🔍 Read the full analysis: Will 722 Proofs Help OpenAI’s AI Mathematics Make Headway? on ThorstenMeyerAI.com
Get audio and creator gear delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
OpenAI has published 722 mathematical manuscripts, arranged in 372 families, generated after its model was given roughly 4,000 problems. The collection includes claims about major open questions, but the results have not been independently confirmed as a body; their significance will depend on verification and whether mathematicians can use the methods.
OpenAI published 722 mathematical manuscripts on Monday, presenting results produced by an unnamed, unreleased model after it was given roughly 4,000 problems. The collection includes claims about longstanding open questions, but OpenAI chief executive Sam Altman said the results have not yet been confirmed by outside mathematicians, leaving their reliability and potential influence unsettled.
The manuscripts are grouped into 372 families of related results and span areas including number theory, geometry, operator algebras, topology, theoretical computer science and mathematical physics. OpenAI said the average result used about three hours of ChatGPT Pro thinking compute. The repository makes the work available under an Apache-2.0 license.
Among the claims are a proof of the Unique Games Conjecture, a resolution of Hilbert’s tenth problem over the rationals, and results concerning free group factors, the Riemann zeta function and the Hodge conjecture for CM abelian varieties. These are claims in manuscripts, not findings independently established by the mathematical community. OpenAI’s repository README cautions that some results without formal verification could have issues.
Many, but not all, results have Lean formalizations, a way to encode proofs so that software can check their logical steps. OpenAI also provided 10 abridged reasoning summaries, rather than one for every family. The company selected the published work from about 4,000 problems; no outside group participated in that selection process, according to the source material.
722 proofs, one question: will any of OpenAI’s AI mathematics actually lead anywhere?
An unreleased, unnamed model produced claimed proofs of results that would each define a career. Sam Altman calls them “claims not yet confirmed by outside mathematicians.” The real question isn’t whether it’s impressive. It’s whether answers nobody understands become discoveries anyone can build on.
Same day: Alon, Bloom, Gowers, Litt, Sawin post a digested, human-verified version. The model for success.
Connes rigidity counterexample challenged within a day — constructed groups fail the required condition. Three rival machine “counterexamples” from different labs now circulate.
~10,000 agents, 88 hours, est. ~$22M at retail. Priority dispute; 25 Fields Medalists sign “A Severe Misalignment” — not saying it’s wrong, saying it’s not understood.
Altman now hedges at announcement — a shift from September. Verification has barely started.
Humans extract the technique, write it up, build on it. This is where downstream discovery comes from.
The question is answered; nobody learns anything reusable. Closes a door without opening a field.
The proof breaks, or proves a statement that doesn’t match the conjecture as mathematicians mean it.
The Unique Games Conjecture is the clearest case. Results like the optimality of Goemans–Williamson for Max-Cut are proved assuming UGC. A correct proof converts them all — no understanding required. A zero-free strip for zeta works the same way for prime-distribution results. Free group factors, Kadison, Mahler would redirect whole programmes — but how depends on the method, which means digestion.
Technology. A Navier–Stokes blow-up proof doesn’t change how anyone designs aircraft; engineering turbulence models never depended on the answer. Near-term consequences are mathematical, not industrial. “AI will cure cancer next” skips several steps.
“Verification abundance, adjudication scarcity” — making proof-checking cheap doesn’t reduce the burden of deciding what’s true and what matters. 722 manuscripts land on a review system built for a trickle, filtered by a selection nobody outside OpenAI made.
Humans re-deriving results, like Alon–Gowers et al. in May
Other people’s work building on these manuscripts
How many unformalized results survive expert checking
Do the Lean statements match the real conjectures?
Do any survive peer review?
Some of it, yes — where a literature is waiting (UGC), a correct proof pays off immediately; where a proof carries a new technique humans digest, it can open a field. Most of it, probably not on its own: at 722 manuscripts with 10 reasoning summaries, the Four Colour pattern is the likely default unless mathematicians are funded and given time. And some will be wrong — OpenAI says so itself. It’s an industry pattern, not one company’s: the forced-Euler result came from an Anthropic researcher, and rival machine-generated Connes “counterexamples” circulate from different labs. The proofs arrived this week. The discoveries, if they come, will arrive at the speed of human understanding.
Verification Before Mathematical Impact
The immediate question is not simply whether the model can produce a proof-shaped document. Mathematicians must determine whether each argument is valid and whether it establishes the exact statement claimed. Formal checking can help test logic, but it does not by itself establish that a result is important, correctly framed, or useful to researchers.
Even a correct proof may have very different effects. A result might give researchers a method they can adapt, settle a question without producing reusable techniques, or turn out to prove something narrower than the original conjecture. The source material describes the Erdős unit-distance result as a case in which mathematicians produced a digested, human-verified account of model output. That step made the work easier for the field to evaluate and build on.
The scale of this release matters because each of its 372 families needs scrutiny, and a handful of headline claims cannot establish the quality of the full catalogue. If the work holds up, it could affect research across multiple fields. If claims fail or remain hard to interpret, the volume may add verification demands without creating new mathematical tools.
formal proof verification software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Earlier Releases Offer Caution
This is OpenAI’s fourth major mathematics release this year, according to the source material. In May, the company’s model produced a counterexample to the Erdős unit-distance conjecture. Five mathematicians posted what they called a digested, human-verified version the same day, providing an example of how machine-generated work can be recast for independent assessment.
OpenAI’s August release, called “Ten Advances,” had a more complicated reception. The source material says a claimed counterexample to Connes’s rigidity conjecture was challenged within a day on the grounds that the constructed groups did not meet the conjecture’s required conditions. That episode illustrates why a claimed proof or counterexample must be checked against the precise mathematical statement.
In September, OpenAI announced a Lean-formalized result concerning finite-time blow-up in the Navier–Stokes equations. The announcement prompted a separate debate about the aims of AI mathematics. A declaration signed by 25 Fields Medalists criticized using famous problems as benchmarks when the work does not produce human understanding. Their objection, as described in the source material, was not that the proof was wrong but that evaluation by problem-solving alone can overlook mathematics’ broader purpose.
Which Claims Will Survive Review?
The release does not establish that the headline results are correct. Outside mathematicians have not confirmed the catalogue, and the source material does not report a completed, independent review of the 372 families. It is also unclear which manuscripts have complete formal verification and how much of the supporting work other researchers can reproduce.
OpenAI’s selection process leaves another limit: the company chose which results from roughly 4,000 problems were included, and no external group took part in that filtering. The provided summaries cover only 10 of 372 families. Readers therefore do not yet have a comparable concise account of the reasoning across the collection, or a clear basis for judging how representative the published manuscripts are.
Nor is it known which correct results, if any, will yield techniques others can reuse. A proof can settle a question without changing the surrounding field. The contribution of this release will depend not only on whether individual arguments check out, but on what researchers can understand and develop from them.
Independent Checks Will Settle the Claims
The next step is independent mathematical review of individual manuscripts: checking the arguments, confirming that conclusions match the stated problems, and examining formalizations where available. That work will take different amounts of time depending on the field and the complexity of each result; the source material gives no timetable for a full assessment.
Researchers will also need to identify whether any proofs contain reusable methods and publish explanations that others can evaluate. The model’s identity and technical details remain undisclosed in the supplied material, so it is not clear when OpenAI may provide more information about how it produced the manuscripts. Until outside mathematicians complete that work, the 722 papers should be treated as a large collection of claims awaiting scrutiny, not as 722 established advances.
Key Questions
What did OpenAI publish?
OpenAI published 722 mathematical manuscripts grouped into 372 families. They were generated by an unnamed model after it was given roughly 4,000 problems.
Have mathematicians verified the results?
Not as a collection. Altman said the claims have not been confirmed by outside mathematicians, and OpenAI’s README warns that some unformalized results could have issues.
What is a Lean formalization?
Lean is a proof assistant that can check formalized mathematical arguments for logical consistency. Many, but not all, of the released results have Lean formalizations.
Why does it matter if a proof is correct?
Correctness can settle a mathematical question, but a proof’s broader value may depend on whether its methods help researchers solve other problems. Independent review is also needed to confirm that it proves the intended statement.
When will the claims be settled?
The source material provides no review timetable. The claims will need assessment by mathematicians, and the time required may vary across the 372 families.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
