AI Just Proved It Can Do Original Math. A Fields Medalist Just Chose to Police It

Column Overview

Two announcements landed a week apart in the summer of 2026, both touching the same field and, as it turned out, the same company. Neither was framed as a response to the other. But read together, they say more about where the AI industry actually stands than either one does alone.

The week in question

On July 23, 2026, at the International Congress of Mathematicians in Philadelphia, Jacob Tsimerman received a Fields Medal for his proof of the André-Oort conjecture — one of the four medals mathematics hands out once every four years, and about as close as the field gets to a lifetime-achievement award granted mid-career. Within hours of accepting it, Tsimerman announced he was taking a leave of absence from the University of Toronto, keeping his tenured position but stepping away from active teaching, to join OpenAI's AI safety team starting in August 2026.

A little over a week later, on roughly August 2, OpenAI published results from Astra, a model it has not yet released publicly. Astra had worked through ten long-open problems in mathematics and theoretical computer science — several of them unresolved for decades, one for 27 years — at a total inference cost OpenAI put at under $2,000. The company released a 249-page manuscript alongside the work and posted machine-checked formal proofs to GitHub.

Nobody scheduled these two events to land in the same window. That's precisely what makes the coincidence worth sitting with: a company shipped evidence that its models can generate mathematics good enough to interest working mathematicians, and in the same stretch of days, one of the most decorated living mathematicians decided the most useful thing he could do with his career right now was go help make sure that kind of system behaves the way it's supposed to.

What Astra actually did

It's worth being specific here, because "AI solves hard math" has been said before about results that turned out to be either not that hard or not quite solved. Astra's list is unusually concrete.

The headline result is a construction of a non-sofic group — an abstract algebraic structure proven to resist approximation by any finite structure, no matter how large. Sofic groups are, loosely, groups that behave well enough that you can approximate them arbitrarily closely with finite pieces; showing a group that provably can't be approximated this way is a different kind of result, because you have to rule out every possible finite approximation, not just fail to find one. No one had produced such a construction in 27 years of people trying.

Astra also pushed forward the upper bound on sphere-packing density in higher dimensions, a problem that had been effectively stuck since the late 1970s, moving it closer to the threshold established by Cohn and Elkies. It produced a counterexample to Alain Connes's rigidity conjecture about von Neumann algebras — disproving, rather than confirming, a standing conjecture from one of the field's more influential figures. It proved the Ehrhart volume conjecture. And it closed out three problems from Paul Erdős's well-known catalog of open questions, including problem 183, on multicolor Ramsey numbers.

The detail that matters most for judging how seriously to take all this: the proofs were formalized in Lean 4, and the "sorry" count in the published repository is zero. In Lean, "sorry" is the placeholder mathematicians insert when a step isn't actually proven yet — a flag for "trust me, this part checks out." A zero count means every logical step in all ten proofs has been mechanically verified, not just written up in prose that reads convincingly. That's a meaningfully higher bar than a model producing a plausible-sounding argument. As of this writing, mathematicians who've reviewed the preprints have responded positively on an informal basis; the results haven't yet gone through formal peer review at a journal.

Why a Fields medalist would choose this, specifically

Tsimerman's own work is proof, in the literal sense — his career output is a stream of statements shown to be true beyond doubt, using tools built for exactly that purpose. AI safety research, as a discipline, has been drifting toward needing the same skill set: verifying that a system behaves within some formally stated specification, checking whether a machine-generated derivation actually holds up, reasoning rigorously about what a model's outputs can and cannot be guaranteed to do. It's not an obvious career pivot on the surface, but it maps onto his actual training closely enough that it isn't really a pivot at all.

He's been direct about the reasoning: "The risks are super high, the stakes are super high, so we need a very high level of assurance." His stated interest is specifically in the mechanics of oversight — understanding how AI agents behave as systems, and deriving the kind of proofs that could establish, with some rigor, that a complex system won't act outside expected bounds.

One detail is worth pulling out because it cuts against the easy reading of this move. Tsimerman previously wrote a paper cataloging, systematically, the range of ways AI could plausibly lead to catastrophic outcomes. This isn't the profile of someone wandering into safety work out of vague unease, or of an uncritical booster taking a well-paid detour. It reads more like someone who did the risk assessment first and chose to work closer to the problem, not farther from it.

What the coincidence actually tells us

The temptation is to read these two stories as evidence for whichever side of the AI debate you already occupy — either "look how capable these systems are getting" or "look how worried the experts really are." Both readings miss what's actually notable, which is that the same week produced both signals, and they trace back to people with the deepest visibility into where AI capability genuinely stands right now. This isn't optimists and pessimists shouting past each other. It's the same population of people who understand the technology best treating capability gains and safety work as parallel, simultaneous priorities rather than opposing camps.

The formal-verification detail ties the two threads together more literally than it might first appear. Astra's math results carry weight because they're machine-checkable — the value isn't that the proofs sound right, it's that a verifier confirmed each step holds. Tsimerman's stated interest in AI safety is built around bringing that same standard of rigor to a different question: not "does this proof hold," but "does this system behave the way we expect it to." One effort pushes outward, expanding what's provably true in mathematics. The other pushes inward, trying to make what's provable about an AI system's behavior more solid than "an expert read the outputs and it seemed fine." Different targets, same underlying method.

Industry observers are likely to read this as evidence that leading AI labs are now running two build-outs in parallel rather than choosing between them: one aimed at demonstrating that model capability is still climbing quickly — which is, not incidentally, the core of the commercial and fundraising narrative — and one aimed at recruiting people capable of rigorously examining what comes with that capability. Framed that way, publishing a capability milestone and hiring a Fields medalist for safety work in the same window isn't two conflicting PR moves. It's two expressions of the same underlying calculation about where the field is headed.

An open question, not a closed one

There's a harder question sitting underneath all of this that neither announcement answers on its own. When a model proves a theorem in a form so intricate that even trained mathematicians need a machine to check it line by line, what does that imply about how we ought to be evaluating whether an AI system itself is safe? Does that evaluation need to move toward the same standard — formal, mechanically checkable claims — rather than resting on domain experts reading through outputs and forming a judgment that they seem reasonable?

Nobody has answered that yet, including Tsimerman. But the fact that a mathematician built for exactly that kind of rigor chose, at the peak of his field's recognition, to go work on it, might be the closest thing to an answer the moment has produced so far.