Read the full analysis: 722 Proofs In, What Future Awaits OpenAI’s AI Mathematics? on ThorstenMeyerAI.com
TL;DR
OpenAI published 722 mathematical manuscripts attributed to an unnamed, unreleased model, covering 372 families of results drawn from about 4,000 problems. The company says the results are claims that still need outside scrutiny; whether they are correct, useful, or capable of generating new mathematics remains unsettled.
OpenAI published 722 mathematical manuscripts on Monday, presenting work by an unnamed, unreleased model across 372 families of related results. The catalogue includes claims about major open problems, but OpenAI says they have not been confirmed by outside mathematicians, leaving the validity and mathematical value of the work unsettled.
The manuscripts span number theory, geometry, operator algebras, topology, theoretical computer science and mathematical physics. OpenAI says they came from roughly 4,000 problems posed to the model, with the company selecting those it considered significant. The average result used about three hours of ChatGPT Pro thinking compute, according to the source material. The collection is published 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, a result on nonabelian free group factors, and a zero-free region for the Riemann zeta function to the right of Re(s) = 11/12. The manuscripts also claim results concerning the Hodge conjecture for CM abelian varieties and Mahler conjectures in convex geometry. These are descriptions of what the papers claim, not established solutions.
OpenAI’s repository includes Lean formalizations for many, but not all, results. Its README cautions that some unformalized results could have issues. The release also provides only ten abridged reasoning summaries for 372 families. The Riemann zero-free-region write-up and the Hodge result received exceptions to the usual procedure; the Riemann paper was edited by humans for readability.
722 Proofs, One Question — Reality Check
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.
LEAN · reported
LEAN · reported
LEAN · reported
STATUS · see repo
STATUS · see repo
STATUS · see repo
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.
When a Proof Becomes Useful
The scale and ambition of the catalogue matter, but a correct proof is not automatically a useful mathematical discovery. A result may settle a question while offering few techniques that other researchers can reuse. The field’s longer-term response will depend on whether mathematicians can verify the arguments, understand how they work, and extract ideas that support further research.
The Unique Games Conjecture illustrates the potential stakes if a claim holds up. Many results in theoretical computer science rely on the conjecture to establish limits on approximation algorithms, including work on the Goemans–Williamson algorithm for Max-Cut. A valid resolution could force researchers to revisit those conditional results. But until the proof is checked, those consequences remain hypothetical.
There is also a practical question about review capacity. Hundreds of manuscripts across specialist fields require time from researchers qualified to assess them. The number of papers does not establish how many are correct or significant, and a formalization can help check a formal proof without, by itself, showing that the mathematical ideas are valuable or broadly understood.
A Mixed Record of AI Proofs
This is OpenAI’s fourth major mathematics release this year, following projects that have produced different kinds of responses. In May, the company’s model generated a counterexample to the Erdős unit-distance conjecture. Five mathematicians—Noga Alon, Thomas Bloom, Tim Gowers, Daniel Litt and Will Sawin—then published a human-verified, digested account. That process showed how machine output can become assessable mathematical work after researchers reconstruct and check it.
An August release called “Ten Advances” included a claimed counterexample to Connes’s rigidity conjecture. A critique argued that the constructed groups did not meet a condition required by the conjecture. The episode shows why a result’s wording and hypotheses need close scrutiny: a proof may fail, or address a nearby statement rather than the one mathematicians meant.
In September, OpenAI announced a Lean-formalized proof concerning finite-time blow-up in the Navier–Stokes equations, generated through a large multi-agent effort, according to the company. The announcement prompted a dispute over research priority alongside work on forced Euler equations by Levent Alpöge and Tristan Buckmaster. Twenty-five Fields Medalists later signed a declaration titled “A Severe Misalignment of AI in Mathematics,” criticizing the use of famous problems as benchmarks when results are not accompanied by human understanding. Their objection, as described in the source material, was about the purpose and practice of mathematics—not a finding that the Navier–Stokes proof was wrong.
“A Severe Misalignment of AI in Mathematics.”
— The 25 Fields Medalists who signed “A Severe Misalignment of AI in Mathematics”
Independent Checks Still Pending
No outside confirmation is established for the catalogue as a whole, and the source material does not provide independent verdicts on each of the 372 families. It is also unclear how many researchers are currently reviewing the papers, which claims will withstand scrutiny, and how long that process will take.
The selection process was controlled by OpenAI: the company filtered about 4,000 problems for what it judged an appropriate level of significance. The release does not establish that the selection was independently assessed or explain in detail how the model’s full reasoning can be evaluated from the ten abridged summaries. For claims lacking Lean formalizations, the company’s own warning leaves additional room for errors.
Even if particular results prove correct, their downstream value is unknown. Researchers will need to determine whether the proofs reveal reusable methods, simply settle statements, or require substantial correction. The available information does not establish that every headline result is correct, that the model has solved all the named problems, or that the manuscripts will lead to new discoveries.
Mathematicians Must Test the Claims
The next step is paper-by-paper review by specialists, including checking the statements against the precise conjectures, examining the arguments, and testing formalized proofs where available. Researchers may also produce clearer, human-readable accounts that identify the central ideas and make the work easier to assess.
For readers, the most informative developments will be independent verification or published critiques of specific manuscripts, not the size of the catalogue alone. OpenAI has released the papers, but the source material gives no timetable for external assessments or for further company updates. Until those checks appear, the 722 manuscripts should be treated as a body of mathematical claims awaiting evaluation.
Key Questions
What did OpenAI release?
OpenAI published 722 mathematical manuscripts attributed to an unnamed model. They are grouped into 372 families and were selected from work on roughly 4,000 problems, according to the source material.
Has the mathematics been independently confirmed?
Not as a collection. OpenAI says the results are claims that have not yet been confirmed by outside mathematicians, and its repository warns that some unformalized results could have issues.
What major problems do the manuscripts claim to address?
The papers include claims involving the Unique Games Conjecture, Hilbert’s tenth problem over the rationals, nonabelian free group factors, the Riemann zeta function, the Hodge conjecture for CM abelian varieties, and Mahler conjectures. These remain claims pending independent review.
Why does a correct proof not guarantee a major discovery?
A proof can settle a question without giving mathematicians methods they can reuse. Its broader value depends on whether researchers can understand the argument and build further work from it.
What happens next?
Mathematicians will need to examine individual manuscripts, verify their reasoning and assess whether they establish the stated results. No timetable for those reviews is specified in the source material.
Source: ThorstenMeyerAI.com