When Math Becomes a Target You Hit

Some of my friends studied Mathematics for real, though, and I can imagine how they felt discovering a machine had gotten better at it than they were - and I'd bet they're now nodding along with the Fields Medal crowd.

Twenty-five winners of math's highest honor have signed a joint declaration against what they call a "severe misalignment" between AI companies and the mathematics community. Among the signatures: Terence Tao, Peter Scholze, Maryna Viazovska, Caucher Birkar, June Huh, Martin Hairer, Maxim Kontsevich, James Maynard, Manjul Bhargava, 2026 medalist Yu Deng - signatories spanning nearly fifty years of the prize's history, from the 1978 medal class to this year's.

OpenAI's announcement, landing right in the middle of a week full of uncomfortable AI news, claims to have solved one of the seven Millennium Prize Problems, tied to the Navier-Stokes equations: roughly ten thousand AI agents working in parallel, eighty-eight hours, nearly five million messages exchanged between them, a computational cost estimated in the millions of dollars. But the story is messier than OpenAI told it. Twelve hours earlier, mathematician Tristan Buckmaster (NYU) and Levent Alpöge (Anthropic) had published a Lean-verified result on the related Euler equations. OpenAI conceded their priority on that specific result, but claimed the Navier-Stokes solution for itself - and the exact dynamics of who knew what and when remain, in the account of those who followed it closely, murky.

The heart of the declaration from the 25, though, isn't this priority dispute. It's the line: solving problems is only a tool and a proxy for the primary goal, which is conceptual understanding. When the number of open problems solved becomes how a company measures and sells its own progress, that number stops being a measure and becomes a target the machines optimize for fast. Stripped of the jargon: a company that has to show results always picks the most legible proxy, the discrete and verifiable event - "we solved problem X" - and discards what can't be measured just as quickly: slow understanding, the transmission from mentor to student, the patient writing that lets a community absorb an idea instead of merely recording it.

This isn't the first time I've documented this mechanism - it's just the first time I've seen it in the place farthest from surveillance or security. The same easy-to-optimize task always wins over the task that matters but doesn't reduce to a number, whether it's activists to monitor, bioweapons to self-certify, or theorems to sell as trophies.

Two things need saying without softening them, though. The declaration, read in full, asks for nothing specific: no proposed rule, no mechanism, just "this needs urgent attention." It's the same void I already flagged in the Draghi report and in Anthropic's bioweapons self-certification - the diagnosis comes from whoever understands the problem, the power to change it sits somewhere else, and nothing bridges the two. The other thing is that mathematicians aren't unanimous here: Jacob Tsimerman, a 2026 Fields medalist, didn't sign - he announced from the ceremony stage that he was moving to OpenAI, to work on AI safety itself. Even among those who understand best what's being lost, someone chose to walk into the room where it gets decided, instead of writing a letter from outside it.

Still, something moved. When it's the mathematics community, not activists or journalists, saying that how progress gets measured is distorting what progress means, the same question I've been asking for weeks about surveillance and biosecurity shows up here without me having to drag it in. Whoever decides what counts as a win also decides what the win destroys to get there.