🤚 The Open-Palm Illumination
The most expensive way to say “please check my work” is apparently to release a mathematics library in one afternoon. In a new Moonshots discussion, Peter H. Diamandis and his guests revisit OpenAI’s extraordinary math announcement: 722 preprints concerning purported progress on 372 problems, published on October 6. The video describes this as solving 722 math problems. That is not quite what the release establishes. A paper is not a problem, a proposed proof is not a checked proof, and the velvet rope between those concepts is not decorative.
The distinction became urgent almost immediately. As Retraction Watch reported, OpenAI withdrew three manuscripts on October 7 after finding a sign error in one argument that also affected two dependent papers. Its own public correction history lists those withdrawals, revisions to 14 other manuscripts, and reference updates to another 13. This is less a story about an omniscient mathematician arriving at the salon than about a very fast one turning up with an unusually large stack of documents and requesting that everyone else find the commas.
For the avoidance of an extremely lucrative misunderstanding: withdrawal of three papers does not mean every other paper is wrong. Nor does the existence of hundreds of files mean hundreds of theorems have been independently accepted. The announcement is an invitation to examine claims, not a certificate of completed peer review.
👐 The Two-Handed Reality Check
What went wrong was specific, not mystical. OpenAI’s October 7 entry says a sign error invalidated a stabilization-trace cancellation argument in a paper on Weil classes; the construction was used by two related manuscripts involving Kuga–Satake correspondences and K3 surfaces. No one needs to pretend to have a working knowledge of K3 surfaces to appreciate the operational lesson: mathematical claims can have dependencies. One bad hinge can leave several magnificent doors leaning against the wall.
The company says its revisions to the 14 other manuscripts include proof repairs, corrected statements, clarified hypotheses and dependencies, and a citation correction. Those categories matter. Fixing a typo is not the same thing as changing the assumptions under which a theorem works. A clean audit trail lets specialists see precisely which kind of repair was required; a breathless headline mostly lets the rest of us purchase a feeling.
Retraction Watch quoted Cornell mathematician Alex Townsend suggesting that formally verified results could have been announced separately from papers still seeking scrutiny. It also quoted MIT researcher Andrew Sutherland welcoming prompt correction while describing a broader trust problem around the rollout. These are assessments from identifiable mathematicians, not proof that the entire project failed. Their point is more awkward for the launch deck: publishing research and earning confidence in it are different jobs.
In the same reporting, an OpenAI spokesperson said roughly half the results were released without confirmation, citing advice to publish before full formalization. The company’s repository history subsequently reports 300 of 719 remaining top-line results formalized, about 42%, as of its October 7 update. Formal verification is not the only way mathematicians evaluate work, and formalization is not a magic appraisal stamp for everything around a result. But it offers a much clearer status label than “AI solved math.”
🌿 The Gentle Awakening
The episode’s broader question is not whether an AI system can propose interesting mathematics. It plainly can produce a formidable amount of material for people to investigate. The question is how an institution should release that material without outsourcing its quality-control budget to an unsuspecting research community. In ordinary academic life, a preprint allows early discussion; it does not command universal attention on the strength of its publisher’s compute bill.
Consider a concrete scenario. A graduate student spots a lemma in one of the new manuscripts that could help a separate geometry paper. Before citing it, the student has to check the actual statement, its hypotheses, any dependencies, and the repository’s correction history. If the result depends on a withdrawn construction, that changes the student’s work immediately. If it survives independent checking, that matters too. The lesson is not “never use AI-assisted math.” It is “do not allow press-release certainty to replace proof-level diligence.”
This is especially important when the output arrives in bulk. The value of a proposed theorem is not measured by how many PDF files fit inside a GitHub repository. At scale, even a modest error rate can create an enormous review burden, and corrections can propagate through citations as efficiently as confident promotional copy. OpenAI’s decision to record withdrawals and revisions publicly is useful. The necessity of doing so immediately after the launch is useful information as well.
👑 The Crown Verdict
The Moonshots conversation packages the event as a preview of AI-driven discovery, and there is a serious possibility behind the spectacle: machines might help researchers find and test ideas that would otherwise take years to encounter. That possibility deserves attention. It also deserves the elementary courtesy of accurate nouns. 722 preprints are not 722 verified solutions. 372 problems are not a single leaderboard score. Three withdrawn manuscripts and 14 revisions are not footnotes to be hidden beneath a gold-foil launch announcement.
Our verdict: treat the repository as a potentially valuable, actively corrected body of research, not as a completed inventory of conquered mathematics. Reward transparent errata; demand transparent verification status; let mathematicians do mathematics before appointing the marketing department chair of the proof committee. The future of discovery may be automated in interesting ways. The obligation to show your work remains stubbornly, gloriously unpremium.
Inspired by Meta’s $14.3B Scale AI Bet Ships Muse, OpenAI Solves 722 Math Problems, & Anthropic Bans AI Cruelty by Peter H. Diamandis.
Your unverified theorem is showing. Cite wisely.