This is a follow-up to our May 22 coverage of what mathematicians are saying about OpenAI’s 80-year geometry proof. The core event, an OpenAI internal reasoning model reportedly disproving the Erdős Unit Distance Conjecture, first posed in 1946, has been covered here. What’s new is who’s vouching for it, and what that choice signals.
According to secondary reporting, the companion verification paper was co-authored with a group of external mathematicians including Fields Medalist Timothy Gowers, who holds the Combinatorics chair at the Collège de France. The specific co-author list and total count couldn’t be confirmed against the paper directly, the companion paper’s arXiv identifier wasn’t available at time of publication, but Gowers’s engagement with the result is corroborated across multiple sources. What makes this more than a credentialing footnote: Thomas Bloom, who according to reports previously identified errors in an earlier OpenAI mathematical claim in October 2025, is also listed as a co-author in secondary coverage. That’s not incidental. A mathematician who caught a prior overstatement lending his name to verification is a different kind of endorsement than a researcher with no prior record of skepticism.
The proof itself, according to OpenAI, spans 125 pages. The companion paper’s authors note it was “first mathematically generated in one shot” by the model, then “expositionally refined through human” involvement, per secondary reporting. Don’t read “one shot” as fully autonomous end-to-end delivery, the human refinement step is part of the documented process, and the distinction matters for anyone evaluating what this actually demonstrates about AI reasoning capability.
Analysis
The 'one shot' framing in OpenAI's announcement describes autonomous generation of the original proof, not fully autonomous end-to-end delivery. Human expositional refinement followed. Research teams evaluating AI-assisted science tools should note this distinction before drawing capability conclusions.
A Princeton mathematician reportedly published a follow-up paper extending the result with a calculated super-linear lower bound. The specific exponent and arXiv link couldn’t be confirmed at publication time, so the numerical details are withheld here until the citation is verified. The existence of a follow-up extension is itself meaningful: independent mathematicians are treating the result as worth building on, not just as a press release worth commenting on.
The catch is this: none of the primary source URLs for this event were accessible during verification. What’s confirmed here comes from secondary reporting, TechCrunch, Hacker News, and the AIchats Substack, all reporting on the same OpenAI announcement. That’s not independent mathematical verification of the underlying proof. It’s secondary coverage pointing at a primary paper that’s not yet directly accessible here.
Evidence
What to watch
When the companion arXiv paper becomes directly accessible, check the co-author list, the verification methodology, and whether Bloom’s specific role is authorial or advisory. That distinction will sharpen what the result demonstrates about AI-assisted mathematical research. Also watch for independent benchmark classification, OpenAI’s official channel has framed this as a breakthrough, but third-party mathematical evaluation is what establishes that framing.
TJS synthesis
The Bloom credibility arc matters more than the co-author count. When a researcher with a documented record of catching AI overstatements becomes a verifier, that’s peer review with stakes. It doesn’t confirm the proof is correct, that requires reading the paper, but it raises the prior probability meaningfully above a standard vendor announcement. Wait for confirmed arXiv access before citing specific technical claims. The result’s direction looks credible; the precise details aren’t independently verified yet.