🔍 Read the full analysis: 722 Proofs Later, OpenAI’s AI Mathematics Still Faces A Question on ThorstenMeyerAI.com
Get smart everyday buys 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 published 722 mathematical manuscripts in 372 families, generated by an unnamed model from roughly 4,000 problems. The results include claims about prominent open problems, but OpenAI says outside mathematicians have not confirmed them; whether the work yields reusable ideas remains unknown.
OpenAI published 722 mathematical manuscripts on Monday, presenting work from an unreleased, unnamed model across 372 families of related results. The collection includes claims about major open problems, but the company and its chief executive Sam Altman have cautioned that outside mathematicians have not yet confirmed the results.
According to OpenAI’s post and its GitHub repository, the manuscripts span number theory, geometry, operator algebras, topology, theoretical computer science and mathematical physics. They came from roughly 4,000 problems posed to the model. OpenAI says it selected problems it judged to have an appropriate level of significance; the selection was made internally. The average result used about three hours of ChatGPT Pro reasoning compute.
The catalogue includes claims that the model proved the Unique Games Conjecture, resolved Hilbert’s tenth problem over the rational numbers, and showed that all nonabelian free group factors are isomorphic. Other manuscripts claim a zero-free region for the Riemann zeta function to the right of Re(s) = 11/12, a result about the Hodge conjecture for CM abelian varieties, and work on the Mahler conjectures. These are claims in manuscripts, not results independently established by the release itself.
OpenAI released Lean formalizations for many, but not all, of the results. Its repository warns that “some of the unformalized results could have issues.” The company also supplied ten shortened reasoning summaries, rather than summaries for all 372 families. The Riemann zero-free-region write-up was edited by humans for readability, and OpenAI identified the Riemann and Hodge results as exceptions to its standard procedure.
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 release matters because some of its claims, if verified, could affect fields well beyond the problems themselves. The Unique Games Conjecture, for instance, underpins a body of theoretical computer science research on the limits of approximation algorithms. A proof could prompt researchers to revisit conclusions that rely on the conjecture. But that consequence depends first on mathematicians checking that the result is correct and proves the statement at issue.
Verification alone would not settle the question of the work’s broader value. In mathematics, a proof can be important not only because it answers a question, but because its methods become tools others can apply. OpenAI’s manuscripts will have greater downstream impact if researchers can extract and explain reusable ideas. If the proofs are correct but offer no usable insight, they may settle questions without changing how the field develops.
That distinction also makes the volume difficult to interpret. 722 manuscripts are not 722 independently verified discoveries. Each family needs expert review, and the catalogue’s internal selection process does not show how external specialists would rank the work or judge its importance.
As an affiliate, we earn on qualifying purchases.
OpenAI’s Earlier Math Releases
This is OpenAI’s fourth major mathematics release this year, according to the source material. Its earlier results show why claims need to be evaluated individually. In May, the company reported that its model found a counterexample to the Erdős unit-distance conjecture. Five mathematicians—Noga Alon, Thomas Bloom, Tim Gowers, Daniel Litt and Will Sawin—posted what they described as a digested, human-verified version on the same day. That process turned machine output into work specialists could evaluate.
OpenAI’s August “Ten Advances” release had a more contested result. A claimed counterexample to Connes’s rigidity conjecture was challenged within a day; a critique argued that the constructed groups did not meet a condition required by the conjecture. In September, OpenAI announced a Lean-formalized proof concerning finite-time blow-up in the Navier–Stokes equations, generated using about 10,000 concurrent agents over 88 hours. The release prompted a dispute about research priorities, rather than establishing that the proof was wrong.
Three days after the Navier–Stokes announcement, 25 Fields Medalists, including Terence Tao, Peter Scholze and Maryna Viazovska, signed a declaration titled “A Severe Misalignment of AI in Mathematics.” Their concern, as described in the source material, was that using famous problems as benchmarks without human understanding could work against mathematics’ broader aims. The current release again puts verification and understanding at the center of the discussion.
““A Severe Misalignment of AI in Mathematics.””
— The 25 Fields Medalists who signed the September declaration
formal verification tools for mathematicians
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Which Claims Will Survive Review
Independent verification remains the central unknown. The release does not establish which of the 372 families have been checked by outside specialists, how many proofs withstand scrutiny, or whether any need substantial correction. Lean formalizations can support verification for some work, but the collection includes results without them, and OpenAI itself cautions that those may have issues.
It is also unclear how the company’s significance filter shaped the catalogue. OpenAI selected from roughly 4,000 posed problems, but the source material does not describe an external review of that selection or provide enough detail to compare the chosen set with the problems that were omitted. The ten abridged reasoning summaries cover only a small share of the 372 families, limiting what readers can infer about the full body of work.
Even if particular claims prove correct, their lasting mathematical value cannot yet be assessed. Researchers will need time to understand the arguments and determine whether the techniques can be reused. The current record does not show which manuscripts, if any, will lead to new lines of research.
AI research notebooks for mathematics
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Checks and Follow-Up
The next step is expert scrutiny of individual manuscripts, including checking whether each proof is valid and whether it establishes the stated result. Formalized proofs may offer a route for machine-assisted checking, while unformalized arguments will require close mathematical review. The release does not set a schedule for that process or identify when evaluations will be complete.
Readers should watch for corrections, independent write-ups, and responses from specialists working in the relevant fields. The May Erdős result offers one example of a process in which mathematicians produced a human-verified, digestible account of machine-generated work. Whether comparable accounts emerge for the new manuscripts remains to be seen.
For now, the catalogue is best treated as a large set of research claims rather than a completed tally of discoveries. The meaningful test will be which results survive scrutiny and what other mathematicians can learn from them.
mathematical problem solving software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did OpenAI release?
OpenAI published 722 mathematical manuscripts, grouped into 372 families and generated by an unnamed, unreleased model from roughly 4,000 problems.
Have outside mathematicians confirmed the results?
Not according to the source material. OpenAI and Sam Altman have framed the manuscripts as claims awaiting outside confirmation; the release does not establish that the collection has been independently verified.
Are all the manuscripts formally verified?
No. OpenAI provided Lean formalizations for many, but not all, of the results. Its repository warns that some unformalized results could have issues.
Why does the Unique Games claim matter?
The conjecture is used in theoretical computer science research on the limits of approximation algorithms. If a proof is verified, researchers may need to revisit work that assumes the conjecture; that impact is conditional on confirmation.
What would make the release valuable beyond solving problems?
Researchers would need to understand the proofs and identify methods they can reuse. The manuscripts’ longer-term significance depends not just on whether claims are correct, but on whether the work helps generate further mathematical insight.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
