The announcement that shook math

On September 8th, 2026, OpenAI said it had produced an AI-generated solution to Navier-Stokes existence and smoothness, one of the seven Millennium Prize Problems.

A Deep Blue moment

Tristan Buckmaster, one of the mathematicians in the Navier-Stokes saga, put it bluntly. The comparison to machines beating humans at games was immediate.

“This is a Deep Blue, Kasparov moment.”

So is math solved now

If AI can answer questions at the frontier, the worry follows fast. Should mathematicians do what chess and Go players did, and just keep playing?

Games are not only about winning

Philosopher C. Thi Nguyen argues the point of a game is the process, not just the aim. Nobody plays Twister to win. A game of Go can be an achievement in defeat.

Do not retreat, reject

De Toffoli and Duede say keep playing concedes too much. The better move is to refuse the premise that mathematics is a game in the first place.

Two bad assumptions

The AI solved math story rests on two claims: that AI really solved a mathematical problem, and that math is only about solving problems. Both are wrong.

Not all answers are solutions

OpenAI answered whether Navier-Stokes can develop a singularity: yes. Deep Thought answered life, the universe and everything: 42. Neither is what anyone actually wanted.

What OpenAI handed over

Two artifacts, not one number. A Lean formalization certifying validity, plus a manuscript that appears to carry the matching informal proof.

Proof has two meanings

The logical notion: deductive validity a machine can check without understanding anything. The intelligible notion: an argument a person can grasp, share and build on.

Lean nails one of them

A formalization meets the logical standard exactly, and that is a real contribution. It secures certainty. It does not yet hand the community ideas it can use.

Mathematicians want to know why

Thurston's point: proof is for understanding. They want to know what makes a proposition true, in ideas they can communicate and connect to what they already know.

The two used to travel together

No human could build a monstrous logical proof without first grasping the shareable ideas underneath. Formalization mostly checked what was already understood.

AI pulls them apart

Now we can get formal proofs floating free of any intelligible one. The two notions of proof, long welded together by human limits, come apart.

The opposite failure is real too

Jaffe and Quinn pointed at Thurston's geometrization work: a great idea with incomplete proof. Hales formalized Kepler precisely to check an argument nobody could verify by hand.

“a roadblock rather than an inspiration”

Why the Clay problems exist

The seven were not chosen because anyone wanted seven answers. They were chosen because the proofs were expected to be fruitful.

“Because a proof gives not only certitude, but also understanding.”

Even genuine proofs would not settle it

Assume future systems deliver proofs that are both certified and fully intelligible. Math still would not be solved the way chess was.

There is no checkmate

Chess players compete for fame and jobs, and they also win chess. Mathematics has no victory condition. Nobody wins math.

Treat it as a tool

Jeremy Avigad's framing: technology built to serve our purposes. We do not race the car we drive, or shout over the phone we use.

“It is misguided to think of mathematicians as competing with AI.”

Math does more than solve

Tao's own list: build theories and techniques, understand the world, sustain a community, train the next generation, add to cumulative knowledge, make beautiful things.

The choice, not the ending

Twenty-five Fields Medallists have warned of a severe misalignment between AI companies and mathematics. If success becomes certified answers, math shrinks to what is easiest to benchmark away.