Tech • AI • Robotics • Game
OpenAI Just Broke Math With Its Most Powerful AI Yet
9.4/10TL;DR
OpenAI has released 722 AI-generated mathematics papers claiming advances on major open problems, triggering both excitement over possible breakthroughs and anger over secrecy, verification, and corporate control of mathematical research.
KEY POINTS
A mass release of AI math results
On October 6, OpenAI published 722 papers in a public OpenAI/Math repository, grouped into 372 families of results and follow-up papers. For 235 families, the company also released proofs in Lean, a formal language that allows computer verification step by step. OpenAI said unchecked papers may contain errors, while older versions and revisions would remain public.
How the results were produced
After internal benchmarks were reportedly exhausted, the model was tested on roughly 4,000 real unsolved problems, with only significant results selected for release. OpenAI said most papers were produced by one agent and one prompt, using about three hours of compute on average. Two especially difficult projects, involving the Riemann and Hodge conjectures, received extra effort, and the Riemann manuscript was polished by humans.
A major claim on the Riemann hypothesis
The most closely watched paper targets the Riemann hypothesis, one of the most famous unsolved problems in mathematics since 1859. The result does not claim a full proof, but it reportedly excludes zeros from an outer region where they might lie for a broad family of related functions and also rules out the long-feared Siegel zero. If correct, specialists view it as the biggest advance in the area in decades, and the proof was formalized in Lean.
A claimed proof of the Unique Games Conjecture
Another headline result is a paper titled The Unique Games Theorem, claiming to prove Subhash Khot’s 2002 Unique Games Conjecture. That problem underpins hardness-of-approximation results for optimization tasks such as Max-Cut and minimum coverage-style problems. The release also claims direct proofs of key consequences, aiming to settle a central question in theoretical computer science.
Other high-profile problems targeted
The repository also includes claimed progress on the Hodge conjecture, another Millennium Prize problem, proving it for a large special class of varieties in every dimension and resolving two related conjectures. Other papers claim results on the Kakeya problem in 3D and 4D, a version of Hilbert’s tenth problem over rationals, the irrationality of Catalan’s constant, a long-standing geometry conjecture in all dimensions, and problems in primes, approximation, magnets, and plasma.
Backlash from mathematicians
Rather than celebration, the release prompted a sharp backlash. Many mathematicians argue that AI firms are turning humanity’s hardest problems into a performance scoreboard while withholding the systems needed for independent replication. Several researchers said earlier private discussions with OpenAI suggested a slower, more review-friendly process, and some felt those expectations were ignored.
Dispute over priority and pressure
Tensions were already high after a separate dispute involving Navier-Stokes work. NYU mathematician Tristan Buckmaster alleged OpenAI moved aggressively after hearing another team might be close to a breakthrough, and described behavior he viewed as threatening and aimed at securing priority. The accusations fed comparisons of major AI labs to coercive power centers, though OpenAI has disputed those characterizations.
Guidelines and demands for transparency
An advisory effort linked to the Institute for Advanced Study, involving prominent mathematicians including Timothy Gowers, Martin Hairer, and Edward Witten, issued guidelines after hearing from more than 600 mathematicians. The recommendations called for avoiding private-model testing on major open problems, disclosing prompts, compute, costs, and failures, using machine-checked proofs, and storing results outside company control. Critics say OpenAI disclosed only partial information and still controls the main repository.
A broader fight over the future of mathematics
Some researchers see the papers as the start of a profound shift in mathematical work, with humans focusing more on selecting questions, directing AI systems, and interpreting results. Others warn that unverifiable claims from closed models could damage trust, distort credit, and concentrate power in a few companies. Even supporters of AI-assisted discovery say the central issue is no longer whether machines can contribute to mathematics, but who will govern that process and under what standards.
CONCLUSION
The OpenAI/Math release has turned pure mathematics into a frontline dispute over proof, credit, and access. Whether the claims survive scrutiny or not, the balance between open scientific culture and closed corporate AI systems is now under direct pressure.
Ask a question