🔍 Read the full analysis: How Far Can 722 Proofs Take OpenAI’s AI Mathematics? 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, grouped into 372 families and generated from about 4,000 problems by an unnamed, unreleased model. The company’s results include claims about major open problems, but they have not been independently confirmed as a body of work. Their value will depend on verification and whether mathematicians can extract methods others can use.
OpenAI published 722 mathematical manuscripts on Monday, presenting work by an unnamed, unreleased model across fields including number theory, geometry and theoretical computer science. The papers report results on several major open problems, but OpenAI chief executive Sam Altman said the claims have not been confirmed by outside mathematicians, leaving their reliability and mathematical value unresolved.
The manuscripts are organized into 372 families of related results and were selected from roughly 4,000 problems posed to the model. OpenAI says the average result used about three hours of ChatGPT Pro thinking compute. The collection is published under an Apache-2.0 license, and many—but not all—of its results have Lean formalizations, which can help check proofs using software.
The claims range from a proof of the Unique Games Conjecture and a proposed resolution of Hilbert’s tenth problem over the rationals to results concerning free group factors, the Hodge conjecture for CM abelian varieties and Mahler’s conjectures. One manuscript proposes a zero-free region for the Riemann zeta function to the right of Re(s) = 11/12. These are claims in the released papers, not established resolutions of the problems.
OpenAI’s repository includes 10 abridged reasoning summaries, out of 372 result families. The source report says the company selected the problems for what it considered an appropriate level of significance, with no outside group involved in that selection. OpenAI’s README cautions that some unformalized results could have issues. The Riemann-region write-up was edited by humans for readability, and the Hodge result and that write-up were exceptions to the usual process described for the collection.
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.
Proofs Need Ideas Others Can Use
The central issue is not just whether a manuscript reaches the right answer. In mathematics, a proof can matter because it provides techniques that other researchers can understand and reuse. If experts can digest a model’s reasoning and build on it, the work may lead to results well beyond the original problem. If a proof checks but offers no transferable method, its effect on the field may be narrower.
The Unique Games claim illustrates the possible stakes. Many results in theoretical computer science rely on the conjecture to establish limits on approximation algorithms. A verified resolution could affect how researchers assess what algorithms can achieve. But that consequence depends on the claim being correct and on specialists understanding its proof; publication alone does not settle either question.
The collection also tests how research communities should evaluate machine-produced mathematics. Formal verification can check that a proof follows specified rules, but it does not by itself show that the result is novel, that the formalized statement matches the problem mathematicians care about, or that the argument yields useful insight. Those judgments require careful review by people with relevant expertise.
Products Worth Considering
FLEXISPOT EN1 Electric Standing Desk, 48 x 24 in, Black
Logitech MX Master 4 for Mac, Ergonomic Bluetooth Mouse with Haptic Feedback, Ultra-Fast Scrolling, USB-C Charging, for MacOS, iPadOS, Windo
HP Sprocket Photo Booth Instant Color Printer Bundle
As an affiliate, we earn on qualifying purchases.
OpenAI’s Earlier Math Releases
This is described in the source report as OpenAI’s fourth major mathematics release this year. An earlier release in May concerned the Erdős unit-distance conjecture. Five mathematicians—Noga Alon, Thomas Bloom, Tim Gowers, Daniel Litt and Will Sawin—posted a human-verified account of the counterexample the same day. That process, in which specialists translated machine output into an argument the field could evaluate, offers one possible model for assessing the new papers.
OpenAI’s August “Ten Advances” release had a more contested result. A claimed counterexample to Connes’s rigidity conjecture was challenged within a day; the critique said the constructed groups did not meet a condition required by the conjecture. The episode shows why a result can appear compelling yet still require close examination of whether it addresses the precise mathematical statement at issue.
In September, OpenAI announced a Lean-formalized Navier–Stokes proof produced by about 10,000 concurrent agents over 88 hours, according to the source report. The announcement prompted a separate debate about the purpose of AI benchmarks in mathematics. Twenty-five Fields Medalists signed a declaration titled “A Severe Misalignment of AI in Mathematics.” Their stated concern, as reported, was that targeting famous problems without human understanding could work against the aims of mathematical research—not simply that the proof was wrong.
“A Severe Misalignment of AI in Mathematics.”
— The 25 Fields Medalists who signed the declaration
Independent Checks Still Needed
No outside verification of the full collection is established in the source material. It is not clear which of the 372 result families experts have reviewed, whether any claims will survive scrutiny, or how long that review will take. The release does not provide enough information to assess every result’s proof, novelty or relation to prior work.
Formalization status also varies: many results have Lean formalizations, but not all. A formalized argument may help establish that a particular proof is internally valid, yet reviewers still need to check that it proves the intended statement and adds something new. For unformalized papers, OpenAI itself warns of possible issues. Neither the model’s identity nor the selection process’s external oversight is specified in the material provided.
The eventual impact is also unknown. Some claims could prove false, some could be correct but difficult to use, and others may lead researchers to new ideas. The source material does not establish which category any particular manuscript belongs to beyond the examples of earlier releases.
Mathematicians Must Test the Claims
The next step is independent, result-by-result review: specialists can examine whether each argument is sound, whether it addresses the stated conjecture and whether its ideas can be made accessible to other researchers. Lean formalizations may assist with checking the parts that have been encoded, while the unformalized results will require other forms of scrutiny.
OpenAI’s earlier Erdős release suggests that a human-written, verified explanation can make machine output easier to evaluate. For this collection, the source report identifies only 10 abridged summaries, leaving open how the other result families will be presented and reviewed. No timetable for outside validation is given. Until such work is available, the manuscripts should be treated as research claims awaiting assessment, not as confirmed solutions to long-standing problems.
Key Questions
What did OpenAI release?
It published 722 mathematical manuscripts, grouped into 372 families of related results. They were generated by an unnamed model from roughly 4,000 problems, according to the source material.
Have mathematicians verified the claims?
The source material says the claims have not yet been confirmed by outside mathematicians as a body of work. Independent review of individual manuscripts remains necessary.
Do Lean formalizations prove the results are correct?
A Lean formalization can help check that an encoded proof follows formal rules. It does not alone establish that the result is novel or that it matches the conjecture researchers intended to test; many manuscripts also lack formalizations.
Why might a correct AI proof still have limited impact?
A proof may settle a question without providing methods that researchers can reuse. Its broader value will depend on whether mathematicians can understand the reasoning and build further work from it.
What happens next?
Mathematicians will need to check the manuscripts individually. The source material gives no timetable for that review or for further explanations from OpenAI.
Source: ThorstenMeyerAI.com
