OpenAI releases 722 AI-written math manuscripts claiming hundreds of open problems
Including quasi-Riemann, Unique Games, Hodge for CM abelian varieties and free group factors
Event confirmedAwaiting review
Importance: historic (5 of 5)Status
- Claim
Event confirmedAwaiting review
- Our reporting
- High confidence
- Verification
- Lean proof, not yet reviewed
- Importance
- Historic (5 of 5)
- Last verified
- 16 November 2026
Your AI and this story
- GPT-6.2 Astra128 days after its cutoff
- Claude Opus 5.598 days after its cutoff
- Gemini 4 Hydrogen189 days after its cutoff
- Grok 4.7128 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 98 days before it.
Key facts
- Xena Project blog, ‘…and the proof is checked in Lean’ (12 Nov; HN front page 16 Nov, 42 points): proposes four checks for AI-generated Lean repositories (statement matches the informal one, the code compiles,
lake comparator, andlake comparator --paranoidagainst kernel exploits); says that of the release’s ~290 papers with Lean code about 38 do not state or prove their main theorem, and warns ‘this does not mean that any corresponding pdf also contains a correct proof!’ - Repo: github.com/openai/math, single ‘Initial commit’ at 2026-10-06 14:58 PDT; ~3.1 GB, 722 manuscripts in preprints/ (PDF + source + BibTeX), overview.pdf (‘OpenAI Research Catalog’, 17 subject sections), CONTENTS.md manuscript map, Lean library in lean/, 10 reasoning summaries
- Manuscript dates run from 10 Sep to 6 Oct 2026; most are dated 23–27 Sep (≈560) and 4–5 Oct (≈136). OpenAI had claimed ‘100+ open problems’ on 21 Sep (see OpenAI says an internal model resolved 100+ long-standing open problems in 24 days of training (list released 6 Oct: 722 manuscripts))
- Procedure (README): ‘The vast majority of results were obtained with the same procedure using an unreleased internal OpenAI model. On average, each result used three hours of ChatGPT Pro thinking compute with that model.’ About 4,000 problems were posed; families were kept if they reached ‘an appropriate level of significance’
- Stated exceptions to the fixed procedure: the zero-free region for the Riemann zeta function and the Hodge conjecture for CM abelian varieties; the Re(s) > 11/12 zero-free-region write-up was ‘human edited for readability’
- Caveats (README): ‘This collection includes results at different stages of verification. Not all have accompanying Lean formalizations… Some of the unformalized results could have issues. We will endeavor to fix any such issues quickly.’ Corrections will be released as new versions, old versions kept
- Lean: lean/formalization.yaml lists 162 manuscripts with a formalized main result and 185 Comparator challenge configurations; scope ‘Partial progress’; review status ‘unchecked’. Challenges allow only the standard axioms (propext, Quot.sound, Classical.choice). It builds on Mathlib, PrimeNumberTheoremAnd (Kontorovich) and other community libraries. Per-family scope notes say exactly which statements were formalized (often the main theorem, not every consequence)
- Reasoning summaries for 10 families: two-point Chowla (007), irrationality exponent of π (017), Mahler conjectures (087), NP-hardness at the basic SDP threshold (102), quasipolynomial Szemerédi bounds (159), Kaplansky direct finiteness in char 2 (197), Mézard–Parisi formula (221), spontaneous magnetization in the quantum Heisenberg ferromagnet (271), free group factor isomorphism (287), 3D relativistic Vlasov–Maxwell (362)
Show 33 more
- OpenAI blog ‘Sharing AI progress in mathematics’: the release drew on the ‘advice and public recommendations’ of the independent Advisory Group on Mathematics and AI (AGMAI, IAS); OpenAI will fund ‘workshops, conferences, and special programs around the understanding of major results produced by AI’ and is ‘working to responsibly release the model that produced these results’
- AGMAI statement (6 Oct): ‘an important event for mathematics’; its advisory role ‘should not be interpreted as a judgment of the impact of these results or an endorsement of the process’; ‘This release is the beginning, not the completion, of the process of human understanding’
- OpenAI’s X post (6 Oct, 22:19 UTC): ~3.97M views, 18k likes on 7 Oct
- Famous claims, Lean-formalized per OpenAI’s catalogue (not yet independently reviewed): quasi-Riemann hypothesis Re s > 7/8 for ζ and all Dirichlet L-functions (003, 199-page paper) and uniform exclusion of Landau–Siegel zeros; Unique Games Conjecture (102); all free group factors isomorphic, L(F2) ≅ L(F3) (287); irrationality exponent of π equals 2 (017); Catalan’s constant irrational (005); Kaplansky zero-divisor conjecture false (196) and direct-finiteness false in char 2 and odd char, via a finitely presented nonsofic group, also refuting Gottschalk’s surjunctivity conjecture (197); Thompson’s group F nonamenable (248); Cannon’s conjecture (246); ω ≤ 9/4 for matrix multiplication (107); symmetric and general Mahler conjectures (087); Falconer distance conjecture in all dimensions (073); Erdős’s reciprocal-sum (Erdős–Turán) conjecture on arithmetic progressions (159); two-point Chowla (007); the plane is not 5-colourable (158); Borsuk fails in dimension 9 (156); Sidorenko’s conjecture false (161); Barnette’s conjecture (180); Seymour’s second-neighbourhood conjecture (173); circulant Hadamard conjecture (179); Artin K(π,1) conjecture (254); Kadison similarity (288); hyperinvariant-subspace counterexample (293); hot-spots conjecture for simply connected planar domains (369); spontaneous magnetization of the 3D quantum Heisenberg ferromagnet (271); 3D relativistic Vlasov–Maxwell global regularity (362); Mézard–Parisi formula (221)
- Famous claims with NO Lean scope note (manuscript only): rational Hodge conjecture for CM abelian varieties, with the Tate conjecture for abelian varieties over finite fields and the Hodge standard conjecture as consequences (032); full BSD formula for Selmer corank 0/1 and Goldfeld’s conjecture (002, 006); Hilbert’s tenth problem over ℚ undecidable (004); L = RL = BPL (103; a Lean formalization was added in the 7 Oct update, see OpenAI claims L = RL = BPL with a Lean proof); modularity of elliptic curves over imaginary quadratic fields (030); Fujita’s freeness conjecture (038); Bloch’s conjecture for surfaces (040); Kakeya maximal conjecture in 3D (074); abelian Zilber–Pink (016); the p-adic section conjecture (019); infinitude in Artin’s primitive root conjecture (029); Hilbert–Smith conjecture (304); uniform limit-cycle bound in Hilbert’s 16th problem (143; only the quintic Liénard case is in Lean); Hadwiger’s conjecture false in fractional form (157; Lean covers only the companion linear list-Hadwiger bound)
- Several families overlap earlier AI-era results already in this dataset (k-server, Crouzeix, Mumford–Shah, Kahn–Kalai, Borsuk, linear Hadwiger)
- Overlap: family 181 (Erdős–Gallai cycle-decomposition conjecture, manuscript dated 24 Sep) proves the same theorem as Jaehoon Kim’s arXiv 2610.07840 (posted 6 Oct, disclosing ‘extensive use of ChatGPT (OpenAI) and Claude (Anthropic)’)
- A model’s reaction: Claude Opus 5.5, run on the cloned repo by @ereliuer_eteer and posted as its output 85 minutes after the release (~90k views; header ‘Opus 5.5 on github.com/openai/math’, and the text addresses the poster as ‘you’), checked the Lean challenge statements (‘Nothing hides in a home-made definition’), flagged that HarmonicGrowth.lean states its theorem as an axiom, and described ‘a mix of awe and vertigo, plus a bit of grief’. We confirmed the flagged details in the repo; its ‘~120 formalized papers’ undercounts formalization.yaml’s 162
- Repo structure (checked 7 Oct): lean/ComparatorChallenges holds 405 challenge statements, each with a JSON config allowing only propext, Quot.sound and Classical.choice. Most state the theorem with
sorry(e.g. quasi-RH:riemannZeta s ≠ 0for7/8 < s.re, against Mathlib’s riemannZeta); HarmonicGrowth.lean instead usesaxiom mainStatement : MainClaimwiththeorem main := mainStatement, presumably a stand-in since the comparator checks the solution module. The quasi-RH solution (OAI.NumberTheory.DirichletL.Nonvanishing) imports a module named Detector.FinalAssemblyUnconditional - Viral readings checked against the repo (7 Oct): ‘kakeya conjecture in R^3 has been resolved’ (Rohan Pandey, ~96k views) mixes up two problems. The 3D Kakeya set conjecture was already proved by Hong Wang and Joshua Zahl in 2025 (arXiv 2502.17655). Family 074 claims the 3D Kakeya maximal conjecture and the 4D Hausdorff-dimension conjecture, both manuscript only. ‘five-colours is no go’ (Peter Gostev, ~59k) is accurate: family 158 (Lean: no_proper_five_coloring, arbitrary colourings) leaves 6 or 7 as the only possible chromatic numbers of the plane, partial progress on Hadwiger–Nelson (lower bound 5 since de Grey 2018)
- WSJ (7 Oct): OpenAI says it set its model on the Millennium Prize Problems, pouring ‘millions of dollars in computing power’ into the effort, only after hearing that Anthropic had already solved at least one. It declined AGMAI’s request to stop testing on proprietary models: ‘It is important to continue to evaluate our internal frontier models on mathematics and other sciences’
- First withdrawals (history.md, 7 Oct): a sign error in ‘Algebraicity of Weil classes on split abelian eightfolds’ ‘invalidates a stabilization-trace cancellation argument and the construction used by two dependent papers’; OpenAI withdrew it together with ‘Algebraicity of Kuga–Satake Correspondences for K3 Surfaces’ and ‘The rational Hodge conjecture for products of K3 surfaces’ (all in the Hodge family 032, none Lean-formalized). The main CM-abelian-variety Hodge paper was not withdrawn. 14 other manuscripts were repaired, 13 updated citations; the catalogue now counts 719 manuscripts, with 300 of 719 top-line results (~42%) formalized after 6 new formalizations
- The flaw was flagged publicly first by Elliot Glazer (2:46 UTC, 8 Oct): ‘Based on a hot tip: I had Astra check the OpenAI paper … and it seems to be flawed. This is one of the nonformalized results.’ (~119k views). OpenAI’s Dan Roberts announced the update at 5:20 UTC (~341k views); HN threads 337 and 224 points
- Independent human-led (often AI-assisted) proofs of problems in the catalogue appeared within days, with priority statements: Kittle & Kogler’s exact overlaps conjecture (arXiv 2610.10511, found by GPT-6 Astra on 27 Sep; OpenAI family 148 dated 24 Sep), Hu, Liu & Wan’s LeBrun–Salamon conjecture (2610.10410, ‘obtained well before’ OpenAI’s family 062), Jaehoon Kim’s Erdős–Gallai cycle decomposition (2610.07840, family 181), and Minzer, Fei & Wang’s 4-to-1 games result posted on 14 Sep after hearing rumours (Quanta)
- More overlaps posted 7–8 Oct (sweep 2026-10-09): Engel & Mauri’s human-written proof of the hyperkähler SYZ conjecture (arXiv 2610.12277; family 041; ‘The structure of the proofs appears similar’); Qin & Sun’s Ω(d² ln² d) Hilbert-number lower bound, obtained ‘by the end of August’ with GPT-5.6 Sol (2610.12218; complementary to family 143’s finiteness claim); Ky’s antipodal Hellinger conjecture (2610.10618; family 119). Family 119’s Courtade–Kumar conjecture had already been proved three times on 21–22 Sep (see Courtade–Kumar ‘most informative Boolean function’ conjecture (2013) proved three times in two days, all with AI: Ky & Tran (ChatGPT), Google + CUHK (Gemini, Lean-verified end-to-end), Mahdavifar & Beirami)
- Expert refusal: Asaf Karagila, the leading specialist on the Partition Principle (family 244), wrote on 8 Oct that the preprint ‘sucked. It is unclear, muddled’ and should be desk-rejected, and that he will not review it; he called the release ‘the equivalent of a Denial of Service’ (HN 172 points; see Set theorist Asaf Karagila refuses to review OpenAI’s claimed solution of the Partition Principle problem (does PP imply AC?): ‘It sucked’; calls the 722-paper release a ‘Denial of Service’)
- Organised opposition: the Association for Human Mathematics (statement dated 6 Oct, reposted as a guest post on Terence Tao’s blog on 7 Oct) called the release ‘not a demonstration of scholarship, but a demonstration of power’ and urged mathematicians to ‘discontinue their work with OpenAI’
- Autoformalisation critique: Bastounis, Circelli & Hansen (arXiv 2610.08144, 6 Oct; HN 334 points) argue that Lean verification of AI autoformalisations ‘may offer no confidence in the original NL argument’ and show that the Lean proof of OpenAI’s Navier–Stokes blow-up does not correspond to its natural-language proof
- ‘Lean checks at 42%’ headlines (tech-insider.org, Developers Digest, 8–9 Oct) are a misreading: the figure is OpenAI’s own README line ‘The repository has ~42% top-line results formalized’, i.e. history.md’s 300 of 719 top-line results after the 7 Oct update. It is a share formalized, not a share that passed or failed an independent check
- Axios (8 Oct), ‘OpenAI’s math breakthrough points beyond math’: maths, like programming, gives AI ‘a way to tell when it is right’, since proofs can be checked by experts and increasingly by computers; the piece quotes a Rutgers mathematician (Kontorovich’s ‘instant Fields Medal’ line) alongside doubts about how original the results are
- 36Kr’s ‘Terence Tao Reposts Boycott Statement: Mathematics Community Fully Cuts Ties With OpenAI’ (9 Oct) overstates events: it refers to the AHM statement Tao ran as a guest post on 7 Oct (already above), not a new statement by Tao or a field-wide boycott; the AHM text sets no enforcement mechanism
- Community reactions collected (Proofs and Prompts, ‘100+ reactions to 100+ solutions’, posted Oct 8, updated Oct 10; roughly 80–100 signed reactions): Tao: ‘many of the AI-generated proofs appear to introduce clever new ideas’ but ‘I am horrified by the many person-years of ongoing patient and deliberately slow research efforts … being casually disrupted or destroyed by such a release’; Scholze: ‘mathematics is a marathon, not a sprint’; Agol: problems ‘I might not have expected to be resolved in my lifetime’; Ben Green: the Re s > 7/8 proof ‘is absolutely shocking to me’; Henry Wilton: ‘If OpenAI wanted to destroy the mathematical community, this would be a great way to go about it.’
- Economist Joshua Gans (‘OpenAI drops a bomb on maths’, Substack, Oct 7): ‘this is probably the biggest day of scientific advancement in history’; argues AI will split mathematicians’ work into problem formulation and verification around machine proving
- The Verge long read (Robert Hart, 9 Oct; ‘more than three dozen mathematicians’): Lichtman counts ‘tens’ of Fields-calibre results (progress toward RH, a Hodge special case, 4D Kakeya) but says all come ‘out of existing methods and techniques’; Scott Armstrong says the targets show OpenAI is going after Yang–Mills and the mass gap; Buckmaster has heard of ‘three people who had entire research programs obliterated’; Colva Roney-Dougal: ‘friends and colleagues whose grant proposals have just been wiped out’
- Tao’s Caltech lecture ‘Math 2.0’ (9 Oct, slides 10 Oct): ‘proof abundance’ makes ‘further blind optimization of problem-solving alone … actively harmful to the long-term health of mathematics’ (see Terence Tao’s Caltech lecture ‘Math 2.0’)
- NBC News (Evan Bush, 11 Oct; mostly paywalled, opening read): headline ‘Some OpenAI math discoveries are “jewels,” if verified. But mathematicians are wary of “slop.”’; one mathematician said the community should be ‘dancing in the streets’, another compared the release to a tsunami hitting her field; NBC counts potential solutions to roughly 370 problems
- Retraction Watch (Alicia Gallegos, Oct 8): an OpenAI spokesperson said the errors behind the three withdrawals were found in an internal audit: ‘Where errors are identified, we will work to correct them promptly and withdraw papers if no fixes can be found.’ Roughly half of the results were released without formal confirmation; OpenAI said its AGMAI advisers recommended releasing before ‘full formalization’. Cornell’s Alex Townsend: ‘I think that OpenAI should have announced the manuscripts that were lean verified first’; he expects more errors. MIT’s Andrew Sutherland called the withdrawals ‘the responsible thing to do’ and said many mathematicians are ‘unhappy with how OpenAI has behaved up to this point’
- Sophie Huiberts, guest post on Tao’s blog (Oct 12), ‘This is normal bad work’: ‘This is poor quality work and these pdfs are all easy rejects’; ‘October 6th was not a watershed moment for mathematics’; ‘These pdfs are exhaust fumes from their internal benchmarking effort’. She criticises attribution, errors, missing code and ‘salami-slic[ing] 372 results over 719 pdfs’, and says a software supplier scooping its own ChatGPT users is ‘a bad look’ (post file 2026-10-12-huiberts-this-is-normal-bad-work)
- Second-wave press, Oct 12: franceinfo (via RTÉ) quotes Cédric Villani calling it ‘the most radical change in mathematics in at least four centuries, perhaps even in 1,000 years’, says Hugo Duminil-Copin called the shock ‘monumental’, said he cried and had been working on 16 of the problems, and quotes Northwestern’s Antonio Auffinger (a year ago there were ‘always hallucinations’) and Benjamin Antieau; Cult of Mac (via Yahoo, citing The Washington Post) reports that NYU’s Tristan Buckmaster now keeps original ideas out of Codex, while OpenAI says its investigation found his prompts could not have influenced its Navier–Stokes work. The Economist (‘AI’s mathematics blitz pushes science into strange new territory’) and Quartz (‘OpenAI’s math data dump wiped out early-career researchers’ programs, professor says’) were not readable by our fetcher
- Critique of one manuscript (Ben Recht, UC Berkeley, arg min blog, 12 Oct; about 150 points on Hacker News): the release’s claimed asymptotically faster Fourier transform is, in his reading, a ‘galactic’ algorithm. ‘The estimated values for C were 10^100 when N=10^12’, so the classical FFT stays faster at any usable size. He calls it ‘less a breakthrough and more a malicious exploit’ of big-O notation. He says he did not read the paper and examined it by questioning ChatGPT about it
- Second wave, Oct 12-13: Le Monde (‘They have annihilated my field of research’, Hugo Duminil-Copin; the article blocks automated access and was not read, quote from its headline and from IBT UK’s summary), The Economist, Quartz (early-career programmes), Tom’s Hardware, TechRadar (‘mathpocalypse’), RTE and The Week. IBT UK’s summary adds that Duminil-Copin said he believed OpenAI wanted to show it could ‘massacre’ mathematicians, and that Timothy Gowers declined to sign the Fields medallists’ letter, saying a flood of results is not in itself a disaster while the loss of open problems could weaken the talent pipeline
- Own entries added 29 Oct for two claim clusters our first coverage missed: log abundance and canonical nonvanishing in characteristic zero, with companion Iitaka-subadditivity papers (OpenAI claims the abundance conjecture; Jihao Liu raises priority; Jihao Liu says a draft he sent to Codex may have fed it), and L = RL = BPL with a Lean proof whose statement definitions a volunteer re-check found identical to the proof’s (OpenAI claims L = RL = BPL with a Lean proof)
- Bryna Kra (Northwestern, former AMS president) on KQED’s podcast Life, Automated (28 Oct; computer-generated transcript): ‘There’s a complete misalignment in the interests of the AI companies and the interests of the mathematical community’; the AI-written proofs ‘were very bad at tracing where the results came from’; ‘The problem is solved, but without any understanding of what was going into it’
What happened
Late on 6 October 2026, OpenAI published the list it had promised on 21 September, when it said an internal model had “resolved” more than 100 open problems. The list turned out to be far larger: a public GitHub repository with 722 manuscripts in 372 “result families”, each family grouping a main result with companions, consequences or alternative proofs, classified into 17 subject areas from number theory to PDE.
The README says the model was posed about 4,000 problems during evaluation, “after performance on our existing mathematical evaluations saturated”. Each kept result used on average about three hours of ChatGPT Pro thinking compute. OpenAI names two exceptions to this fixed procedure, the zeta zero-free region and the Hodge conjecture for CM abelian varieties, without saying exactly how they differed. It also publishes abridged “summarized chain of thought” documents for ten families. These show the model trying and discarding approaches; the free-group-factor summary, for example, lists failed attempts with rigidity, Gaussian crossed products and free entropy before the final construction.
Verification is uneven, and OpenAI says so. A Lean 4 library ships with the release. formalization.yaml lists 162
manuscripts with a formalized main result and 185 Comparator challenges, and 235 of the 372 families in CONTENTS.md link
a Lean scope note. Each challenge states a theorem, and a checker confirms that OpenAI’s proof proves exactly that statement
using only Lean’s standard axioms. Whether each formal statement faithfully captures the famous problem is still a question
for human reviewers, and the catalogue’s own review field reads “unchecked”. The other families, among them the Hodge, BSD,
L = BPL and Hilbert’s tenth claims, are unrefereed manuscripts.
The release followed advice from the independent Advisory Group on Mathematics and AI (AGMAI) at IAS, whose 29 September guidelines asked labs to disclose compute, prompts and reasoning, formalize proofs and fund human understanding. AGMAI’s statement the same evening called it “an important event for mathematics” but said its role “should not be interpreted as a judgment of the impact of these results or an endorsement of the process”.
The most famous claims
Status below: Lean means OpenAI’s catalogue gives a formal proof of the main statement (machine-checked against a published statement, not yet independently reviewed); paper means manuscript only.
- Quasi-Riemann hypothesis: ζ and every Dirichlet L-function have no zeros with Re s > 7/8, plus uniform exclusion of Landau–Siegel zeros (Lean). Separate entry: OpenAI model claims the quasi-Riemann hypothesis.
- Khot’s Unique Games Conjecture, plus direct optimal Max-Cut and Vertex Cover hardness (Lean). Separate entry: OpenAI model claims a proof of Khot’s Unique Games Conjecture, with Lean formalization.
- Rational Hodge conjecture for all CM abelian varieties, which would give the Tate conjecture for abelian varieties over finite fields (paper). Separate entry: OpenAI release proves the rational Hodge conjecture for all CM abelian varieties, giving the Tate conjecture for abelian varieties over finite fields; Lean proof added 26 Oct, human-written proof by van Hoften and Sempliner, and Milne calls it proved.
- Free group factors: L(F₂) ≅ L(F₃), so all interpolated free group factors are isomorphic (Lean). Separate entry: OpenAI model claims the free group factor problem.
- Kaplansky’s conjectures: a torsion-free group algebra over 𝔽₂ with zero divisors (196); direct finiteness fails in characteristic 2 and odd characteristic, through a finitely presented nonsofic group, which also refutes Gottschalk’s surjunctivity conjecture and the group-ring Determinant Conjecture (197). All Lean.
- Irrationality exponent of π is exactly 2, which implies the Flint–Hills series converges (Lean for the exponent); Catalan’s constant is irrational (Lean).
- Thompson’s group F is nonamenable; Cannon’s conjecture; Artin K(π,1) conjecture; a torsion-free hyperbolic group that is not residually finite (all Lean).
- Matrix multiplication ω ≤ 9/4 (Lean); integer multiplication and the DFT below n log n (109, 130); L = RL = BPL (paper); NP-hardness of 3-colouring with any constant number of colours (Lean).
- Combinatorics: Erdős’s reciprocal-sum conjecture, i.e. any set of integers whose reciprocals diverge contains arbitrarily long arithmetic progressions (Erdős’s best-known prize problem; Lean); the plane is not 5-colourable, Borsuk fails in dimension 9, Sidorenko’s conjecture false, Seymour’s second-neighbourhood conjecture, Barnette’s conjecture, the circulant Hadamard conjecture (all Lean); Hadwiger’s conjecture false in fractional form (paper; Lean covers only the companion linear list-Hadwiger bound, compare Linear Hadwiger conjecture proved).
- Analysis and geometry: Falconer distance conjecture in all dimensions (Lean); Kakeya maximal conjecture in 3D (paper); symmetric and general Mahler conjectures (Lean); Hilbert’s sixteenth problem, uniform bound on limit cycles in each degree (paper; Lean covers only the quintic Liénard case); hot-spots conjecture for simply connected planar domains (Lean); global regularity for 3D relativistic Vlasov–Maxwell (Lean).
- Number theory and arithmetic geometry (paper): full BSD formula for Selmer corank ≤ 1 and Goldfeld’s conjecture; Hilbert’s tenth problem over ℚ (undecidable); modularity over imaginary quadratic fields; abelian Zilber–Pink; the p-adic section conjecture; infinitude in Artin’s primitive root conjecture.
- Operator algebras: Kadison similarity, strong Kadison–Kastler, the generator problem for II₁ factors, a counterexample to the hyperinvariant-subspace problem on Hilbert space (Lean); counterexamples to Baum–Connes (coefficient-free reduced) and Kadison–Kaplansky (paper).
Why it matters
If even a fraction of these results hold up, this is the largest single jump in mathematical knowledge on record, produced by an AI system in about six weeks. Machine-checked Lean proofs for famous statements such as Unique Games and the quasi-Riemann hypothesis make parts of it verifiable far faster than refereeing. The scale also creates a new problem, the one AGMAI flagged: the community must now absorb, check and understand hundreds of results it did not choose. Until that happens, treat each individual claim as unverified unless a Lean proof exists and experts have confirmed that its formal statement matches the problem.
What is disputed or not yet verified
| Verification | Mixed: Lean formalizations for 162 manuscripts / 235 family scope notes (OpenAI-produced, review status ‘unchecked’); the rest are unrefereed preprints |
|---|
Reactions
First ~7 hours; X view counts from fxtwitter, early 7 Oct
- Alex Kontorovich (number theorist): “Quasi-RH?!?!???! Are you kidding me? If a human did this, it would be an instant Fields Medal, no questions asked… They got a zero free strip!!!! Insane” (X56, ~420k views). A follow-up: “Yeah, no Siegel zeros either. So I guess two Fields medals…”
- Levent Alpöge (Anthropic): “Big, big, big, big props for quasiriemann and no Siegel zeroes”. He called it “obviously the most significant moment in mathematical history”, and also wrote: “There are some sad stories related to their users getting scooped / conflicts of interest (kakeya maximal comes to mind)” (X57, ~112k). The Kakeya scooping allegation is unverified.
- Daniel Litt (Toronto): “Looks like mathematicians have a lot of exciting work to do”. He called some framings of “progress” on famous conjectures “a stretch”, and flagged that the Bloch’s-conjecture paper “seems to cite some wrong existing work without mentioning it’s wrong… but AFAICT doesn’t meaningfully rely on it” (X58, X).
- Jared Duker Lichtman: “This output is absolutely historic… However, it is not the Millenium problem rumor”. He said the cubic family of L-functions “appear to be essential for the proof” of the zero-free strip (X).
- Steven Strogatz, on ω ≤ 9/4: “This leap in progress is like Bob Beamon’s long jump” (X60, ~118k).
Show 46 more reactions
- Joe Bebel (TCS): the release “includes a claimed proof and Lean formalization of the Unique Games Conjecture, which I have been personally working on for the last year and a half” (X59, ~99k). Chris Peikert: “Unique Games Conjecture, proved. L=RL, proved.” (X).
- Kiwamu Watanabe (algebraic geometer, in Japanese) called the algebraic-geometry list (Hodge for CM abelian varieties, Fujita, Nagata, log abundance, MMP termination) “quite amazing”, but said he cannot tell whether it is correct (X, ~70k). Isaac Kim called the quantum-information claims “a shocking list of problems” (X).
- Ken Ono: “AI-assisted mathematical discovery at the highest level is no longer hypothetical” (X). Gil Kalai: “Certainly this is an amazing milestone for mathematics, and the results will needs to be verified” (blog64).
- Terence Tao, about 30 minutes after WIRED reported the coming release (before it went live), on Mathstodon: solutions “are now being harvested at large scale in an unsustainable fashion”, and a “Math 2.0” should “decenter the role of raw problem solving” (Mathstodon63). He had not commented on the content by early 7 Oct.
- Andrew Sutherland (MIT), to Scientific American: treat one-shot claims as unverified “until and unless they release the model… We should ask for receipts.”
- OpenAI side: Sam Altman, “We are entering a new era of discovery now” (X62, ~520k); Greg Brockman (X); Codex lead Thibault Sottiaux (X, ~356k). Former OpenAI contributor @AcerFur on family 109: “we can do integer multiplication faster than n log n” (X, ~264k).
- Commentary: Zvi Mowshowitz, “simultaneously the biggest day in mathematical history, and also felt kind of slow”; François Chollet asked whether the jagged frontier is “mainly math + code” driven by RL with verifiable rewards; Gary Marcus said it “uses symbolic AI (Lean, etc)”. High-reach posts by Dan McAteer (~388k), Will DePue (~282k) and AISafetyMemes (“90 of the 500 most important open problems”, based on an LLM-ranked list at proofatlas.ai, not a mathematicians’ ranking).
- Independent Lean re-check: within hours, a public repo (davegoldblatt/openai-zeta-proof-check, created 7 Oct 01:36 UTC) re-ran the Comparator check of the 7/8 zeta theorem. It reports acceptance by two kernels (Lean’s and nanoda) with only the three standard axioms; the proof spans ~2,900 files and ~500k lines. This is one person’s run; the paper itself was not reviewed.
- Whole-library re-check (Oct 10–25): Neal McBurnett rebuilt the entire OAI Lean library (122,458 AI-generated files, 26.1M lines, about eleven times Mathlib) at openai/math commit fd4aeeb2ee and ran Comparator on all 416 challenge configurations (github.com/nealmcb/openai-math-lean-tooling, built with Claude Code). “Everything checks, with Lean’s kernel”: all 416 accepted, negative controls rejected. The independent nanoda kernel accepted 401 of 416 (with an enlarged stack); its failures were nanoda bugs, not proof errors, reported as nanoda_lib issue 55 (linked from the quasi-Riemann entry; opened Oct 15, closed Oct 21; 14 of 15 pass on newer nanoda, one SquareDifference declaration still overflowed). In his Oct 15 run both kernels accepted QuasiRiemannHypothesis (Lean replay ~6 min, nanoda ~81 s), a second independent two-kernel pass for the 7/8 theorem. Cost: ~3.8 days and ~410 CPU-hours on a laptop, or 13.8 hours and ~$85 on one 32-vCPU cloud machine. He stresses that a pass shows the Lean proof matches the Lean statement, not that the statement is faithful to the paper. We found no Lean Zulip or other forum thread re-verifying the 7/8 proof beyond these two repos.
- High-reach explainers and memes (views at 11:31 UTC on 7 Oct): kimmonismus posted an accurate summary of the README, “HOLY, the rumors were true” (X, ~195k). Hesamation wrote “mathematicians: frontier labs need to publish real proofs so we can check them, not blog posts / OpenAI: fine. here’s 722 papers” (X, ~122k). The meme “math is solved / it’s so over” by @cgtwts reached ~182k (X).
- Kakeya and Hadwiger–Nelson, as read on X: Rohan Pandey’s “kakeya conjecture in R^3 has been resolved” (X, ~96k) conflates the 3D Kakeya set conjecture, proved by Hong Wang and Joshua Zahl in 2025 (arXiv 2502.176559), with what family 074 claims: the 3D maximal conjecture plus the 4D dimension conjecture, both manuscript only. Peter Gostev said he had spent “300 billion tokens on Hadwiger–Nelson problem (GPT-5.6 and Astra) and got absolutely nothing”. He read family 158 correctly: “five-colours is no go… Not a full solution yet, but it’s a biggie” (X, ~59k). The manuscript “The Euclidean plane is not five-colorable” (dated 23 Sep) has Lean comparator statements for no proper five-colouring and for a proper seven-colouring, so the chromatic number of the plane would be 6 or 7. The lower bound had been 5 since de Grey (2018).
- No expert had reported a mathematical error in a main proof by early 7 Oct.
- First follow-up work: on 7 Oct Doug Colkitt (@0xdoug) published conditional improvements to family 109’s integer-multiplication exponent, from κ = 2^-182 to 2^-59 in three releases, done “with assistance from OpenAI Codex” and assuming OpenAI’s unverified theorem (GitHub8, X). Separate entry: Outside researchers using AI crowdsource a tighter exponent in OpenAI’s integer-multiplication result #109, from 2^-182 to about 2^-8.4 (conditional).
Press
- WIRED (6 Oct, before the release): “OpenAI is pissing off a bunch of mathematicians again”. It reports that mathematicians felt their advice was ignored (Bryna Kra: “Apparently, that input was ignored”; Nestor Guillen: “there’s a perception of mobster behavior”).
- NYT (Kenneth Chang, Siobhan Roberts): “OpenAI Releases Findings on 377 Math Problems, Further Roiling Field”. Paywalled; we have only the preview, and the 377 figure is the NYT’s count. Gizmodo also says 377; the repo itself says 372 families.
- Scientific American (Joseph Howlett), The Verge (Robert Hart; quotes AGMAI asking labs not to use releases as “marketing vehicles”), Gizmodo (Mike Pearl; notes AGMAI’s 29 Sep request that labs “stop testing advanced mathematical problems on proprietary models”).
- WSJ, “AI Solved One Math Problem and Everyone Freaked Out. It Just Cracked Hundreds More.” (7 Oct, 01:09 UTC). We read it in full through HT Tech’s syndicated copy; the WSJ original is paywalled. It describes “advances related to three of the five remaining Millennium Prize Problems, including the Riemann hypothesis”. That is the WSJ’s framing; Lichtman said the release is “not the Millennium problem rumor”. According to the WSJ, OpenAI says it set its frontier model on the Millennium Prize Problems, “and poured millions of dollars in computing power into the effort”, only after hearing that Anthropic had already solved at least one. OpenAI turned down AGMAI’s request to stop testing on proprietary models: “It is important to continue to evaluate our internal frontier models on mathematics and other sciences, so we can accelerate developing the tools to advance those fields.” The WSJ also reports that mathematicians rushed to post their own AI-assisted work before they could be scooped. It quotes Kontorovich by name and Alpöge’s “obviously the most significant moment in mathematical history” as the words of “a researcher inside Anthropic”.
Day two (7–8 Oct): first withdrawals, rival proofs and a boycott call
- Withdrawals. On 7 Oct OpenAI updated the repository and its new
history.md. A sign error in “Algebraicity of Weil classes on split abelian eightfolds” invalidated “a stabilization-trace cancellation argument and the construction used by two dependent papers”, so OpenAI withdrew that paper, “Algebraicity of Kuga–Satake Correspondences for K3 Surfaces” and “The rational Hodge conjecture for products of K3 surfaces”. All three belong to the Hodge family 032 and none had a Lean formalization. The headline CM-abelian-variety paper was not withdrawn. Fourteen other manuscripts were repaired (Lipschitz heights/Ashkin–Teller, Kähler MMP and abundance, taming/hypersymplectic, incompressible box transport, a stale citation in the BSD paper), and 13 more had their citations updated. Six new formalizations brought the count to 300 of 719 top-line results (~42%). Elliot Glazer had posted about two and a half hours before OpenAI’s announcement that, “based on a hot tip”, he had GPT-6 Astra check the eightfolds paper and “it seems to be flawed” (X65, ~119k views). OpenAI’s Dan Roberts announced the update: “We will continue to update the repo with new formalizations and with any errata we notice” (X11, ~341k views). This was the first confirmed error in the release, and it fell in the unformalized part, as OpenAI’s README had warned. - Rival and independent proofs. Within a day, mathematicians posted papers on problems that also appear in the catalogue, each with a priority note. Kittle and Kogler proved the exact overlaps conjecture on the line (arXiv 2610.10511). Their disclosure says GPT-6 Astra found the first proof on 27 Sep, while OpenAI dates its family 148 manuscript 24 Sep. Hu, Liu and Wan’s LeBrun–Salamon proof (2610.10410) says its argument was “obtained well before” OpenAI’s family 062. Earlier, Jaehoon Kim had done the same for Erdős–Gallai (family 181). Quanta’s Ben Brubaker reported that Dor Minzer’s MIT group rushed out a 4-to-1 games result on 14 Sep, three days after hearing rumours of OpenAI’s proof. Minzer: “There is a lot of value in failing and knowing why you failed. Using AI takes all of this out.” Braverman and He (2610.10108) note that OpenAI’s sub-n log n Fourier-circuit result also implies their refutation of the multiple-unicast conjecture.
- Backlash and reflection. The Association for Human Mathematics (see Quanta’s math editor asks ‘Is AI the End of Math As We Know It?’) issued a statement dated 6 Oct: “Releasing over 700 files at once is not a demonstration of scholarship, but a demonstration of power”. It urged mathematicians to “discontinue their work with OpenAI and to return to a vision of science that centers human understanding”. Terence Tao reposted it as a guest post on 7 Oct, followed by guest posts from Raghu Meka (“The barriers of perception”: “the downside of not believing there is a barrier is far less than that of believing there is one”) and Álvaro Lozano-Robledo (“What should we tell our students?”: “Keep calm and carry on studying math”). Bastounis, Circelli and Hansen (arXiv 2610.08144) argued that Lean checks of AI autoformalisations need not certify the natural-language proof, using OpenAI’s Navier–Stokes formalization as an example (see OpenAI claims a Millennium Prize problem). On 8 Oct Tao also shared the update to a parody he had posted before the release, “OpenAI Releases the Final Ten Minutes of 500 Previously Unreleased Films” (@Endings500), which now read “We Are Withdrawing Three Endings from Endings-500, Reaffirm Our Full Confidence in the Remaining 497” (Mathstodon). TechCrunch’s 8 Oct headline summed up the second-day mood: the solutions “aren’t meeting the field’s standards yet”.
- Outside mathematics. Ethereum researcher Justin Drake cited the release (“mathematical superintelligence is upon us”) in a 5.5M-view call for crypto “bunker mode”; see After OpenAI’s math release, Justin Drake urges crypto ‘bunker mode’ over AI-broken ECDSA.
Day three (8–9 Oct): expert triage, Lean doubts and the press wave
- Erdős problems. Thomas Bloom, who runs erdosproblems.com, listed the claims relevant to his site on 8 Oct (blog71). He counts significant claims on 25 Erdős problems (#3, 90, 120, 138, 148, 172, 174, 181, 184, 293, 304, 371, 416, 431, 508, 551, 604, 802, 821, 952, 968, 970, 986, 1083, 1150). These include 12 problems from the FrontierMath Erdős list (#508 only partly), 21 of the 25 have Lean formalizations, and the PDFs total 1,347 pages. He has “not yet looked at in any serious way any of the proofs”. His verdict: the papers are “very poorly written, in the usual AI fashion”, and he expects most to be “(a) correct yet (b) capable of being substantially simplified”, “very rich ‘raw material’”. He flags a gap for #3: the PDF claims quasipolynomial bounds for Szemerédi’s theorem, but the Lean challenge file states only Erdős’s qualitative conjecture, and the Lean code seems to prove a weaker bound. (OfficeChai’s earlier count of 23 came from a first version of the list.)
- Scientific American (8 Oct) gave two pieces. In “The most exciting claims…” (Peter Hall, Joseph Howlett), Hector Pasten (PUC Chile) said of the Siegel-zero result: “This is not just progress on Siegel zeros—it settles the question. This is amazing.” Roland Bauerschmidt (NYU) said of the sigma-model/Yang–Mills-type papers: “The second paper is horrible. If these results had been sent to me by a nobody, I would have deleted the e-mail.” In “Mathematicians marvel, and grumble…”, Michael Douglas (Harvard) said “Yesterday will be the day when a new era of math and mathematical physics began”, and Andrew Sutherland (MIT) said “We should expect some of the proofs to contain mistakes, possibly serious ones”. SciAm also noted that OpenAI has not said which of the ~4,000 problems posed the model failed on.
- Unique Games, first-hand. Scott Aaronson’s “The Mathocalypse” (7 Oct) called the release “surely one of the biggest days in mathematical history”. He noted that the claimed Unique Games proof lands directly on the research of his wife, Dana Moshkovitz. The Decoder (8 Oct) quotes her text message about the proof: “It feels like something written by someone who’s on psychedelics.”
- Fortune (Jeremy Kahn, 7 Oct): Daniel Litt called the results “great for mathematics”. OpenAI gave reasoning summaries and compute estimates for only 10 problems. Its promise to fund “workshops, conferences, and special programs” still had no dates, hosts or budget by 9 Oct.
- Boycott framing. The Decoder (Manuel Uth, Matthias Bastian, 8 Oct) headlined that “some mathematicians call for OpenAI boycott”, meaning the AHM statement. It describes Tao as chairing the AHM. Other reports call the statement the AHM communications group’s text, run as a guest post that is not Tao’s own position (XenoSpectrum, Futurism), and the AHM site lists no officers. Tao’s chairmanship is unverified. Futurism (9 Oct) headlined “Mathematicians React With Fury”.
- Lean’s own reliability. On 9 Oct Thomas Hales posted a guest post on Tao’s blog on what mathematicians should know about Lean. It recalls the July–August 2026 “Summer of Soundness Bugs”, which included an AI-assisted Lean “disproof” of Collatz that passed two kernels, and says Lean proofs need a human audit “to ensure statement fidelity”. See Lean kernel soundness bug #14576. Zvi Mowshowitz’s 9 Oct roundup (“New Math from OpenAI”) collects similar worries, plus Josh Frisch’s “main emotion thus far… has been loss: loss of meaning, loss of purpose”.
- The Conversation (Melissa Lee, Monash, 9 Oct; syndicated by Phys.org): a colleague found one paper so unintelligible that as a journal editor “it would have gone straight into the bin”. She quotes reactions ranging from “obviously the most significant moment in mathematical history” to “a giant pile of turd dumped at our doorstep”, and suggests the release also serves OpenAI’s AGI narrative ahead of its IPO.
- Headlines we could not read (paywalled or blocked): NYT, “‘Breathtaking,’ ‘Devastating’: Mathematics Reels After New OpenAI Release” (8 Oct); The Atlantic, “OpenAI Just Carpet-Bombed Mathematics”; Science, “AI for research mathematics has arrived”; New Scientist, “OpenAI has dumped 722 maths papers – now it must clean up the mess” and “OpenAI mistranslated mathematics into code for its Navier-Stokes proof” (about the Bastounis–Circelli–Hansen critique); The Guardian, “OpenAI’s release of mathematical findings draws concerns from experts”; Sky News, “‘mathpocalypse’”. They are listed in links by title only.
- Misreadings to avoid. Some blogs (e.g. Pasquale Pillitteri, “solved parts of two more Millennium Problems”) present the quasi-RH and the Hodge-for-CM-abelian-varieties results as Millennium Prize progress. Neither claims the Riemann hypothesis or the Hodge conjecture itself, and Lichtman called the release “not the Millennium problem rumor”. Separately, Chinese and SEO sites re-ran older Anthropic results as new during the week (36Kr on θ(p_c)=0, see Claude-written Lean proof claims the dying percolation conjecture θ(p_c)=0 in every dimension; tech-insider on the 67.2% zeta-zeros result, see Claude proves more than two-thirds of Riemann zeta zeros are simple and on the critical line (up from 41.6%)).
Day four (9–10 Oct): the long reads
- The Verge (Robert Hart, 9 Oct) spoke to “more than three dozen mathematicians” (“Staggering”, “Surreal”, “Pure insanity”). Most said understanding the release could take years. Kevin Buzzard found about six results in algebraic number theory that “stood out”, few if any formalized: “I either have to read possibly-not-correct slop, or wait for others to do the same, or wait for someone to formalise them”. Brendan Hassett (Brown): “If this had been written by a person, I wouldn’t spend any more time trying to understand it.” Nalini Joshi (Sydney) noted “short bibliographies”. Jared Duker Lichtman put “tens” of results at Fields-Medal level (progress toward RH, a Hodge special case, the 4D Kakeya conjecture) but said they all come “out of existing methods and techniques”; Álvaro Lozano-Robledo agreed. Scott Armstrong said “there were definitely some targets”: Hong Wang’s area and several Millennium problems, with a cluster in mathematical physics pointing at Yang–Mills and the mass gap. On the damage: Armstrong knew of a group whose programme was “wiped out”, Buckmaster of “three people who had entire research programs obliterated”, Colva Roney-Dougal (St Andrews) of colleagues “whose grant proposals have just been wiped out”, and Francesco Fournier-Facio (Heriot-Watt) said parts of group theory were “bulldozed”. Constantin Kogler (IAS) called it “the most important single moment in the history of mathematics”, and Yang-Hui He (LIMS) compared the year to Euclid’s Elements. The Verge notes OpenAI did not name the model or disclose prompts or the full problem list, against AGMAI’s advice, and did not comment on the record.
- Business Insider (Katherine Li, 10 Oct): Buckmaster: “So we’ve had whole research programs wiped out on Tuesday by their dump”. He also asked whether unpublished grant proposals, often run through LLMs during review, ended up in training data (a question, not evidence). Bryna Kra feared mathematicians will become more secretive. Kontorovich: “I don’t think we have any interest in giving prizes to someone who pressed the button.” OpenAI did not respond. Buckmaster repeated the “chilling effect” argument on MS NOW (Jacob Soboroff, 10 Oct).
- The Independent (Andrew Griffin, 9 Oct) read OpenAI’s blog as “almost apologetic”. It quoted the Proofs and Prompts reactions of Henry Wilton (“If OpenAI wanted to destroy the mathematical community, this would be a great way to go about it”) and Hugo Duminil-Copin (“The shock is immense. I am paralysed”), and Alex Meiburg’s blog on the “intrinsic value” of human comprehension. The Daily Mail, Chosun Ilbo (Duminil-Copin, Martin Bridson: “We should not limit ourselves to following in their footsteps”, Buzzard, Tobias Osborne) and The Decoder (“How much beauty have we lost?”, Sam Hughes; Matt Zaremsky’s “blasted … with TNT”) re-ran the same collection.
- Opinion. In Inside Higher Ed (9 Oct), Ahmed Abbes (CNRS) and Haynes Miller (MIT) called the wave “very much like an occupation by a foreign power” and said advisory committees give labs “a useful cover”. They asked for government-guaranteed equal access to frontier models. Tom Whipple’s Times column (“Has AI just killed maths?”) is paywalled and was not read.
- Older voices. Fields Medallist Maxim Kontsevich told China Media Group (aired 10 Oct): “It’s really faster than I expected”, comparing AI proofs to “winning complicated chess in complicated position”. The interview does not mention OpenAI. Tom Verhoeff (TU/e), who used Claude Opus 5.5, DeepSeek and Aristotle on Lehmer’s permutation conjecture (see Lehmer’s 1965 permutation conjecture proved), told IO+ the OpenAI results look “truly esoteric”.
- Kakeya. 36Kr/新智元 (10 Oct) pointed out that the catalogue’s 4D Kakeya manuscript runs 175 pages and the 3D maximal-function version 97 pages, against Wang and Zahl’s 127-page 3D set proof (Feb 2025). It reports no comment from Wang or Zahl.
- Rehashes, not new findings. “Mathematicians flag translation error in OpenAI’s Navier–Stokes proof” (news.lavx.hu, 10 Oct; Chosun, 9 Oct, from TechCrunch) is the 6 Oct Bastounis–Circelli–Hansen preprint already covered above and in OpenAI claims a Millennium Prize problem. Crypto Briefing’s “OpenAI claims breakthrough in solving Navier–Stokes” and Cape Times’s “Touched on their hypotenuse” (Michael Sherman, 7 Oct) restate the September dispute.
- Tao’s “Math 2.0” lecture (Caltech, 9 Oct) is a separate entry: Terence Tao’s Caltech lecture ‘Math 2.0’.
Follow-up papers by outside mathematicians (8–9 Oct arXiv)
- Krzysztof Burdzy, co-author of the 1999 Burdzy–Werner counterexample to the general hot spots conjecture, posted a ChatGPT-generated note (arXiv 2610.12622) extending OpenAI’s smooth-domain hot-spots theorem to non-smooth simply connected domains. His commentary: “The OpenAI model that generated [3] missed the theorem presented in this note”, and “It is highly surprising that [3] did not cite the spectacular paper of de Dios Pont”, which shows that the conjecture fails for convex domains in high dimensions. Rodrigo Bañuelos produced a near-identical result with ChatGPT at the same time.
- Baiying Liu re-tuned the parameters of the quasi-Riemann proof to get Re s > 0.874957 in place of 7/8, with a Lean formalization (arXiv 2610.12234).
Expert expositions and follow-ups (10–13 Oct arXiv)
Several specialists posted human rewrites of release proofs, the first outside checks of individual families:
- Thompson’s group F (248): Sang-hyun Kim and Yash Lodha, “a human written exposition of OpenAI’s recent proof that Thompson’s group F is nonamenable”, with a generalization; “No part of this writing … was produced by AI” (arXiv 2610.18360; own entry Thompson’s group F is nonamenable).
- Landau–Siegel zeros (003): William D. Banks, explicit constant (1 − β) log q ≥ 0.003, written with Claude and checked line by line (arXiv 2610.15483; own entry William Banks writes out OpenAI’s Landau–Siegel proof with explicit constants).
- Deligne–Drinfeld conjecture (008): Thomas Willwacher, “a simplified and more human-friendly version of the proof … No originality is claimed”, written “for the convenience of myself and fellow human readers who found [Ope26] hard to follow”; AI tools helped him understand and restructure the argument (arXiv 2610.15592).
- Builds on release results: Goldbach exceptional set X^{7/12+ε} from the 7/8 half-plane (Pintz & Althöfer, 2610.15559); 7/8 → 0.8748867 (Sang, 2610.15291); unconditional Ellenberg–Venkatesh ℓ-torsion bounds (Frei, Sofos & Widmer, 2610.18491); natural length equals the exact Hausdorff measure for SLE (Zhan, 2610.15066, family 230); truth-table degrees rigid under Baire-measurable automorphisms, adapting the Turing-degree rigidity proof (Kjos-Hanssen, 2610.16059, family 241, with Claude and a Lean proof by Aristotle); a dynamical two-point Chowla generalization (Wang & Yi, 2610.16394, family 007); Erdős #824 conditionally on the “Weighted dilation graphs” preprint (Li & Zhuang, 2610.15674, Lean by ChatGPT/Codex).
- 14 Oct listing: follow-ups to the release (Penrose/Bray et al., Hardy–Littlewood–Chowla, Shlyakhtenko and others) are in the mid-October catalogue (moved there on 2026-10-15 to keep this page within the size budget).
- Reading the release as a source of tools: Ying Xie, “What the Machine Invented” (arXiv 2610.17090), extracts new lemmas and constructions from 38 families (extraction with Claude Fable 5.1). The qtBench paper (Cainelli, Luccioli, Iraci, Paolini, 2610.16555) reports OpenAI researchers’ “promising candidate solutions” to several of its 29 algebraic-combinatorics problems and asks whether “AI discovery at scale” is ethical and how results should be released.
Sources
80 sources from 53 sites. Numbers match the chips in the text.
80 sources: 13 primary, 39 press, 28 reactions
Primary
- OpenAI on X: ‘We’re releasing a broad range of new mathematical results…’x.com, official
- GitHub: openai/math (722 manuscripts, Lean library)github.com, code
- OpenAI: Sharing AI progress in mathematicsopenai.com, official
- OpenAI Research Catalog (overview.pdf)github.com, paper
- Lean formalization catalogue (formalization.yaml)github.com, code
- AGMAI: On OpenAI’s Release of Mathematical Results (6 Oct)agmai.org, official
- Independent Lean re-check of the 7/8 zeta proof (davegoldblatt)github.com, code
- Douglas Colkitt: conditional κ = 2^-59 improvement of family #109 (integer multiplication), with OpenAI Codexgithub.com, code
- Wang & Zahl (2025): the Kakeya set conjecture in three dimensions (arXiv 2502.17655)arxiv.org, paper
- openai/math history.md: 7 Oct withdrawals, fixes and formalizationsgithub.com, official
- Dan Roberts (OpenAI) on X: 6 new Lean formalizations, 19 modifications, 3 withdrawalsx.com, official
- Bastounis, Circelli & Hansen: Navier–Stokes lost in translation (arXiv 2610.08144)arxiv.org, paper
Show all 13 sources
- openai/math README (‘~42% top-line results formalized’)github.com, official
Press
- NBC News: Some OpenAI math discoveries are ‘jewels,’ if verified. But mathematicians are wary of ‘slop.’ (11 Oct)nbcnews.com, press
- WIRED: OpenAI is pissing off a bunch of mathematicians againwired.com, press
- NYT: OpenAI Releases Findings on 377 Math Problems, Further Roiling Fieldnytimes.com, press
- Scientific American: OpenAI unleashes hundreds more math resultsscientificamerican.com, press
- WSJ: AI Solved One Math Problem and Everyone Freaked Out. It Just Cracked Hundreds More.wsj.com, press
- Quanta: As AI Closed In on ‘Unique Games’ Proof, Researchers Raced to Beat the Machinesquantamagazine.org, press
- TechCrunch: OpenAI’s math solutions aren’t meeting the field’s standards yettechcrunch.com, press
- Implicator: OpenAI posts 372 AI math results, then withdraws three papers over a sign errorimplicator.ai, press
- Scientific American: The most exciting claims from OpenAI’s heap of new proofs (8 Oct)scientificamerican.com, press
- Scientific American: Mathematicians marvel, and grumble, at OpenAI’s trove of new results (8 Oct)scientificamerican.com, press
- Fortune: OpenAI publishes solutions to more than 370 outstanding math challenges (7 Oct)fortune.com, press
- The Decoder: Some mathematicians call for OpenAI boycott after AI-generated proofs flood their field (8 Oct)the-decoder.com, press
Show all 39 sources
- Futurism: Mathematicians React With Fury as OpenAI Releases Hundreds of New AI-Generated Proofs (9 Oct)futurism.com, press
- The Conversation (Melissa Lee): Is this the ‘mathocalypse’? (9 Oct)theconversation.com, press
- NYT: ‘Breathtaking,’ ‘Devastating’: Mathematics Reels After New OpenAI Release (paywalled, not read)nytimes.com, press
- The Atlantic: OpenAI Just Carpet-Bombed Mathematics (not read)theatlantic.com, press
- Nature News (Davide Castelvecchi): OpenAI posts 700 maths preprints online: mathematicians are up in arms (7 Oct; standfirst: ‘The firm used an unreleased model to tackle hundreds of high-profile problems.’)nature.com, press
- Nature News (Kaia Glickman): Will AI scoop your science? Some researchers see a gloomy future (8 Oct)nature.com, press
- Science: AI for research mathematics has arrived (not read)science.org, press
- New Scientist: OpenAI has dumped 722 maths papers – now it must clean up the mess (not read)newscientist.com, press
- The Guardian: OpenAI’s release of mathematical findings draws concerns from experts (not read)theguardian.com, press
- Sky News: OpenAI sparks ‘mathpocalypse’ (not read)news.sky.com, press
- Axios: OpenAI’s math breakthrough points beyond math (8 Oct)axios.com, press
- The Verge (Robert Hart): ‘Pure insanity’: Mathematicians will need years to make sense of OpenAI’s latest drop (9 Oct)theverge.com, press
- The Independent (Andrew Griffin): Mathematicians are very, very angry at OpenAI. Here’s why (9 Oct)independent.co.uk, press
- Business Insider via Yahoo (Katherine Li): NYU professor says OpenAI’s mass release wiped out early-career researchers’ projects (10 Oct)tech.yahoo.com, press
- Inside Higher Ed (Ahmed Abbes, Haynes Miller): The AI Invasion Has Come for Mathematics (opinion, 9 Oct)insidehighered.com, press
- Bastille Post: Fields Medalist Kontsevich says speed of AI advances in math took him by surprise (CMG interview, 10 Oct)bastillepost.com, press
- Chosun Ilbo: AI’s math breakthroughs frustrate Fields Medalists (9 Oct)chosun.com, press
- Daily Mail: The AI ‘mathpocalypse’ (9 Oct)dailymail.com, press
- The Times (Tom Whipple): Has AI just killed maths? (column, 9 Oct; paywalled, not read)thetimes.com, press
- Retraction Watch: OpenAI withdraws three preprints a day after releasing 722 manuscripts on unsolved math problems (Oct 8)retractionwatch.com, press
- franceinfo via RTÉ: Why mathematicians are in ‘crisis’ due to AI advancements (Oct 12)rte.ie, press
- Cult of Mac via Yahoo Tech: A mathematician shared his ideas with AI. Now, he’s keeping his research away from OpenAI (Oct 12)tech.yahoo.com, press
- The Economist: AI’s mathematics blitz pushes science into strange new territory (Oct 12; paywalled, not read)economist.com, press
- Quartz: OpenAI’s math data dump wiped out early-career researchers’ programs, professor says (Oct 12; not readable, 403)qz.com, press
- Le Monde: OpenAI’s latest blow to mathematicians: ‘They have annihilated my field of research’ (Oct 12; not read, automated access blocked)lemonde.fr, press
- IBT UK: Fields Medallist says OpenAI’s 700 AI-generated maths papers have ‘annihilated’ his research field (Oct 13)ibtimes.co.uk, press
- KQED Life, Automated (28 Oct): Is math a canary in the AI coal mine? With Bryna Krakqed.org, press
Reactions
- Xena Project: …and the proof is checked in Lean (12 Nov 2026)xenaproject.wordpress.com, discussion
- Post-Cutoff: all 373 result families, graded and linked (audit 2026-10-29)postcutoff.com, discussion
- Hacker News discussion (652+ points)news.ycombinator.com, discussion
- Alex Kontorovich on X: ‘instant Fields Medal’x.com, discussion
- Levent Alpöge on Xx.com, discussion
- Daniel Litt on Xx.com, discussion
- Joe Bebel on X (Unique Games)x.com, discussion
- Steven Strogatz on X (matrix multiplication)x.com, discussion
- Claude Opus 5.5 reads the repo (posted by @ereliuer_eteer)x.com, discussion
- Sam Altman on X: ‘a new era of discovery’x.com, discussion
- Terence Tao on Mathstodon (pre-release)mathstodon.xyz, discussion
- Gil Kalai: Sharing AI progress on mathematics – amazinggilkalai.wordpress.com, discussion
Show all 28 sources
- Elliot Glazer on X: Astra check finds the abelian-eightfolds paper flawedx.com, discussion
- HN: OpenAI withdraws three mathematical results (224 pts)news.ycombinator.com, discussion
- AHM Statement on OpenAI’s October 6 Release of Mathematical Documentsahmath.org, discussion
- AHM statement as guest post on Terence Tao’s blog (7 Oct)terrytao.wordpress.com, discussion
- Raghu Meka (guest post on Tao’s blog): The barriers of perceptionterrytao.wordpress.com, discussion
- Álvaro Lozano-Robledo (guest post on Tao’s blog): What should we tell our students?terrytao.wordpress.com, discussion
- Thomas Bloom: Erdős problems in the OpenAI proofs (8 Oct)erdosproblems.com, discussion
- Scott Aaronson: The Mathocalypse (7 Oct)scottaaronson.blog, discussion
- Zvi Mowshowitz: New Math from OpenAI (9 Oct)thezvi.substack.com, discussion
- Thomas Hales (guest post on Tao’s blog): What mathematicians should know about the Lean Theorem Prover (9 Oct)terrytao.wordpress.com, discussion
- Developers Digest: OpenAI Math Repo, what the Lean proofs do and don’t verify (8 Oct)developersdigest.tech, discussion
- Proofs and Prompts: 100+ reactions to 100+ solutions (Oct 8)proofsandprompts.com, discussion
- Joshua Gans: OpenAI drops a bomb on maths (Oct 7)joshuagans.substack.com, discussion
- Sophie Huiberts (guest post, Terence Tao’s blog): This is normal bad work (Oct 12)terrytao.wordpress.com, discussion
- Ben Recht (arg min): Faster Fourier Transformsargmin.net, discussion
- Hacker News: A look at the math that OpenAI claims computes faster Fourier transformsnews.ycombinator.com, discussion
Also cited in the text
- x.com/littmath/status/2107616410897678765
- x.com/jdlichtman/status/2107609204722892924
- x.com/ChrisPeikert/status/2107617657415782731
- x.com/Kiwamu_Watanabe/status/2107621501193408756
- x.com/Isaac__kim/status/2107607429437657563
- x.com/KenOno691/status/2107612779264000261
- x.com/gdb/status/2107640886482104524
- x.com/thsottiaux/status/2107597780352950624
- x.com/AcerFur/status/2107606747972309163
- github.com/nealmcb/openai-math-lean-tooling
- x.com/kimmonismus/status/2107597320028065793
- x.com/Hesamation/status/2107600759953179030
- x.com/cgtwts/status/2107597257684169010
- x.com/khoomeik/status/2107603513438335334
- x.com/petergostev/status/2107603534308864146
- de Grey (2018)arxiv.org
- x.com/0xdoug/status/2107917690568327225
- Mathstodonmathstodon.xyz
- arXiv 2610.12622arxiv.org
- arXiv 2610.12234arxiv.org
- arXiv 2610.18360arxiv.org
- arXiv 2610.15483arxiv.org
- arXiv 2610.15592arxiv.org
- 2610.15559arxiv.org
- 2610.15291arxiv.org
- 2610.18491arxiv.org
- 2610.15066arxiv.org
- 2610.16059arxiv.org
- 2610.16394arxiv.org
- 2610.15674arxiv.org
- arXiv 2610.17090arxiv.org
- 2610.16555arxiv.org
Changes
- Added the Xena Project’s 12 Nov post on what ‘checked in Lean’ means (science agent; found via HN)
- Science agent run 2: own entries for families 288 (Kadison similarity), 369 (hot spots), 362 (Vlasov–Maxwell), 156 (Borsuk in dimension 9), 155 (periodic tiling in dimension 3), 008 (Deligne–Drinfeld, confirmed), 235 (random k-SAT threshold, with Carenini’s priority) and 311 (Chenglong Ma’s provenance claim)
- Noted the Lean formalization of L = RL = BPL (7 Oct update) and linked the new abundance and L = BPL entries (science agent)
- OpenAI’s 26 Oct repository update adds two manuscripts with Lean formalizations, ‘The Exponential Time Hypothesis is False’ (family 378, to be examined by a working group led by Impagliazzo, Paturi and Williams) and ‘Local smoothing for the Euclidean wave equation’ (family 079, all dimensions, with Kakeya corollaries), plus a Lean formalization of the rational Hodge conjecture for CM abelian varieties (family 032) and restored C++ sources for the MUB-in-dimension-six computation; separate entries for the first two
- Added Neal McBurnett’s whole-library re-check (all 416 Comparator configs pass Lean’s kernel; nanoda 401/416, failures traced to nanoda bugs, issue 55; QuasiRiemannHypothesis accepted by both kernels). Prompted by Gemini citing ‘forum posts’ re-verifying the 7/8 proof on nanoda; no forum thread found
- Video added: SAIR interview with Terence Tao and Yu Deng (Oct 19), in which Tao says commercial labs treat maths benchmarks as byproducts and that about a dozen problems in the release interest him but poor exposition makes reading them a low priority
Show all 32 changes
- Added video(s) rest-is-science-mathematicians-canary-ai-coal-mine, ai-search-700-math-breakthroughs-ai-news
- Added video ellie-sleightholm-openai-latest-maths-breakthrough
- Follow-ups moved to the mid-October catalogue (page size budget); added follow-ups from the 14 Oct arXiv listing (Penrose/Bray et al. with provenance concerns, Hardy–Littlewood–Chowla, Shlyakhtenko, ℝ³ chromatic numbers, Solomon–Yau application, 3D hot spots, Visser, Castillo–Gordillo-Núñez critique); sweep 2026-10-15
- Added expert expositions (Kim–Lodha on Thompson F, Banks on Siegel zeros, Willwacher on Deligne–Drinfeld) and follow-up papers of 10–13 Oct (arXiv)
- Added the Le Monde headline quote (article not readable) and IBT UK’s summary (Duminil-Copin, Gowers) (Google News)
- Related: the catalogue that lists this result is now October 2026: the arXiv stream of conjectures settled with disclosed AI help continues (the October blocks were split out of the summer catalogue)
- Added Ben Recht’s critique of the faster-Fourier-transform manuscript as a galactic algorithm (HN front page)
- Retraction Watch on the withdrawals (Townsend, Sutherland), Huiberts’s guest post on Tao’s blog, franceinfo/RTÉ and Cult of Mac second-wave coverage; Economist and Quartz linked but not read
- Added follow-up papers by Burdzy (hot spots, with criticism of the OpenAI manuscript) and Liu (quasi-RH constants)
- Science share of our automated search: family 017 (irrationality exponent of π) now has its own entry; outside papers building on the release appeared on arXiv on 8 Oct: Gaofeng Huang’s holomorphic cancellation counterexample cites the release’s Zariski-cancellation preprint (arXiv 2610.12581), and Qiang Gao used Claude Opus 5.5 to raise the constant in the release’s Laughlin-gap proof from 0.046 to 0.219 (arXiv 2610.12700)
- Added NBC News (11 Oct, opening only). Korea Daily (‘US scholar criticizes rushed proofs’), Soap Central (‘NYU professor … incentive system’, a rehash of the 10 Oct Business Insider piece already cited), Vocal and ascendants.in (‘boycott calls’) add no new facts
- Added Nature News coverage (Castelvecchi, 7 Oct; Glickman, 8 Oct) from the Nature feed; earlier runs had skipped both (titles and standfirsts only, articles not read)
- Day-four section (Fork 4): The Verge long read (read in full: Buzzard, Hassett, Joshi, Lichtman, Armstrong, Roney-Dougal, Fournier-Facio, Kogler, He), Business Insider (Buckmaster, Kra, Kontorovich), The Independent, Daily Mail, Chosun, Inside Higher Ed op-ed (Abbes & Miller), Kontsevich CMG interview, Verhoeff (IO+), 36Kr Kakeya page counts; the Navier–Stokes ‘translation error’ headlines flagged as the known Bastounis et al. preprint; linked Tao’s Math 2.0 entry and the MS NOW video
- Added the Proofs and Prompts collection of signed mathematician reactions (Tao, Scholze, Agol, Green, Wilton) and Joshua Gans’s essay
- Explained the ‘42% Lean’ headlines (README figure = 300/719 formalized, not a pass rate), added Axios’s ‘beyond math’ framing and flagged 36Kr’s ‘Tao reposts boycott’ headline as the 7 Oct AHM guest post. The Verge (‘Pure insanity’, 9 Oct), The Atlantic, The Times and the NYT’s ‘Breathtaking’ piece remained unreadable (blocked or paywalled)
- Added overlaps with Engel–Mauri (SYZ), Qin–Sun (Hilbert number), Ky (Hellinger) and the September Courtade–Kumar proofs, and Karagila’s refusal to review the Partition Principle paper
- Day-three section: Bloom’s 25-Erdős-problem list and #3 Lean gap, SciAm expert quotes (Pasten, Bauerschmidt, Douglas, Sutherland), Aaronson/Moshkovitz, Fortune, Decoder boycott framing (Tao-chairs-AHM claim flagged unverified), The Conversation, Hales on Lean soundness bugs (new entry Lean kernel soundness bug #14576), paywalled headlines listed, Millennium misreadings corrected
- Day-two section: 7 Oct withdrawals (3 Hodge-family papers, sign error), 14 repairs, 300/719 formalized; Glazer and Roberts posts; rival proofs (exact overlaps, LeBrun–Salamon, Minzer); AHM boycott call, Tao guest posts, Navier–Stokes autoformalisation critique; Drake bunker-mode link
- Linked Doug Colkitt’s conditional κ = 2^-59 follow-up on family 109 (new entry Outside researchers using AI crowdsource a tighter exponent in OpenAI’s integer-multiplication result #109, from 2^-182 to about 2^-8.4 (conditional))
- Science block: kind: batch (many results in one entry) (not a long-open problem; listed with other AI science results, not in the open-problems count)
- Added high-reach X posts (kimmonismus, Hesamation, Rohan Pandey, Peter Gostev, cgtwts reach), corrected the viral Kakeya R^3 claim against family 074 and Wang–Zahl 2025, stated family 158’s Hadwiger–Nelson claim exactly; WSJ read via HT Tech syndication (title fixed; OpenAI’s Anthropic motive and its reply to AGMAI)
- Summary and science result now give the manuscript dates (a model reading /science/ concluded they were written in ‘one afternoon’)
- Made the attribution of the Opus 5.5 text explicit (presented by the poster as model output; see the post file)
- Added Claude Opus 5.5’s read of the repo (@ereliuer_eteer) and our own checks of the comparator files (405 challenges; HarmonicGrowth axiom; FinalAssemblyUnconditional)
- Added first reactions (X, Mathstodon, blogs), press and the independent Lean re-check
- Filed from the repository README, overview.tex, CONTENTS.md, lean/formalization.yaml and Lean scope notes, the OpenAI blog post and the AGMAI statement