🔍 Read the full analysis: OpenAI’s AI Mathematics: 722 Proofs, And The Question Of What Follows on ThorstenMeyerAI.com
Get monitors, keyboards and dev gear 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, grouping them into 372 families drawn from about 4,000 problems. The manuscripts include major claimed results, but outside mathematicians have not confirmed them; the release does not yet show whether the work will yield reusable mathematical ideas.
OpenAI has published 722 mathematical manuscripts produced by an unnamed model that the company has not released, presenting claims that include results on several prominent open problems. The papers are arranged into 372 families and were selected from work on roughly 4,000 problems, but OpenAI chief executive Sam Altman has cautioned that mathematicians outside the company have not confirmed the claims.
OpenAI’s post and accompanying GitHub repository describe manuscripts spanning number theory, geometry, topology, operator algebras, theoretical computer science and mathematical physics. The company says the average result used about three hours of ChatGPT Pro reasoning compute. The collection is published under the Apache-2.0 license. Lean formalizations are included for many, though not all, of the results.
The catalogue includes claimed proofs or resolutions involving the Unique Games Conjecture, Hilbert’s tenth problem over the rationals, the Hodge conjecture for CM abelian varieties and the Mahler conjectures in convex geometry. Other manuscripts claim a result about nonabelian free group factors and a zero-free region for the Riemann zeta function to the right of Re(s) = 11/12. These descriptions report what the manuscripts claim; they are not evidence that the results have been accepted by the mathematical community.
OpenAI says its selection process started with roughly 4,000 problems and filtered them for what the company considered an appropriate level of significance. The repository provides ten abridged reasoning summaries for the 372 families, rather than a summary for every family. OpenAI’s README warns that some results without formal verification could have problems. The Riemann-region write-up was edited by people for readability, and OpenAI identifies the Riemann and Hodge manuscripts as exceptions to its usual 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 central question is not only whether a proof is correct, but whether mathematicians can understand its method and use it. A result can settle a conjecture without supplying a technique that advances the surrounding field. The collection therefore matters both as a set of testable claims and as a test of whether AI systems can contribute ideas that people can inspect, explain and build on.
The Unique Games Conjecture illustrates the potential stakes if the claim holds up. Many results in theoretical computer science rely on the conjecture to establish limits on approximation algorithms. A verified proof could affect how researchers interpret those conditional results. But that consequence is conditional: the manuscript must first withstand scrutiny, and its proof must be understood well enough to show what follows from it.
OpenAI’s earlier mathematics releases offer contrasting precedents. In May, a model produced a counterexample to the Erdős unit-distance conjecture; five mathematicians then posted a human-verified account. That process turned machine output into work the field could evaluate. By contrast, a claimed counterexample to Connes’s rigidity conjecture in an August release was disputed over whether the constructed groups met the conjecture’s conditions. The examples show why a large catalogue is not, by itself, a measure of validated progress.
As an affiliate, we earn on qualifying purchases.
A Record of Promising and Disputed Results
The new collection follows several public mathematics announcements by OpenAI this year. The May Erdős result was rapidly translated into a human-verified explanation. An August release called “Ten Advances” included a Connes-related claim that drew a prompt challenge over the statement being addressed. Those episodes make independent checking and clear descriptions of the exact theorem especially relevant to the latest release.
In September, OpenAI announced a Lean-formalized proof concerning finite-time blow-up in the Navier–Stokes equations, produced through a large multi-agent effort. That announcement also prompted discussion about how AI mathematics should be judged. A declaration signed by 25 Fields Medalists criticized using famous problems as benchmarks without adequate human understanding; the stated concern was about mathematical practice and comprehension, not simply whether a particular proof was false.
Mathematical proofs have often mattered because their methods travel beyond the original question. The source material contrasts that kind of influence with computer-assisted work such as the Four Colour Theorem, which settled a problem through a large case analysis that people could not survey in the same way as a conventional proof. The comparison is not a verdict on OpenAI’s new papers; it explains why researchers may ask what a proof teaches as well as whether it checks out.
mathematical proof verification software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Checks Still Needed
The main unresolved issue is whether the claimed results are correct. No independent confirmation is reported for the new catalogue, and the source material does not identify outside mathematicians who have completed a review of its 372 families. Formal Lean verification may make some claims easier to check, but many results are not formalized, and formalization alone does not explain a result’s broader significance.
Other details also remain open: OpenAI has not named or released the model, and the company has not published the complete reasoning record for every family. It is unclear how many manuscripts will survive peer review, how quickly mathematicians can assess them, or whether the proofs contain methods that can be reused. The selection was made by OpenAI, so the collection does not establish how the system performed across all problems it was given.
formal proof assistants for mathematicians
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Mathematicians Test the Claims
The next milestone is outside scrutiny: mathematicians will need to check each manuscript’s exact statement, assumptions and proof, with particular attention to results that are not formally verified. For claims that do check out, researchers may then try to produce clearer accounts and identify whether the methods connect to existing work or lead to new results.
That process will take place result by result, rather than through a single verdict on the entire release. The available material does not set a review timetable or identify an independent body coordinating validation. Until that work is done, the 722 manuscripts are a substantial body of AI-generated mathematical claims, not 722 established discoveries.
AI tools for solving open math problems
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did OpenAI publish?
OpenAI published 722 mathematical manuscripts, organized into 372 families and drawn from roughly 4,000 problems posed to an unnamed model.
Have outside mathematicians confirmed the results?
Not according to the source material. Altman described the results as claims not yet confirmed by outside mathematicians, and the repository warns that some unformalized work could have issues.
Does the release include formal proofs?
Lean formalizations are available for many, but not all, of the results. OpenAI’s repository cautions that some unformalized results may contain problems.
Why could a correct proof still have limited impact?
A proof may settle a question without revealing a method that researchers can reuse. Its wider contribution depends in part on whether mathematicians can understand and build on the reasoning.
What happens next?
Mathematicians need to check the manuscripts independently. The release does not specify a review schedule, and the number of claims that will be validated remains unknown.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
