ResearchPapers, evidence & method
OpenAI releases about 720 AI-generated math manuscripts, including claimed results on the Kakeya, Hodge and Riemann problems
OpenAI posted roughly 720 manuscripts in 372 result families from an unnamed internal model, including a claimed four-dimensional Kakeya proof. Only about 42% of top-line results have Lean proofs, and mathematicians expect verification to take years.
Reach: not measured yet
AT Pulse research deskPublished · 5 min read · AI-assisted analysisOpenAI published an Apache-2.0 repository of manuscripts grouped into 372 families. The repository counts 719 manuscripts, while other reports say 722. They came from about 4,000 attempted problems, and the average result used roughly three hours of ChatGPT Pro-equivalent thinking compute. According to OpenAI, the results include a four-dimensional Kakeya proof, a special case of the Hodge conjecture and progress toward the Riemann hypothesis. [2] [9] [10] [1]
Verification is thin. About 42% of top-line results are formalized in Lean, and some Lean statements do not cleanly match what the papers claim. Within two days, three papers were withdrawn over a sign error. Because the model, the prompts and the failed attempts are undisclosed, outsiders cannot measure the success rate. [1] [2]
Expert reactions are early and mixed. One Stanford mathematician put tens of results at Fields Medal-contender level, while others describe clever uses of existing techniques and write-ups that are hard to follow. An MIT researcher wants the results treated as unverified until they can be replicated. A human audit found that OpenAI's earlier batch of ten results held up, but it covered only that batch. [1] [9] [3]
Mathematicians are now debating norms for releasing AI-generated results. A new advisory group called the release a first step and declined to endorse it; it is pressing for formal proofs and disclosed models and prompts. Another group urged mathematicians to stop collaborating with OpenAI. The debate follows an earlier credit dispute over OpenAI's Navier–Stokes claim. [7] [8] [4] [5]
the results were cheap to generate but are slow to check, and some research programs report being overtaken.
Read the full assessment
Implication: verification capacity and disclosure norms, rather than generation, may decide whether AI-produced science is trusted, in mathematics and beyond.
Executive brief
On roughly October 6, 2026, OpenAI published about 720 AI-generated manuscripts grouped into 372 "families" of results. According to OpenAI, they include progress toward the Riemann hypothesis, a special case of the Hodge conjecture, and a proof of the four-dimensional Kakeya conjecture. Only about 42% of the top-line results come with Lean proofs that a computer can check. The openai/math repository says the results sit at "different stages of verification." More than three dozen mathematicians interviewed by The Verge said that sorting the real advances from the "slop" could take years. Some research groups say their entire programs have already been overtaken.
What changed and event timeline
First batch of ten results
OpenAI announced ten mathematical results. A later found no confirmed substantive error left in any principal result, though review depth varied and some results were never independently reproduced.
Navier–Stokes claim and credit dispute
OpenAI claimed to have resolved the Navier–Stokes Millennium Prize problem. NYU's Tristan Buckmaster said OpenAI used his approach and pressured him, and OpenAI denied copying.
More detail
AGMAI formed
A nine-member advisory group, the Advisory Group on Mathematics and Artificial Intelligence (AGMAI), was announced on Terence Tao's blog to advise on responsible release. Its first recommendation was that labs stop testing advanced problems on proprietary models.
The mass release
The repository went live under an Apache-2.0 license with 719–722 manuscripts in 372 families. AGMAI called the release "a first step" and explicitly said it was not endorsing it.
Corrections and backlash
The repository's change log showed revisions to more than a dozen manuscripts and three papers withdrawn over a "sign error." A group called AHM urged mathematicians to stop collaborating with OpenAI.
Capabilities and access
- Model: an "unreleased internal OpenAI model." OpenAI has not named it (openai/math).
- Compute: about 4,000 problems were attempted. The average result used the equivalent of about three hours of ChatGPT Pro thinking compute.
- Access: OpenAI says it is "working to responsibly release" the model but has given no date.
Read the full section
- Model: an "unreleased internal OpenAI model." OpenAI has not named it (openai/math).
- Compute: about 4,000 problems were attempted. The average result used the equivalent of about three hours of ChatGPT Pro thinking compute.
- Access: OpenAI says it is "working to responsibly release" the model but has given no date. It did not publish the prompts or the full list of problems attempted (The Verge).
- Engadget's description: an unreleased "pioneer" model (Engadget).
Technical analysis for researchers and developers
- Structure: each family groups a principal result with companion arguments, consequences or alternative proofs, and is tagged by discipline (openai/math).
- Verification: about 42% of top-line results are formalized in Lean.
- Version history: corrections are added as new versions, and earlier versions stay accessible.
Read the full section
- Structure: each family groups a principal result with companion arguments, consequences or alternative proofs, and is tagged by discipline (openai/math).
- Verification: about 42% of top-line results are formalized in Lean. Mathematicians say some Lean statements do not map neatly onto what the manuscripts claim, so the formal statement has to be checked against the paper (The Verge).
- Version history: corrections are added as new versions, and earlier versions stay accessible.
- Reproducibility: with no named model, no prompts and no record of failed attempts, outsiders cannot measure the success rate or the selection bias. The 4,000 attempted problems were filtered for significance before inclusion.
Claims and evidence
- Major problems: the Kakeya, Riemann-adjacent and Hodge results are OpenAI's claims (Unite.AI). Stanford's Jared Lichtman put "tens" of results at Fields Medal–contender level (The Verge).
- Replication: MIT's Andrew Sutherland says claims should be treated as unverified until the model is released and others can replicate the results (Engadget).
- Counts disagree: the repository says 719 manuscripts. Engadget, AGMAI and Unite.AI report 722.
Read the full section
- Major problems: the Kakeya, Riemann-adjacent and Hodge results are OpenAI's claims (Unite.AI). Stanford's Jared Lichtman put "tens" of results at Fields Medal–contender level (The Verge). That is an expert's first impression, not a completed review.
- Replication: MIT's Andrew Sutherland says claims should be treated as unverified until the model is released and others can replicate the results (Engadget).
- Counts disagree: the repository says 719 manuscripts. Engadget, AGMAI and Unite.AI report 722.
- Prior track record: the August batch held up under a human audit. That audit covers only the earlier ten results, not this release.
Context and prior work
- Formal verification: Lean has become the main tool for checking AI-generated proofs.
- Competing work: Buckmaster and Anthropic researcher Levent Alpöge posted Lean-verified blow-up proofs for related fluid equations one day before OpenAI's Navier–Stokes claim (NYU Shanghai RITS).
- Fields Medal: Hong Wang's 2026 Fields Medal recognized work that included the three-dimensional Kakeya conjecture. The four-dimensional case is now among OpenAI's claims (The Verge).
Read the full section
- Formal verification: Lean has become the main tool for checking AI-generated proofs. The arXiv audit argues that confidence should combine formal checking, human reconstruction, independent use of the result and a public record (arXiv).
- Competing work: Buckmaster and Anthropic researcher Levent Alpöge posted Lean-verified blow-up proofs for related fluid equations one day before OpenAI's Navier–Stokes claim (NYU Shanghai RITS).
- Fields Medal: Hong Wang's 2026 Fields Medal recognized work that included the three-dimensional Kakeya conjecture. The four-dimensional case is now among OpenAI's claims (The Verge).
Limitations, safety and contested findings
- Retractions: three papers were withdrawn over a sign error that invalidated an argument.
- Write-ups and credit: mathematicians report write-ups that are hard to follow, unusually short and with thin bibliographies, which renews concerns about attribution.
- Methods: Lichtman and UConn's Álvaro Lozano-Robledo say the results use existing techniques in ingenious ways rather than new methods (The Verge).
Read the full section
- Retractions: three papers were withdrawn over a sign error that invalidated an argument.
- Write-ups and credit: mathematicians report write-ups that are hard to follow, unusually short and with thin bibliographies, which renews concerns about attribution. Brown's Brendan Hassett said the write-up he knew best "made little sense" on a quick read.
- Methods: Lichtman and UConn's Álvaro Lozano-Robledo say the results use existing techniques in ingenious ways rather than new methods (The Verge).
- How the community should respond: AHM calls the release "a demonstration of power." Commenters on its statement dispute whether AHM speaks for most mathematicians (AHM statement).
Business and practitioner implications
- Checking is now the bottleneck. Tools that align Lean statements with paper claims, and services that audit formalizations, are an obvious gap.
- Release standards are being set. AGMAI's norms ask for formal proofs, disclosed models and prompts, and no marketing-driven releases (AGMAI).
- Funders and departments are exposed. Grant proposals and research programs in probability, combinatorics and theoretical computer science were reportedly overtaken by the release (The Verge).
Read the full section
- Checking is now the bottleneck. Generating results is cheap (about three compute-hours each), but checking them is not, and that cost falls on the field. Tools that align Lean statements with paper claims, and services that audit formalizations, are an obvious gap.
- Release standards are being set. AGMAI's norms ask for formal proofs, disclosed models and prompts, and no marketing-driven releases (AGMAI). They are likely to shape expectations for AI science claims in other fields.
- Funders and departments are exposed. Grant proposals and research programs in probability, combinatorics and theoretical computer science were reportedly overtaken by the release (The Verge).
- Treat these as demonstrations, not a product. None of these capabilities can be bought or used yet.
The source trail.
Sources (10)
‘Pure insanity’: Mathematicians will need years to make sense of OpenAI’s latest drop
Article text retrieved; extracted text may omit tables or interactive elements.
theverge.com