On October 6, OpenAI pushed 722 mathematical manuscripts — organized into 372 result families — into a public GitHub repository called openai/math, alongside Lean formalizations for many of the results and abridged summaries of the model’s reasoning. The model behind the work is internal, unnamed, and unreleased. By the following morning the Hacker News thread had run to 1,252 points and 1,421 comments, which is the internet’s way of announcing that everyone had already chosen a side.
The two sides are the wrong two sides. One camp calls it a landmark: the machine is doing real mathematics. The other calls it a stunt, because an unnamed model and a Git commit are not evidence. Both are arguing about whether the theorems are true and whether the lab can be trusted. Neither is arguing about the thing that actually broke on October 6.
The Objection Was Never That the Proofs Are False
Look at what the mathematical community actually filed.
On September 29, an advisory group on mathematics and artificial intelligence published responsible-release recommendations after collecting more than 600 replies from working mathematicians. Buried in the document is a footnote explaining what the survey had asked about: a situation in which OpenAI announced the existence of a large number of results without giving details.
Read that footnote twice. The objection is not “your theorems are wrong.” The objection is to sequencing — announce first, disclose later, at a pace the announcing party controls.
This matters because mathematics is the one field where the standard complaint about AI cannot get traction. For three years the standing demand has been: stop showing us benchmarks, show us something checkable. Math has a mechanical checker. It is called Lean, and OpenAI used it — formalizations for many of the result families, sitting in the repo for anyone with the patience to run them. The demand was met. The mood did not improve. That tells you more than the release does: the demand was never really about verification.
Mathematics Pays in Citations, Not in Truth
The apparatus of the field — arXiv, journals, referee reports, tenure letters, the medal with the Scandinavian name — is a credit-allocation machine. A theorem’s value to a career is its timestamp and its byline. OpenAI’s byline is “OpenAI.” The company’s September 21 advisory post claimed its internal model had resolved more than a hundred long-standing open problems. If you are a 34-year-old assistant professor who has spent five years on one of them, Tuesday’s announcement did not read as a gift. It read as a revaluation of your inventory.
That is not envy. It is a solvency problem, and the right should understand it better than anyone, because a credit system is an incentive system. Kill the reward and you do not get faster mathematics. You get fewer people choosing the hard problem, because the hard problem was never the fun part — it was the part that paid.
A postdoc in combinatorics, between sessions at a conference, phone pressed to one ear in a hotel corridor, put it in nine words: “Nobody thinks the proofs are fake. They think the byline is.”
The ‘Release the Model’ Chorus Is an Industrial Demand
Now the part that should make the critics squirm. The loudest complaint — release the weights — is not a verification request. Verification already happened, or it can, in a Lean kernel. Weights are a competitor’s asset. When a rival lab demands the model, it is not asking for epistemics; it is asking for a head start, with a philosophy footnote stapled to it.
There is one honest gap in the release, and it is smaller than either camp admits. A formalization certifies that a statement follows from axioms. It does not certify that anyone understands why the result is interesting, or how it connects to the next thing. Abridged reasoning summaries are a thin substitute for an idea a human can build on. The fix for that is not open weights. It is archival access for referees under embargo, which is cheap, boring, and would have prevented most of this.
The Fix Costs OpenAI Almost Nothing
Here is what the company should have done, and can still do. Publish each result family as a timestamped, citable note with a stable identifier. Name human collaborators where any exist. Give referees embargoed access to the full reasoning traces rather than abridged ones. And commit publicly to a simple rule: if an independent preprint appears first, you don’t claim the family.
None of that surrenders the model, weakens a moat, or costs a meaningful dollar. All of it converts a Git commit into a citation — which is the only currency a mathematician can spend.
The uncomfortable version for the guild: a citable corporate byline beats a rival lab’s secret model, and the community should take that trade without pretending the alternative is purity. The uncomfortable version for OpenAI: you cannot buy a seat in a credit system by being right about the mathematics. You buy it by doing the tedious parts — attribution, timestamps, referees, patience.
The proofs appear to be real. That was never the fight. The fight is over who gets to hand out the receipts, and on October 6 the answer was: the people who hadn’t asked.