🔍 Read the full analysis: OpenAI’s AI Mathematics: What Might 722 Proofs Make Possible? on ThorstenMeyerAI.com
Get office and shipping supplies delivered free — and shop member deals
- Fast, free delivery on millions of items
- Access to Prime Big Deal Days deals on October 6–7
- Prime Video, Amazon Music and more included
TL;DR
OpenAI published 722 mathematical manuscripts produced by an unnamed, unreleased model, covering results selected from roughly 4,000 problems. The company says the results have not been confirmed by outside mathematicians; their value will depend on verification and whether researchers can extract ideas others can use.
OpenAI published 722 mathematical manuscripts on Monday, reporting that an unnamed, unreleased model generated them from roughly 4,000 problems. The catalogue includes claims about major open problems, but OpenAI says the results have not been confirmed by outside mathematicians, leaving their correctness and potential impact unsettled.
The manuscripts are organized into 372 families of related results across fields including number theory, geometry, topology, operator algebras, theoretical computer science and mathematical physics. OpenAI says the average result used about three hours of ChatGPT Pro thinking compute. The papers are published under the Apache-2.0 license, and many, but not all, have Lean formalizations, which can help check mathematical arguments with software.
The catalogue includes claimed proofs of the Unique Games Conjecture, a resolution of Hilbert’s tenth problem over the rationals, and a result on whether nonabelian free group factors are isomorphic. It also includes a proposed zero-free region for the Riemann zeta function to the right of Re(s) = 11/12, a result on the Hodge conjecture for CM abelian varieties, and claims involving the Mahler conjectures. These are claims in manuscripts, not results established by independent review.
OpenAI selected the published work from the larger problem set, describing its filter as seeking an appropriate level of significance. The company supplied ten abridged reasoning summaries, rather than summaries for all 372 families. Its repository warns that some results without formalization may contain issues. The Riemann-region manuscript was edited by people for readability, according to the source material; OpenAI also identified the Riemann and Hodge results as exceptions to its standard process.
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 Will Shape the Payoff
The number of manuscripts and the prominence of some conjectures make the release a substantial claim, but volume is not the same as verified progress. Each result must be checked to establish whether its proof works and whether it addresses the mathematical statement researchers care about. A formalization can help with checking, but it does not by itself show that a result is important or that its method will be useful beyond the specific problem.
The distinction matters because mathematics gains much of its long-term value from reusable methods and explanations, not just from settling a question. If researchers can understand and build on an argument, it may open further work. If a proof checks but offers no transferable insight, its contribution may be narrower. If it fails, or proves a different claim than intended, the headline result will not stand.
One example in the source material is the Unique Games Conjecture. Many theoretical computer science results rely on it to characterize the limits of approximation algorithms. If an independent review establishes a proof, researchers could revisit conclusions built on the conjecture. Until that review happens, any downstream implications remain conditional.
As an affiliate, we earn on qualifying purchases.
A Mixed Record This Year
This is described as OpenAI’s fourth major mathematics release this year. In May, the company said its model had produced a counterexample to the Erdős unit-distance conjecture. Five mathematicians later published what they called a digested, human-verified version. That episode offers a model for how machine-generated work can become legible and useful: researchers inspect the output, translate it into a form the field can assess, and verify the argument.
An August release, called “Ten Advances,” had a more contested result. A claimed counterexample to Connes’s rigidity conjecture was challenged within a day; the critique argued that the constructed groups did not meet the conjecture’s required conditions. The source material also reports that separate machine-generated counterexamples to the same conjecture were in circulation. That history shows why independent scrutiny is necessary, even when an output appears technically detailed.
In September, OpenAI announced a Lean-formalized proof concerning finite-time blow-up for the Navier–Stokes equations, another major open problem. The source says the work used about 10,000 concurrent agents over 88 hours and was accompanied by a priority dispute over related work. Days later, 25 Fields Medalists signed a declaration criticizing the use of famous problems as AI benchmarks without sufficient human understanding. Their stated concern was about the aims and practice of mathematics, not a finding that the proof was wrong.
Which Claims Will Survive Review?
No outside verification of the 722 manuscripts is established in the source material. It is not clear which papers independent mathematicians have examined, how many proofs will withstand detailed review, or whether any statements will need correction. OpenAI’s selection criteria also leave unanswered how the company judged significance among the roughly 4,000 problems and what was excluded.
The release includes ten abridged reasoning summaries for 372 families, and the source says some manuscripts are not formalized. That makes the amount of material available for evaluation uneven. Even a correct proof may not lead to further discoveries if its core ideas are hard to interpret or do not generalize. At this stage, claims about the catalogue’s influence are predictions, not demonstrated outcomes.
Independent Checks and Research Follow-Up
The next test is independent mathematical review: researchers will need to check the arguments, compare each statement with the problem it claims to solve, and clarify any gaps or assumptions. For results with Lean formalizations, reviewers can inspect the machine-checked proof alongside the underlying definitions; unformalized work will require conventional scrutiny and may need to be rewritten or formalized.
OpenAI’s release does not establish a timetable for reviews or say which manuscripts will receive priority. The clearest evidence of lasting impact will come if mathematicians publish verified accounts, identify reusable techniques, and use those ideas in subsequent work. Until then, the 722 papers should be treated as a large collection of proposed results awaiting assessment, not as 722 established advances.
Key Questions
Did OpenAI prove 722 mathematical theorems?
OpenAI published 722 manuscripts containing claimed results. The source material says outside mathematicians have not yet confirmed them, so they should not all be described as established theorems.
What model produced the manuscripts?
The model is described as unreleased and unnamed. The materials say the average result took about three hours of ChatGPT Pro thinking compute, but do not identify the model publicly.
Are the proofs formally verified?
Many, but not all, have Lean formalizations. OpenAI’s repository warns that some unformalized results could have issues; formalization also does not settle whether a result is significant or broadly useful.
Why could a proof of the Unique Games Conjecture matter?
Many theoretical computer science results rely on the conjecture to assess the limits of approximation algorithms. If a proof is independently verified, researchers could revisit those results, but that consequence remains conditional.
What would show that the AI work has lasting value?
Researchers would need to verify the arguments and show that their methods can support further work. A result’s long-term value may depend on what mathematicians can understand and reuse, not only on whether its conclusion is correct.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
