🔍 Read the full analysis: Where Could OpenAI’s AI Mathematics Lead After 722 Proofs? on ThorstenMeyerAI.com
Get hardware and tech essentials 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 produced by an unnamed, unreleased model, covering claims that include solutions to long-standing open problems. The claims have not been confirmed by outside mathematicians, and the collection’s significance will depend on independent checking and whether researchers can extract reusable ideas.
OpenAI published 722 mathematical manuscripts on Monday, reporting that an unnamed, unreleased model generated them while working on roughly 4,000 problems. The collection includes claims about major open questions, but outside mathematicians have not confirmed the results; whether any will lead to further discoveries remains unsettled.
The manuscripts are organized into 372 families of related results and cover fields including number theory, geometry, operator algebras, topology, theoretical computer science and mathematical physics. OpenAI says the average result took about three hours of ChatGPT Pro thinking compute. The materials are published under an Apache-2.0 license, and the repository includes Lean formalizations for many, but not all, of the results.
The catalogue includes claims concerning the Unique Games Conjecture, Hilbert’s tenth problem over the rationals, the isomorphism of nonabelian free group factors, a zero-free region for the Riemann zeta function to the right of Re(s) = 11/12, the Hodge conjecture for CM abelian varieties, and the Mahler conjectures in convex geometry. These are claims in the released manuscripts, not independently established breakthroughs. OpenAI’s repository warns that some results without formalizations could have issues.
OpenAI says it selected the published work from about 4,000 problems for what it judged to be an appropriate level of significance. The selection was made internally, and the public material includes only ten abridged reasoning summaries across the 372 families. The Riemann zero-free-region write-up and the Hodge result received exceptions to the usual process; the Riemann manuscript was edited by humans for readability.
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.
From Machine Proofs to Mathematical Tools
The scale of the release is striking, but a correct answer alone does not show that a result will change mathematics. A proof can settle a question without giving researchers techniques they can reuse. The longer-term test is whether mathematicians can verify the arguments, understand their methods and build on them.
The Unique Games claim illustrates why verification could matter beyond mathematics. The conjecture is connected to theoretical computer science results about the limits of approximation algorithms. If a proof is correct and accepted, researchers would need to reconsider conclusions that rely on the conjecture, including claims about how well certain optimization problems can be approximated. That impact is conditional: the manuscript first needs careful independent scrutiny.
The collection could also add pressure to a debate over what counts as progress when AI systems solve famous problems. Some mathematicians value a proof for the concepts and explanations it contributes, not just for closing a question. If the model’s work is difficult to interpret, the field may gain verified answers without gaining a clear account of why they are true or what else the methods can do.
There is also a practical question of review capacity. Hundreds of manuscripts across multiple specialties require expertise and time to assess. Formal verification can check that a formalized argument follows specified rules, but it does not by itself establish that the formalized statement matches the intended conjecture, or that the result has broader mathematical value. Human review remains central to those judgments.
Products Worth Considering

Ledger Stax Crypto Wallet with Touchscreen

Trezor Safe 7 Crypto Hardware Wallet with Bluetooth and Color Touchscreen

Ledger Nano X Bluetooth Crypto Wallet
As an affiliate, we earn on qualifying purchases.
Earlier Releases Offer Mixed Evidence
This is OpenAI’s fourth major mathematics release this year, following projects that show both potential and unresolved problems. In May, an OpenAI model produced a counterexample to the Erdős unit-distance conjecture. Five mathematicians—Noga Alon, Thomas Bloom, Tim Gowers, Daniel Litt and Will Sawin—posted a human-verified account on the same day. That episode is a model for how machine-generated work can become usable: researchers distilled and checked the output.
An August release, called “Ten Advances,” had a more complicated reception. A claimed counterexample to Connes’s rigidity conjecture was challenged within a day. The critique said the groups constructed in the argument did not meet a condition required by the conjecture. The dispute highlights a recurring verification risk: an argument may address a related statement rather than the precise problem researchers intend to resolve.
In September, OpenAI announced a Lean-formalized proof concerning finite-time blow-up in the Navier–Stokes equations, produced, the company said, by about 10,000 concurrent agents over 88 hours. The work became part of a debate about research priorities and the use of famous problems as AI benchmarks. A declaration signed by 25 Fields Medalists criticized that approach, arguing that it could conflict with mathematical understanding. Their objection was about the purpose and practice of AI research, not a finding that the proof was incorrect.
These episodes provide useful precedent, but they do not establish whether the new manuscripts are sound. Each result needs to be assessed on its own statement, proof and supporting material. The Erdős episode shows a route to validation; the Connes dispute shows why scrutiny matters.
“A Severe Misalignment of AI in Mathematics.”
— The declaration signed by 25 Fields Medalists
Independent Checks Remain Pending
It is not yet clear which of the 372 result families will withstand independent review, how long that review will take, or whether researchers outside OpenAI can reproduce the work. The source material does not report that external mathematicians have validated the headline claims. The company’s selection from roughly 4,000 problems also means the public catalogue is a curated subset, not a complete account of the model’s attempts.
Formalizations are available for many results, but not all. Even where a Lean proof is provided, questions remain about whether the formal statement accurately captures the intended mathematical claim and what insights the proof offers. OpenAI has supplied only ten abridged reasoning summaries, leaving much of the process behind the collection difficult for readers to assess from the summaries alone.
The release does not establish that the results are wrong, nor that they are accepted discoveries. Verification, interpretation and field-wide significance are separate questions, and the available information does not settle them.
Mathematicians Must Test the Claims
The next step is independent, specialist review of individual manuscripts. Researchers will need to check both the statements and the arguments, with formal proof files offering one avenue for scrutiny where they exist. For results that survive, mathematicians can then ask whether the methods reveal reusable ideas or support further work.
OpenAI has not provided a public timetable for reviews or a schedule for further releases in the source material. The clearest near-term signal will be detailed assessments by mathematicians, including corrections, confirmations or explanations of how a result fits into existing theory. Until then, the 722 manuscripts are a large set of published claims awaiting evaluation, not 722 established advances.
Key Questions
What did OpenAI publish?
OpenAI published 722 mathematical manuscripts, grouped into 372 families and generated by an unnamed, unreleased model, according to the company’s materials.
Have outside mathematicians verified the results?
The source material says the claims have not yet been confirmed by outside mathematicians. Publication and internal selection do not amount to independent verification.
Which major problems are covered?
The manuscripts include claims about the Unique Games Conjecture, Hilbert’s tenth problem over the rationals, the Hodge conjecture for CM abelian varieties and other long-standing questions. Their correctness remains to be assessed.
Why does the Unique Games claim matter?
The conjecture is used in theoretical computer science to support results about the limits of approximation algorithms. If a proof is verified, researchers may need to revisit work that depends on it; that consequence is conditional on the claim being correct.
Does a formalized proof settle whether a result is important?
No. A Lean formalization can help check that an argument follows formal rules, but researchers still need to confirm that it proves the intended statement and decide whether its methods contribute reusable understanding.
Source: ThorstenMeyerAI.com
