What happened in this episode?
Lex Fridman convenes Steven Brunton (University of Washington), Ankur Moitra (MIT) and Rutgers number theorist Alex Kontorovich to assess recent AI-released mathematical claims attributed to OpenAI and Anthropic, and what they mean for research culture. The guests describe reported results spanning fluid dynamics, number theory and complexity theory, while stressing that correctness, interpretability and community verification remain separate problems.
Kontorovich focuses on formal proof assistants, explaining Lean and the Mathlib library, and argues that “autoformalisation” — producing machine-checked proofs of existing human mathematics — may now be effectively solved. He cites an Anthropic release he says contains a “sorry-free” Lean formalisation of Fermat’s Last Theorem running to roughly 13 million lines, while noting substantial work would remain to refactor such proofs into reusable, canonical library form. He also outlines a “quasi-Riemann hypothesis” claim, describing a new infinite zero-free strip for the zeta function between real parts 7/8 and 1.
Brunton explains the Navier–Stokes regularity problem and recounts an AI-produced construction he says generates finite-time blow-up in a forced setting via an increasingly thin, stretched vortex filament, while arguing this has limited immediate impact on day-to-day engineering turbulence problems.
Moitra discusses the Unique Games Conjecture and says OpenAI’s release asserts it is true, with broad implications for hardness of approximation. He also raises concerns about institutional strain, potential secrecy in idea-sharing, and AI safety, describing an “owl bias” fine-tuning thought experiment as an example of unintended behaviour transfer.