When the Fields Medalist Admitted AI Beat Math

The Quiet Earthquake in a Room That Hasn’t Shouted Since Babylon

Terence Tao did not say artificial intelligence was better at math than he was. He said something sharper — and harder to parse for people who have never watched their life’s work become obsolete overnight.

On October 8, 2026, Tao wrote through a communal blog called Proofs and Prompts, which gathers individual commentary from mathematicians around the world. His paragraph was short. Its weight registered slowly.

He called it a “deep sense of frustration” that nobody was there to explain the proofs, to answer the questions, to teach students what the reasoning actually meant. Not because the proofs were wrong — Tao acknowledged they contained genuinely useful new ideas. But because mathematics has never been only about answers. It is about the transmission of understanding, the chain of intuition passed from one human mind to another, the conversations at blackboards and conference corridors that turn a correct derivation into something a student can replicate, extend, fight with.

What OpenAI had just dumped — 722 manuscript-length documents covering more than 300 unsolved problems across number theory, algebraic geometry, and differential equations — was technically a breakthrough. Culturally, it was a rupture.

Who OpenAI Actually Interviewed

Before we get carried away by the numbers, consider what OpenAI chose to publish. These are not elementary exercises. They are active frontiers — problems that doctoral candidates spend years circling, that tenured faculty include in grant proposals, that Fields Medalists reference in keynote speeches as examples of the field’s living challenges.

Hugo Duminil-Copin, the 2022 Fields Medalist, put it plainly: his own lecture notes, published papers, and funding applications all listed problems that AI had now solved before anyone in the mathematical community had finished thinking through them. He did not say this gleefully. He described it as “huge shock” and immediately flagged the downstream damage to colleagues and students whose career trajectories depend on publishing incremental steps toward these same targets.

Peter Scholze — the youngest Fields Medalist ever, now at the height of his influence — skipped the philosophical complaint entirely and went straight to the practical hazard. Some encryption systems protect data by relying on number-theoretic problems that computers find excruciating to solve. If AI can generate efficient paths through those problems, the mathematical assumption that keeps your banking password secure may have just been invalidated by a model that nobody fully understands.

Scholze did not specify which systems were at risk. That omission is itself a signal: the people best positioned to assess the damage are also the ones most reluctant to announce it publicly, because revealing vulnerabilities is how adversaries learn where to aim.

The Algorithm Doesn’t Sit Down With You

The deeper anxiety Tao and his peers are articulating is not about competition. It is about the hollowing out of a process.

Mathematical research, at its best, is conversational. A graduate student proves a lemma, defends it at a seminar, gets chewed apart by a postdoc from another university, revises, submits, receives a referee report that is partly hostile and partly charitable, and in that friction the work becomes legible to other human minds. The value is not merely that the theorem is true — every correctly stated theorem is true regardless of who found it — but that the pathway to truth leaves a trail other humans can follow.

A model that generates a correct proof without the capacity to narrate its reasoning leaves a different kind of artifact: a black box with impressive output. The proof works, but the intuition does not transfer. Students cannot learn from it. Junior researchers cannot build on its methods. The intellectual genealogy is severed.

This is why Tao’s complaint about the absence of an explainer matters more than any technical assessment of the proofs themselves. The model proved theorems. Nobody proved why the theorems were worth proving, or how the underlying strategy might generalize to the next generation of problems. That narrative — the scaffolding that turns a result into a research program — is almost entirely human work. OpenAI did not build it.

What DeepMind Actually Published in Science

Two days before Tao’s blog entry, Google DeepMind’s own contribution appeared in peer-reviewed form in Science, after an earlier May release. Their system solved 9 Erdős problems and 44 conjectures about integer sequences by generating candidate proofs, running computational verification, and iterating. The journal’s editorial note was remarkably even-handed: it observed that even failed AI attempts at proof can catalyze human understanding and subsequent research.

That claim deserves scrutiny. Failed attempts are useful precisely because they leave visible tracks — wrong turns that human researchers can examine, critique, and learn from. A successful AI-generated proof that arrives without trace is not obviously more or less useful than a failure. It is simply opaque. Whether that opacity inhibits or accelerates human progress depends entirely on whether other researchers can reverse-engineer the model’s strategies fast enough to build on them. So far, there is no evidence they can, and plenty of reason to suspect they cannot.

Who Wins, Who Loses, What Happens Next

Winners: AI companies that publish results faster than any single human team could, creating an impression of velocity that attracts further investment. Researchers who already work at the boundary of what AI can assist with, because they are best positioned to extract usable insights from high-dimensional outputs. Students entering a field where the floor has shifted — those who learn to collaborate with AI tools will pull ahead; those who treat them as replacement mechanisms may find their training misaligned with what the field actually requires.

Losers: The junior researcher whose first publication was supposed to be the incremental step on Problem X, now preempted by a system that generated 300 incremental steps in a single afternoon. The cryptography community, which operates on the assumption that certain computational barriers exist, and which cannot afford to be caught off guard. Graduate programs built on the traditional apprenticeship model, where learning happens through the gradual reconstruction of human reasoning — a model that assumes someone already knows the reasoning well enough to pass it on.

The immediate future looks like this: more proof dumps. More initial celebration followed by careful technical auditing. Some retractions, as OpenAI itself has already acknowledged, when explanations prove thin or derivations contain hidden gaps. A period of uncertainty during which mathematicians assess what the published results actually contain versus what they only appear to contain at first glance.

Then, the harder question — the one Tao’s paragraph pointed toward without stating directly. If AI can generate correct proofs at scale but cannot transmit the human reasoning that makes those proofs intelligible, the profession must decide whether to treat mathematics as a production of valid statements or as a discipline of shared understanding. Those are not the same thing. They were always distinct. AI has simply made the distinction impossible to ignore.