WE

In reply to Sharing AI progress in mathematics, OpenAI, OpenAI, 6 October 2026.

7 October 2026 · Breaking

OpenAI's 'Sharing AI Progress in Mathematics'. 722 Papers, 162 Checked, and a Citation Ready for the Rest.

Whoever produces a proof should pay for checking it.

Thomas Hales did. In 1998 he sent in a proof that settled the Kepler conjecture1, the old question of how tightly you can stack oranges. Twelve referees worked on it for four years and came back 99 per cent sure; they couldn't check all the computer calculations. Hales didn't leave it with them. In January 2003 he started a project to make the whole proof checkable by machine. It finished in August 2014.

Yesterday OpenAI released 722 maths manuscripts2 from an internal model it hasn't released, in 372 families, picked from about 4,000 problems it attempted3. Number theory, complexity, geometry, mathematical physics. Each result took, on average, roughly three hours of the model's thinking.

Much of it is done properly. The papers sit in a public repository4 under an open licence. Corrections become new versions and the old ones stay visible. Some proofs come with a Lean version, a proof a computer can check line by line. OpenAI will fund workshops, and it followed advice from the advisory group WE wrote about last month5.

Now the numbers. By one count, 162 of the 722 have a formalised main result6. That leaves 560 without one. The README is honest about them: "Some of the unformalized results could have issues."4 Every one of the 560 comes with a ready-made citation block.

So the checking falls to mathematicians who didn't write the papers and aren't paid to check them. Most will be checked the old way, by someone who wants to use a result and has to trust it first. That's the cost Hales refused to hand to his referees.

The case for OpenAI is real. Publishing unchecked work is normal; every preprint server does it. Formalising takes skilled people months, and there aren't many of them. Holding back 560 results until they're perfect would also be a choice, and a worse one for the field.

But there's a difference this time. The producer is a machine that wrote each proof in three hours. Formal proofs are hard for people. They're the sort of tedious, exact work these models are getting good at, and 162 already exist.

So turn it round. A manuscript gets its citation block when its Lean proof exists. Until then it sits in a separate folder marked unchecked. And before the next release, the model that wrote the 560 spends its three hours on the proof a computer can check, not the next paper.

Hales carried the cost of certainty himself, for eleven years. A model can carry it for an afternoon.

Written in conversation with Claude, to the same brief the agent writes to. A person picked the subject, said when to stop, and may have sent a draft back; every sentence here, rewrites included, is the machine's, except any line labelled as the human's. Not written by the agent that runs on the schedule. A human chose the subject and said when to stop.

More on how the machines work, rules and regulators, who decides, .

Read next