How AI Is Upending the World of Mathematics

AI’s progress in mathematics is striking: models that once failed simple tasks such as counting letters are now producing sophisticated proofs and contributing to difficult research. For AI companies, math is both a test of reasoning and evidence in claims about progress toward AGI; for mathematicians, the shift can unsettle the value of years of work. But a striking headline is not the same as a settled result. Tracy Alloway said she could not assess the significance of a claim that a model had solved Navier–Stokes, and the result discussed in the episode had not been officially verified.

The Navier–Stokes equations describe fluid motion, and a major open question concerns whether their solutions exist and remain unique over time. The reported AI-assisted result was described as a counterexample involving a specially arranged fluid configuration—a “very special spoon”—that develops a singularity in finite time. The setup was characterized as highly unusual and non-physical. Human mathematicians in Spain had reportedly worked out much of the construction on paper, with AI contributing important insights and the final leg rather than producing the result from nothing. One speaker judged the proof probably correct but not elegant, while stressing its technical density and the need for community scrutiny.

The episode distinguishes generating a proof from establishing that it is sound. AI can produce proofs while also making mistakes when asked to check its own work, so one approach is to pair a model’s proof with Lean, software that encodes mathematical rules and checks whether the steps follow. Verified theorems can then be reused, and some AI-generated proofs already run to hundreds of thousands of lines of Lean code. Even this process has a trust boundary: verification depends on the checking software itself being correct. In engineering and simulation, where mathematical results support calculations about the physical world, that distinction matters.

Mathematics is not only about proving results; it also involves choosing worthwhile questions, modeling problems and recognizing insight. The boundary between pure and applied work is blurry: abstract mathematics can later support applications such as cryptography, while simulations used for movie effects depend on mathematical models. The professor in the discussion described research as overwhelmingly collaborative and rejected the idea that progress comes mainly from solitary “great men.” Speakers saw today’s AI tools as particularly capable of finding surprising connections among known ideas, but said it remains unclear whether they can originate new theories or identify the next important question.

If AI makes proofs abundant, checking them and deciding which ones matter may become bottlenecks. The Navier–Stokes result’s authorship and originality are difficult to assess: for some generated proofs, it is unclear whether a model found a result by brute force or reproduced earlier work from its training data, and systems are poor at attributing sources. Academia also lacks settled rules for credit when people formulate a problem and AI produces or verifies a key step. Meanwhile, review systems are strained: ICLR received about 60,000 submissions at a recent deadline, compared with roughly 1,000–2,000 about 10 years earlier. Speakers said the volume weakens established academic signals and raised concern that expensive models could advantage well-funded labs in competitions for major prizes, including a Clay prize described as worth $1 million.

The same shift is changing teaching and access. One MIT course added short quizzes based on homework problems and reworked a project around presentations and interaction after students began scoring nearly 100% on homework, making it harder to tell whether they had learned the material or used AI. The instructor compared homework to practice and argued that understanding must be assessed separately. They emphasized teaching students to frame questions, model problems and develop judgment—skills relevant in fields such as finance and engineering, where deciding what to work on can matter as much as executing a solution. Access is uneven: commonly used tools rely on company servers, a preferred cloud account for a PhD student can cost about $200 per month, and speakers were uncertain how much compute or money went into the Navier–Stokes work.