TLDR
Three leading mathematicians discuss the seismic impact of AI systems solving major open problems (Navier-Stokes, quasi-Riemann Hypothesis, Unique Games Conjecture, Fermat's Last Theorem formalization) and debate whether mathematics as a human endeavor is ending or entering a golden age. They explore the 'solved axiom' — whether AI will eventually solve any problem humans+AI can solve — and argue that mathematics' value lies in human understanding, storytelling, and asking the right questions, not just theorem production. They emphasize the need to rethink education, publishing, and career incentives while preserving curiosity and human connection. The conversation spans number theory, fluid dynamics, complexity theory, and the philosophy of mathematical practice. The Fermat's Last Theorem formalization in Lean (13M lines by 2-3 people) and quasi-RH proof (vertical zero-free strip at 7/8) are highlighted as watershed moments. They conclude that AI commodifies intelligence but human meaning comes from relationships and intrinsic motivation. The doorman fallacy illustrates how reducing mathematicians to 'theorem provers' misses their true role as builders of conceptual frameworks. Formal verification (Lean/Mathlib) is now essential for trusting AI-generated proofs. P vs NP remains stubbornly hard. Advice for students: be curious, learn fundamentals, pursue intrinsic motivation, and pretend you already have tenure.
For you
PRO members can sign in and request more information — or their own version of this digest, tuned to what they care about.
Sign in to unlockSurprising Insights
The Fermat's Last Theorem formalization in Lean (13 million lines, sorry-free) was completed by 2-3 people with no domain expertise in number theory, mostly prompting 'keep going' — not by experts carefully curating each step.
The quasi-Riemann Hypothesis proof establishes a vertical zero-free strip (real part between 7/8 and 1) using methods that fundamentally cannot reach the critical line (real part 1/2), meaning a genuinely new idea is still required for full RH.
The Navier-Stokes blowup proof requires smooth external forcing and constructs an infinitely thinning vortex filament — a regime where the continuum equations break down physically (molecular effects dominate), so it does not imply real fluids blow up.
AI-generated proofs do not resemble human proofs: they chain known 'moves' in long cascades (semigroup action) that humans could not explore due to breadth/depth limits, and the proofs often lack the conceptual narrative humans value.
The Unique Games Conjecture was widely doubted (sounded 'too nice to be true') but turned out true; its proof organizes hardness of approximation for vast classes of problems, yet the AI proof is not yet readable for extracting new human-understandable techniques.
Autoformalization (translating natural-language proofs to Lean) is essentially solved for existing mathematics, but 'canonization' — refactoring jerry-rigged AI code into reusable, canonical Mathlib lemmas — remains a massive unsolved bottleneck.
Terence Tao and other Fields Medalists are 'freaking out' — a mix of awe, terror, and grief — because the release of AI solutions destroys the motivational landscape for graduate students who can no longer explore problems profitably once a solution exists.
The panelists' own predictions about AI capabilities were wildly wrong within months: Alex gave a colloquium stating AI could only build 1-2 layers on Mathlib; two days later, the 13M-line FLT formalization dropped, proving him wrong.
Key takeaways
AI systems have recently solved or made major progress on several of mathematics' hardest problems — Navier-Stokes existence/uniqueness with smooth forcing, the quasi-Riemann Hypothesis (zero-free strip at real part 7/8), the Unique Games Conjecture, and a 13-million-line Lean formalization of Fermat's Last Theorem — shocking experts who expected decades more work.
The 'solved axiom' posits that any problem solvable by humans interacting with AI could also be solved by AI alone; the panelists largely reject this, arguing that human mathematics is about building conceptual frameworks, asking the right questions, and developing understanding — not just producing theorem statements.
The 'doorman fallacy' warns against reducing a mathematician's role to its most visible task (proving theorems); mathematicians also build theories, make conceptual connections, mentor, and develop abstractions that generalize across fields — value that AI has not yet replicated.
Formal verification in Lean/Mathlib has become essential: AI-generated natural-language proofs cannot be trusted (stochastic processes), but Lean proofs checked against Mathlib provide certainty. Autoformalization of existing human proofs is now essentially solved, though making proofs 'Mathlib-quality' (canonical, reusable) remains a massive human effort.
The quasi-RH proof uses a semigroup-like chaining of known techniques (cubic theta functions, sextic character twists) rather than a fundamentally new idea; its methods provably cannot reach the full Riemann Hypothesis (real part 1/2), illustrating that AI may exhaust 'tiles' within the current conceptual boundary without expanding it.
Navier-Stokes blowup with smooth forcing is a mathematical tour de force but operates in a non-physical regime (infinitely thinning vortex filaments); the trillion-dollar engineering challenge — predicting turbulence for inverse design of wings, hearts, turbines — remains untouched and may require new coordinate systems discovered via AI-human collaboration.
P vs NP remains the panel's consensus hardest problem; worst-case complexity analysis may be the wrong lens for natural problems (which are often 'secretly easy' for diverse reasons), and understanding why deep learning works requires modeling the structure of natural data, not just worst-case bounds.
The crisis in mathematical careers is real: graduate students face a destroyed blueprint (solve a known hard problem → get recognition), journals are overwhelmed, and the incentive structure rewards flooding arXiv with AI papers. The panel urges recentering on intrinsic motivation, curiosity, and 'pretending you already have tenure' to find fulfilling problems.
Mathematics education is bifurcating: students using AI as a tutor (quizzing, explaining) excel, while those using it for homework answers fail. The panel stresses learning fundamentals (linear algebra, calculus) even if AI computes them, because mastery builds the intuition needed to ask good questions and call bullshit on AI outputs.
Human meaning in mathematics — and life — comes from relationships, curiosity, and the struggle itself. Andrew Wiles' tears at solving Fermat's Last Theorem illustrate that the journey, not the destination, is the point. AI commodifies intelligence but cannot replicate the human experience of understanding, mentorship, and shared discovery.
Notable quotes
“It's a chance to rethink what it is we're trying to accomplish, and it's the chaos of opportunity that some people are reacting negatively to, some people are reacting positively to, and sometimes those reactions are happening in the same person.”
▸ Notable quotes · 2:12“Now we have a flashlight.”
▸ Notable quotes · 2:44“The part that pulls me in the other direction is the recognition that, you know, the playbook and the blueprint for success that we had for so long is now thrown out the window.”
▸ Notable quotes · 4:52“I feel very excited, but maybe more intellectually lonely than I felt before, in part because, you know, a lot of times great results come from asking very humble questions, at least for me.”
▸ Notable quotes · 5:14“If you knew it was possible and you could reverse engineer a little bit, we would be curious, we would dive into it, and we would figure it out. So knowing there's a solution also is an incredible motivation to get there much faster.”
▸ Notable quotes · 10:01“I don't think the solution to theorems living in ones and zeros in data centers is valuable to human beings. I think the value of pure mathematics is that it lives in the heads of human beings.”
▸ Notable quotes · 13:15Chapters
0:00 — Introduction and emotional reactions: Host introduces the emergency episode on AI solving major math problems. Three mathematicians (Brunton, Moitra, Alex) share initial feelings: awe, terror, intellectual loneliness, hope. Terence Tao's quote about mourning disrupted research efforts is discussed.
11:02 — The solved axiom and human role in mathematics: Alex proposes the 'solved axiom': any problem solvable by human+AI could be solved by AI alone. Panel debates whether this destroys mathematics. Consensus: mathematics' value is human understanding and conceptual frameworks, not just theorems in data centers. The doorman fallacy illustrates the mistake of reducing mathematicians to theorem provers.
71:47 — Fermat's Last Theorem formalization in Lean: Alex describes the shock of the 13-million-line Lean formalization of FLT by Anthropic (2-3 people, minimal domain expertise). Autoformalization is now essentially solved for existing proofs, but canonization into reusable Mathlib lemmas remains a massive human effort.
66:06 — Quasi-Riemann Hypothesis and prime numbers: Alex explains RH, the zeta function as signal processing for primes, and the quasi-RH proof (zero-free strip at real part 7/8). The proof chains known techniques (semigroup) but its methods cannot reach full RH (real part 1/2). The moment of seeing the result at MoMath gala is recounted.
90:06 — Navier-Stokes and fluid dynamics: Brunton explains Navier-Stokes as conservation of momentum for fluids. The AI proof constructs a blowup with smooth forcing via an infinitely thinning vortex filament — mathematically profound but physically non-physical. The trillion-dollar engineering challenge of turbulence prediction and inverse design remains.
125:04 — Unique Games Conjecture and complexity theory: Moitra explains UGC, PCP theorem, hardness of approximation. UGC was doubted but proven true by AI. Human progress (Two-to-One Games Theorem) paved the way. Moitra worries about lost human effort and enthusiasm for field-defining problems now solved overnight.
160:24 — P vs NP, deep learning, and worst-case analysis: Moitra discusses P vs NP as the hardest remaining problem. Worst-case analysis may be wrong lens for natural problems (often 'secretly easy'). Deep learning works because natural data avoids computational hardness; understanding this requires modeling the world, not just proving bounds.
198:04 — Advice for students and the future of mathematics: Panel gives advice: be curious, learn fundamentals, pursue intrinsic motivation, pretend you already have tenure. Education is bifurcating (AI as tutor vs. answer key). Human meaning comes from relationships and the struggle itself, not just intelligence. The cherry on top vs. the main thing.
Speakers
Host — Interviewer
Steven Brunton — Mathematician, fluid dynamics expert
Ankur Moitra — Mathematician and theoretical computer scientist
Alex — Mathematician, number theorist