' (linear combinations of existing ideas) versus the 'semigroup' (much further reach through combining known techniques), arguing most recent solutions are semigroup progress but questioning whether they represent genuinely new mathematical ideas.
  • Ankur explains how natural data has structure that makes computationally hard problems practically easy—'nature abhors computationally hard problems'—suggesting deep learning works not because of algorithmic breakthroughs but because evolved data evades worst-case hardness.
  • Steven notes that while Navier-Stokes blow-up was mathematically proven, it operates in a 'non-physical regime' and 'day-to-day engineering fluids and applied math analysis is almost completely unchanged,' distinguishing between solving theoretical math problems and solving trillion-dollar engineering problems.
  • Alex argues that the value of pure mathematics lies in having proofs 'live in the heads of human beings' through understanding and story, not in theorems existing 'in ones and zeros in data centers,' reframing what constitutes valuable mathematical progress.

Topics

AI solving major mathematical conjectures (Riemann Hypothesis, Navier-Stokes, Unique Games Conjecture)Impact on mathematics community and graduate student motivationPhilosophy of mathematics: invented vs discovered, nature of mathematical progressFormalization in Lean and automated theorem provingComplexity theory, P vs NP, and worst-case vs average-case analysisDeep learning theory and why neural networks work on natural dataHuman vs machine mathematical reasoning and the 'doorman fallacy'Fluid dynamics, turbulence, and real-world engineering challengesCareer advice and finding meaning in academic work during technological disruptionCentaur mathematics: human-AI collaboration

Transcript

[0:00] - All right, so this is a kind of emergency episode I think it's fair to say, about the state of mathematics and artificial intelligence. A kind of crisis moment in the history of mathematics. Over the last few weeks, machines claimed to have solved some of the hardest problems in the history of math. Navier-Stokes, the quasi- Riemann hypothesis, the unique games conjecture, and hundreds of other results. So here with me, we have Steven Brunton, [0:33] who's a mathematician, fluid dynamics expert at University of Washington, who can speak to Navier-Stokes, turbulence, all things fluid, and many other topics. Also, here's Ankur Moitra, a mathematician and theoretical computer scientist at MIT, visiting Institute for Advanced Studies in…

Full transcript available for MurmurCast members

Sign Up to Access

More from Lex Fridman

Get AI summaries like this delivered to your inbox daily

Get AI summaries delivered to your inbox

MurmurCast summarizes your YouTube channels, podcasts, and newsletters into one daily email digest.