Former OpenAI Post-Training VP Pushes Back on Fields Medalists' Warning Over AI in Mathematics

In a post on X, Fedus said he shares Terence Tao's emphasis on conceptual understanding but disagrees with using the boundaries of today's models to define what is possible later. He argued that future AI systems will not only solve problems but also produce useful insights and build new conceptual frameworks that mathematicians can then explore further. If training objectives extend beyond getting the right answer to explaining why it is right, extracting methods and helping humans understand, he wrote, AI could become a research partner that supplies both answers and new lines of thought.

Tao published the statement, titled "Severe Misalignment of AI in Mathematics," on his blog on Sept. 11. It was co-signed by 25 Fields Medalists whose awards span nearly half a century, from Pierre Deligne in 1978 to Yu Deng in 2026. The signatories argue that treating the solution of famous problems as a key yardstick of model capability diverges from the mathematical community's goals of conceptual understanding, method extraction and the transmission of knowledge, and could damage the field itself.

The statement makes five main points: that solving a hard problem is only one part of mathematical research, which aims above all at understanding underlying structure and developing concepts and methods usable elsewhere; that a result must be understood before it becomes shared knowledge, passing through discussion, simplification and systematic organization before entering textbooks; that hastily released AI results may neglect prior contributions and raise questions of attribution and plagiarism; that the training of students cannot be replaced by final answers; and that AI has real potential to help mathematics, but how it is used is decisive.

The dispute has surfaced now because AI's ability in mathematics has crossed a threshold, according to the account in the statement and the public record it cites. In 2024, notable AI results were largely confined to competition mathematics such as the International Mathematical Olympiad, where systems climbed from silver-medal level toward gold. Between 2025 and 2026, models moved into open problems, including Erdős problems, geometry conjectures and Millennium Prize problems. In May, an internal OpenAI model disproved the Erdős unit distance conjecture, an open problem in discrete geometry about how many pairs of n points in a plane can be exactly one unit apart. The model constructed infinitely many counterexamples and introduced tools from algebraic number theory into the proof; the result was checked by outside mathematicians and called a milestone in AI mathematics by Fields Medalist Tim Gowers.

In early September, OpenAI said roughly 10,000 autonomous AI agents working with an internal advanced model found a proof that the three-dimensional Navier–Stokes equations develop a singularity, in about 88 hours, and formalized it in Lean. If the mathematical community ultimately accepts the result, AI would be working not on human-designed examination problems but on questions that have resisted mathematicians for decades.

Questions of credit followed immediately. In early September, Tristan Buckmaster, a mathematics professor at New York University, and Levent Alpöge, a mathematician at Anthropic, made public a study of singularities in fluid equations, while OpenAI announced a closely related proof completed by its agents. Buckmaster questioned whether OpenAI began working on the problem after learning of their unpublished direction, and the two sides argued over authorship, priority and whether the model could have come into contact with relevant user data.

Fedus also endorsed a comparison with Go. AlphaGo could not explain its moves, yet human players learned by studying its play and win rates, and he cited Noam Brown's assessment that human players improved markedly after AlphaGo appeared, suggesting mathematics could see something similar.

Mathematical bodies have begun to stake out positions. The American Mathematical Society said the recent Navier–Stokes singularity progress was not produced by AI out of nothing, but rested on generations of work by Navier, Stokes, Leray and, in recent years, Córdoba, Martínez-Zoroa, Buckmaster and Alpöge. It stated that the purpose of mathematics is human understanding. The London Mathematical Society described AI as a powerful new tool that can accelerate exploration and discovery, while posing questions, creating new concepts and forming genuine mathematical understanding remain human work. The Clay Mathematics Institute, which administers the Millennium Prize problems, did not confirm the Navier–Stokes proof, saying results still need to be public, peer reviewed and verified over time, and reiterating that the value of a hard problem lies not only in the final answer but in the new ideas, methods and theoretical tools produced along the way.