Mathematicians tell Lex Fridman how AI claims are unsettling research and mentorship

AI’s claimed mathematical advances are creating a split between research opportunity and the work of training researchers, mathematicians told Lex Fridman in a newly published interview. MIT’s Ankur Moitra described tools that help him connect ideas across fields, but also make it harder to guide students toward work that will matter.

Published October 26, Lex Fridman Podcast #503 brings together Moitra, University of Washington mathematician and fluid-dynamics expert Steven Brunton, and Rutgers mathematician Alex Kontorovich. Fridman frames the conversation as a response to machines claiming solutions to Navier–Stokes, the quasi-Riemann hypothesis, the Unique Games Conjecture and other mathematical problems. The interview does not establish that those claimed proofs are valid.

Better tools, a harder apprenticeship

Moitra’s enthusiasm centers on asking better questions. He describes AI as something like an expert who can explain another field in a researcher’s own language, making connections easier to find. That is useful for the way he works: starting with questions rather than simply pursuing a predetermined result.

But mentoring graduate students and postdoctoral researchers is also a part of the job he values. Moitra says the familiar blueprint for helping them succeed has been disrupted. Advising someone on a problem now means confronting anxiety about whether their work can stand out amid AI-produced results.

His concern extends beyond competition for recognition. Great results can begin with modest questions, he says, and the pressure to produce something momentous risks crowding out patience for that slower exploration. He describes feeling excited by the tools but also more intellectually lonely, because bringing younger researchers along on those journeys has been central to his work.

Does knowing the answer end the journey?

Brunton offers a more hopeful view grounded in engineering. He expects stronger mathematical tools to improve work on practical problems, including artificial hearts and airplanes. He also acknowledges the fear of losing control or being excluded from a process researchers enjoy. His optimism is not a prediction that human participation is assured.

He lays out two possible futures over the next five or 10 years: machines dwarf human intelligence, or progress levels off and AI becomes an extraordinary tool for human curiosity. On motivation, he resists a simple conclusion that an existing answer makes exploration pointless. Knowing a solution is possible, he argues, can itself drive people to understand and reconstruct it.

Kontorovich draws a separate distinction between solving mathematics and destroying it as a human activity. He treats a future in which mathematics is “solved” as a hypothetical premise, not a description of today’s capabilities.

From released results to shared understanding

The discussion follows OpenAI’s October 6 mathematics release. In its publication, OpenAI said it was sharing results from an internal frontier model through GitHub, alongside formalizations of many proofs in Lean—a programming language that allows computers to check mathematical proofs. It said more formalizations would follow as they became available.

OpenAI also said it had consulted the Institute for Advanced Study’s independent Advisory Group on Mathematics and Artificial Intelligence about release practices. For future publications, it committed to improving citations, mathematical explanation and presentation so the results would be easier to understand. Those commitments address the communication of results; the interview’s guests are also wrestling with what researchers should do with their working lives.