By ByteBot
On October 6, OpenAI dropped 722 AI-generated mathematics manuscripts on GitHub — produced by an unnamed, unreleased internal model — covering 372 open problem families including a claimed quasi-Riemann hypothesis result. Released under Apache-2.0, no peer review, no named author: just a black-box model and a dump of papers that promptly hit #1 on Hacker News with 751 comments. The openai/math repository is live today. The math community’s response has been close to fury. Twenty-five Fields Medalists signed an open letter. Terence Tao warned publicly about Goodhart’s Law. And developers now have 722 advanced math papers they can legally use to train AI models right now.
OpenAI Math Manuscripts: What’s in the Repository
The repository contains 722 manuscripts organized into 372 families, spanning pure mathematics, theoretical computer science, and mathematical physics. The model was given roughly 4,000 problems; 722 (18%) were judged publishable by OpenAI. Each includes an abridged reasoning summary and a BibTeX citation block. The four biggest claimed results: a quasi-Riemann hypothesis result (proving the zeta function has no zeros with real part above 7/8), the Unique Games Conjecture, the rational Hodge conjecture for CM abelian varieties, and the free group factor problem. None have been confirmed by independent mathematicians yet.
The compute cost tells its own story. The September Navier-Stokes claim required 10,000 parallel agents running over 88 hours at a cost of millions of dollars. These 722 manuscripts averaged approximately three hours of ChatGPT Pro thinking compute per paper — a single agent, a single prompt. AI math research is getting dramatically cheaper, and fast.
The Verification Problem Is Real
Only 162 of the 722 manuscripts — roughly 22% — have fully formalized Lean proofs. Another 73 have partial Lean coverage. The remaining 487 papers are unverified beyond OpenAI’s internal judgment. OpenAI’s own README cautions that “some of the unformalized results could have issues,” which is an unusual degree of candor about incompleteness in a flagship release.
For developers, the practical split is clear: the 162 Lean-formalized papers are machine-checkable and safe to use as reliable training data or research starting points. The remaining 560 require human verification before you can trust them for anything beyond noise-robust training pipelines. The formalization breakdown from cellcog.ai is the clearest guide to which papers are in which camp — the openai/math README also has a formalization catalogue per paper.
Related: arXiv Caps Submissions: AI Flood Breaks Science Publishing
Why Mathematicians Are Not Applauding
This is the second major OpenAI math release in a month — the Navier-Stokes claim in September was the first — and the response from working mathematicians has been closer to alarm than applause. Twenty-five Fields Medalists signed an open letter calling the approach “harm to the science of mathematics.” Nearly 4,000 researchers signed the Leiden Declaration demanding responsible integration of AI tools in research. Scientific American described the field as “already in shock.”
Terence Tao’s concern cuts deeper than academic gatekeeping. At the 2026 ICM plenary, he warned that “the inherently ungrounded nature of generative AI, combined with the financial incentives of AI companies, makes the use of these tools particularly vulnerable” to Goodhart’s Law — the tendency of any metric, once made a target, to stop measuring the thing it was meant to measure. In other words: if “solving famous math problems” becomes the benchmark for frontier model capability, labs will optimize for that benchmark specifically, regardless of whether the proofs are actually correct or useful to the field.
The peer review bypass is the structural issue underneath all of this. By releasing 722 manuscripts directly to GitHub, OpenAI sidesteps the editorial and review infrastructure that mathematics has used for validation since Euler. There is no human author to hold accountable. The generating model cannot be audited — it is explicitly unnamed and unreleased. Moreover, the pace — 722 papers overnight — outruns the field’s capacity to check the work before the next dump arrives.
Key Takeaways
- OpenAI released 722 Apache-2.0 AI-generated math manuscripts from a secret model on October 6; the repository is at github.com/openai/math and is immediately usable for AI training
- Only 22% (162 papers) have full Lean formal verification — treat the other 78% as unverified until independent mathematicians confirm them
- The four biggest claimed results — quasi-Riemann hypothesis, Unique Games Conjecture, Hodge conjecture, free group factor problem — have not been independently confirmed as of today
- AI math compute costs are collapsing: Navier-Stokes needed millions of dollars; these 722 papers averaged roughly three hours of ChatGPT Pro thinking compute each
- The mathematics community’s concerns are legitimate, not territorial — peer review bypass and an unauditable secret model create genuine epistemological problems that extend well beyond mathematics
