MATERIAL: "722 solved problems", Erdős problems and the "mathematicians are cooked" talk

Research notes for a writer, compiled 2026-11-01. Not an article. [S#] = numbered source in the last section. "Unverified" = primary text not reached.

The claim, corrected

The owner's framing: "OpenAI published those 722 solved problems". What the record says:

• 722 is a count of manuscripts, not of solved problems. On 6 Oct 2026 OpenAI pushed github.com/openai/math: 722 machine-written manuscripts in 372 "result families" across 17 fields, every paper authored "OpenAI", plus a blog post "Sharing AI progress in mathematics" [S1][S2][S3][S10]. The README now reads 721 / 373 after withdrawals [S2]; TechCrunch counted 719 on 8 Oct [S15].
• Only a small slice is Erdős problems. Six family titles [S3]; Thomas Bloom (erdosproblems.com) counted "significant claims relevant to 25 Erdős problems" [S5].
• "Solved" is OpenAI's claim, not a community verdict. README: "Some of the unformalized results could have issues" [S2]; 3 withdrawn and 14 revised within a day [S4]; Bloom: "I have not yet looked at in any serious way any of the proofs in this release"; "none of these should be viewed as 'closing the problem'" [S5].
• "Mathematicians are cooked" appeared in no fetched source; origin unverified, social-media shorthand rather than a mathematician's quote.

What happened (dated timeline)

Act 1: the October 2025 overclaim (the precedent)

• 12 Oct 2025: Sébastien Bubeck (OpenAI) posts that GPT-5 "just solved Erdos Problem #339" ... "by realizing that it had actually been solved 20 years ago" [S19].
• 17 Oct 2025: Mark Sellke posts a tally of Erdős problems listed as open for which GPT-5 found published solutions [S18]. Kevin Weil (then OpenAI VP of science) posts: "GPT-5 just found solutions to 10 (!) previously unsolved Erdős problems, and made progress on 11 others. These have all been open for decades." [S18]
• 17 Oct 2025: Bloom replies: "this is a dramatic misrepresentation." "GPT-5 found references, which solved these problems, that I personally was unaware of." "The 'open' status only means I personally am unaware of a paper which solves it." [S19]
• 18 Oct 2025: Weil deletes the post; Bubeck: "I deleted the post, I didn't mean to mislead anyone obviously, I thought the phrasing was clear, sorry about that." Demis Hassabis: "this is embarrassing". Yann LeCun: "Hoisted by their own GPTards" [S18][S19].
• 20 Nov 2025: OpenAI's paper "Early science acceleration experiments with GPT-5" has a section "Erdős problems (part 2/2) – Mehtaab Sawhney and Mark Sellke"; abstract: "four new results in mathematics (carefully verified by the human authors)" [S20]. Scientific American later summarised the Sawhney/Sellke work as digging up "lost solutions to nine other Erdős problems, plus partial solutions to 11 more" [S21].

Act 2: AI solutions become routine on erdosproblems.com (Nov 2025 to Aug 2026)

• 22 Nov 2025: Tao, on Problem #367: "AI assistance is now becoming routine." [S30]
• 25 Dec 2025: Problem #333 posted as AI-solved on Christmas morning, later retracted because Erdős had already solved it in 1977 [S25].
• 4 Jan 2026: GPT-5.2 Pro proves Problem #728; Aristotle (Harmonic) formalises it; Tao: solved "more or less autonomously" [S29]. 14 Jan: Tao on the share of open Erdős problems simple enough for current AI with minimal human help: "(My guess is that this proportion is on the order of 1-2%.)" [S29]
• 29 Jan 2026: Google DeepMind: Gemini evaluated "700 conjectures labeled 'Open' in Bloom's Erdős Problems database"; 13 addressed, "5 through seemingly novel autonomous solutions", "8 through identification of previous solutions in the existing literature" [S28].
• 12 Feb 2026: Scientific American, citing Tao's tracking page: "AI tools have helped transfer about 100 Erdős problems into the 'solved' column" since October; "The bulk of this assistance has been a kind of souped-up literature search." [S21]
• 20 May 2026: OpenAI announces an internal model's disproof of Erdős's 1946 planar unit distance conjecture (Problem #90); Alon, Wood and Bloom back it in OpenAI's companion remarks [S31]. 2 Jun: Leiden Declaration calls for "tight guardrails" (1,590 signatures by 5 Jun) [S32]. 29 May: Problem #1196 solved with GPT-5.4 Pro [S33].
• 30 Jun 2026: Tao's wiki "AI contributions to Erdős problems" frozen [S22][S23]. 3 Aug 2026: Quanta: 565 solved / 652 open on the site [S25].
• 1 Sep 2026: Epoch AI and Bloom launch FrontierMath Erdős (68 open problems): best model 3%, all others 0% [S26][S27] (details in table).

Act 3: the bulk release (Sep to Oct 2026)

• Aug 2026: OpenAI convenes about 40 mathematicians, hinting at hundreds of solved conjectures (Wikipedia, secondary) [S9].
• 8 Sep 2026: OpenAI claims finite-time blowup for 3D Navier–Stokes with a multi-agent system (at one point 10,000 sub-agents). Tristan Buckmaster (NYU) and Levent Alpöge (Anthropic) say the model used the same line of attack as their in-progress work; Buckmaster says Bubeck offered priority only if Alpöge's name were removed; Bubeck: "We did not use their prompts or proofs to prompt our models or direct our agents" [S17].
• 11 Sep 2026: declaration "A Severe Misalignment of AI in Mathematics" signed by Fields Medalists (28 per Wikipedia, 25 per TechCrunch) [S9][S14].
• 21 Sep 2026: OpenAI says "the same internal model has resolved more than 100 additional open problems across most areas of mathematics" and announces the Advisory Group on Mathematics and AI (AGMAI), nine unpaid mathematicians hosted at IAS (incl. Gowers, Hairer, Witten, Wood) [S13][S14].
• 29 Sep 2026: AGMAI guidelines: release "the name of the model, prompts, chains of thought, number of attempts, compute used, and costs"; use "scholarly repositories that are not controlled by the posting AI lab"; "Proofs should be formalized where possible, unless it leads to unacceptable delays." [S13]
• 1 Oct 2026: Kevin Buzzard: "I personally call for OpenAI to simply dump their collection of theorems upon the world" [S37].
• 6 Oct 2026, 6 p.m. EDT: the release [S10]. Same day: Bloom posts "Changes": comments and proof claims frozen, "The site will not display the statuses of any problems (e.g. open, solved, etc.)", solved count and percentage removed; "The main point is to disincentivise people who are simply glory-chasing"; "Use AI as you like -- but do not abandon mathematics as a human activity." [S6] Same day, 18:00 UTC: Tao's four-post thread (see quotes) [S7].
• 7 Oct 2026: OpenAI withdraws three manuscripts over "a sign error" (one argument and two dependent algebraic-geometry papers), revises 14, corrects one citation; Dan Roberts (OpenAI): will "continue to update the repo with new formalizations and with any errata we notice." Association for Human Mathematics (AHM): "Mathematicians did not ask for this work to be done." "Releasing over 700 files at once is not a demonstration of scholarship, but a demonstration of power."; urges members to "discontinue their work with OpenAI" [S4][S12].
• 8 Oct 2026: Bloom's "Erdős problems in the OpenAI proofs" lists the 25 problems [S5]. TechCrunch: 719 manuscripts, 10 with chain-of-thought summaries, 42% not formalised [S15].
• 9 Oct 2026: Quanta: Johan Commelin (Mathlib) says "some of the formalizations [appear to] prove weaker statements than what's claimed in the paper" (not confirmed firsthand); AHM, "over 800 mathematicians", calls for a boycott [S11].
• 10 Oct 2026: Tao's "Math 2.0" post and Caltech slides [S8].
• 15 Oct 2026: OpenAI adds to its post: "$5 million to a new Simons Foundation initiative to help mathematicians examine, understand, and build on results produced by AI" [S1 via mirror].
• 21 Oct 2026: a separate site, erdosproblems.ai (Plasma AI), shows 494 open / 110 claimed / 617 solved; same on 1 Nov [S23][S24].

What "solved" meant, in numbers

The 722 release as a whole

• Problems posed to the model: about 4,000; result families kept: 372 (now 373); manuscripts: 722 (now 721); compute: about three hours of "ChatGPT Pro thinking compute" per result on average; 10 reasoning summaries published; no model name, no prompts beyond those 10 [S2][S3][S15].
• Formalisation share, by source (different denominators, not reconcilable): README "about 42% of top-line results are formalized in Lean" [S2]; 235 of 372 families have a Lean statement, but formalization.yaml covers 162 papers and is marked "review: unchecked" [S3]; Retraction Watch: "roughly 50% of results released unconfirmed" [S4]; Science (snippet): 300 of 719 [S16]; TechCrunch: 42% not formalised [S15].
• Errors in the first 48 hours: 3 withdrawn, 14 revised, 1 citation fixed [S4]. A blogger found no sorry, new axioms or native_decide in the Lean, but did not build it [S3].
• Lean checks the stated theorem, not whether it matches the named problem: the Erdős #3 ($5,000 prize) family is listed "proved" but the Lean covers only the reciprocal-sum version [S3]; Bloom: the Lean "appears to prove the weaker bound" [S5]; see also Commelin [S11].

The Erdős slice

• 25 Erdős problems: 3, 90, 120, 138, 148, 172, 174, 181, 184, 293, 304, 371, 416, 431, 508 (partial), 551, 604, 802, 821, 952, 968, 970, 986, 1083, 1150. "21 out of the 25 proofs have been formalised in Lean." 12 are on the FrontierMath Erdős list [S5].
• Bloom's position: not checked, not an endorsement, "does not mean to make any judgements about novelty or attribution"; expectation that "the vast majority of them are (a) correct yet (b) capable of being substantially simplified"; #986: "the original question was answered already by Bradač" [S5].
• Accepted by the list's maintainer today: none, because the list no longer shows statuses at all since 6 Oct 2026 [S6]. The last public count was 586 solved (48%) on 5 Oct 2026 [S23].

AI-attributed Erdős solutions before the release (Tao's wiki, frozen 30 Jun 2026) [S22][S23]

• Categories (exact): 1(a) AI standalone, "Literature presence: comparable literature unknown. Human involvement: non-significant."; 1(b) AI alongside literature, "comparable literature discovered afterwards"; 1(c) AI building on literature, "comparable literature known beforehand"; 1(d) AI collaborating with humans, "Human involvement: significant."; 2(a) literature search, 2(b) formalization, 2(c) rewriting, 2(d) computation. Notice: "Sections 1(a)-1(d) are equally important." "This page is not a benchmark."
• Full solutions per section (Papex's tally of the wiki tables): 1(a) 18, 1(b) 14, 1(c) 5, 1(d) 35, 2(a) 31; 69 distinct problems with at least one AI-led full solution [S23]. Row counts incl. partial progress: 1(a) 57 unique problems, 1(b) 24, 1(c) 37, 1(d) ~133, 2(a) ~84, 2(b) ~185 [S22].
• Reading: about 69 AI-led full solutions out of roughly 1,200 problems by mid-2026; literature search is the largest single category (31 full), formalising existing proofs the busiest (~185 rows).

How hard were they?

• Tao, Jan 2026: order of 1-2% of open Erdős problems are simple enough for current AI with minimal human help [S29].
• Epoch/Bloom, Sep 2026: "Many Erdős problems never attracted much attention from mathematicians"; on the 68 curated hard ones, best model 3%, all others 0%; Bloom estimated about 3 to 5 problems of that calibre had been solved by AI as of Aug 2026; GPT-6 Astra solved 5 more off-benchmark at over $220,000 of compute vs about $20,000 for the benchmark run [S26].
• On #90: Bubeck: "this proof isn't exactly the spark of genius that we see sometimes in mathematics."; correct in 50% of trials, data not released [S32].

The site itself

• Launched 28 May 2023 with "just over 200 problems"; "The site now hosts 1221 problems, over 9000 comments, and almost 2000 registered users."; 10,000 to 25,000 unique visitors a day [S6]. 565 solved / 652 open on 3 Aug 2026; 111 open-to-solved changes in 2024 + Jan-Aug 2025 (pre-AI baseline, includes rediscoveries) [S25]. 1,217 listed / 652 unsolved, Aug 2026 [S26]. Last count 586 solved, 5 Oct 2026 [S23]. Bloom published no per-year "solved" series; the 111 figure is the only pre-AI rate found.

What mathematicians said (exact, with links)

• Thomas Bloom, 17 Oct 2025: see timeline (x.com/thomasfbloom/status/1979254235075059732, via [S19]). 20 May 2026: "AI is helping us to more fully explore the cathedral of mathematics we have built over the centuries" [S31]. Jun 2026: "It could be right. It could be nonsense. Who's going to be able to check this?" [S32]. 6 Oct 2026: "I think it is against the spirit of Erdős to regard any problem as 'closed'" [S6].
• Terence Tao, 6 Oct 2026 (mathstodon.xyz/@tao/117395268889583356, 18:00:44 UTC): "This process cannot be easily reversed" ... "solutions to open problems are now being harvested at large scale in an unsustainable fashion" ... "leaving entire fields of mathematics much less fertile than when such problems were solved in the traditional "Math 1.0" fashion". Next post (…325940185): ""Math 1.0" placed a premium on being the first to solve an open problem" ... "optimized to the point of unsustainability" ... the community will need to "re-evaluate its criteria for education, publication, and career advancement" [S7]. Second post of the thread notes some open directions are being withheld for fear of being scooped [S7].
• Tao, Sep 2026: "The indiscriminate strip-mining of open problems for solutions may destroy the ecosystem" [S17]. Oct 2026 (TechCrunch): "Problems are being solved autonomously by AI prompters who have no interest in the broader field itself" [S15]. 10 Oct 2026, Math 2.0: against "the excessive emphasis on automated solving of open problems by AI" [S8].
• Tao, 14 Jan 2026: "(My guess is that this proportion is on the order of 1-2%.)" (mathstodon.xyz/@tao/115891257393270694) [S29]. Tao, 22 Nov 2025: "AI assistance is now becoming routine." [S30]
• Andrew Sutherland (MIT): "We should ask for receipts." and that one-shot claims should be treated as unverified "Until and unless they release the model and people can replicate their results" [S10]. Daniel Litt (Toronto): "I see no reason why we should ask the company to keep them secret from us" ... "To me, it's going to be a good thing for mathematics." [S10]
• Melanie Wood (Harvard, AGMAI): "there is not human understanding of them at the point of release, and now the work begins" [S15].
• Johan Commelin (Mathlib): "some of the formalizations [appear to] prove weaker statements than what's claimed in the paper"; Mihalis Dafermos: "the papers are written in a very synoptic style" [S11].
• Alex Townsend (Cornell): OpenAI "should have announced the manuscripts that were lean verified first" [S4]. Kevin Buzzard, 1 Oct 2026: "Lean will not accept proof by authority or proof by intimidation" [S37].
• Hugo Duminil-Copin (Fields Medal): "They have annihilated my field of research" (Le Monde via Wikipedia, secondary) [S9].

On "without letting anyone know" (closest sourced items)

• Science (snippet, not fetched): Kogler found in the release a proof of the exact overlaps conjecture, "a result he and a colleague had been working flat out to write up since cracking it 10 days earlier" [S16].
• Buckmaster/Alpöge priority dispute on Navier–Stokes, Sep 2026 [S17]. Tao: some open directions are being withheld for fear of being scooped [S7].
• OpenAI consulted AGMAI, which it says "recommended releasing results without waiting for 'full formalization'" [S4]; AGMAI posted its own note on 6 Oct [S13]; OpenAI told Scientific American it is not bound by AGMAI [S10].
• No source says OpenAI contacted authors of individual open problems, nor that it had to; the lists are public.

Data points table (benchmark, score, date, source, caveat)

What is unsure / could not verify

• OpenAI's blog text: openai.com returned 403; its two sentences come from a mirror (CellCog) and an RSS timestamp [S1]. Verify before quoting.
• Science article [S16]: 403; Kogler and "300 of 719" are from search snippets.
• Nature's #90 piece and Tao's "job description is changing" line: paywalled, not fetched.
• Origin of "mathematicians are cooked": not found.
• Tao's Oct 2025 remark on literature search ("most productive near-term adoptions of AI...") exists only as an X repost; primary not located.
• The 1(a)-1(d) counts are third-party hand tallies of wiki tables; the wiki gives no totals.
• Lean coverage of the release: five figures from five sources; none authoritative.
• 28 vs 25 Fields Medalists on the Sep 2026 declaration: sources disagree.
• Wikipedia used only for the timeline skeleton; not all of its citations were fetched.
• No published verdict yet on whether any of the 25 Erdős claims is wrong or already known, beyond #986 (Bradač) and the #3 weaker-bound issue.

Sources

1. https://openai.com/index/sharing-ai-progress-in-mathematics/ — "Sharing AI progress in mathematics", OpenAI, 6 Oct 2026 (update 15 Oct). Headline release; $5M Simons update. Read only via mirror https://www.cellcog.ai/blog/openai-math-results/.
2. https://github.com/openai/math — README, OpenAI, Oct 2026. 721/373, ~4,000 problems, ~3 h compute, error caveat, ~42% formalised, Apache-2.0.
3. https://quanglee.dev/blog/openai-722-math-manuscripts/ — quanglee.dev, Oct 2026. Clone-based figures: 722/372/17 fields, 6 Erdős families, 235 Lean statements, 162 in formalization.yaml, #3 caveat; Tao/Sutherland/Litt quotes.
4. https://retractionwatch.com/2026/10/08/openai-withdraws-preprints-722-manuscripts-unsolved-math-problems/ — Retraction Watch, 8 Oct 2026. 3 withdrawn, 14 revised, Roberts, AHM, Townsend, ~50% unconfirmed.
5. https://www.erdosproblems.com/forum/thread/blog:10 — Bloom, "Erdős problems in the OpenAI proofs", 8 Oct 2026. 25 problems, 21 formalised, non-endorsement.
6. https://www.erdosproblems.com/forum/thread/blog:9 — Bloom, "Changes", 6 Oct 2026. Statuses removed, 1221 problems, site history.
7. https://mathstodon.xyz/@tao/117395268889583356 and /117395269325940185 (thread also 117395267721642920, 117395268130901862) — Tao, 6 Oct 2026 18:00 UTC. "harvested", Math 1.0/2.0.
8. https://terrytao.wordpress.com/2026/10/10/math-2-0/ — Tao, "Math 2.0", 10 Oct 2026; slides teorth.github.io/tao-web/slides/math-2-0-caltech-2026.pdf.
9. https://en.wikipedia.org/wiki/OpenAI_bulk_release_of_mathematical_papers — Wikipedia, to 31 Oct 2026. Timeline skeleton; secondary.
10. https://www.scientificamerican.com/article/openai-unleashes-hundreds-more-math-results-upon-a-field-already-in-shock/ — Howlett, Scientific American, 6 Oct 2026. Sutherland, Litt, OpenAI "not bound".
11. https://www.quantamagazine.org/updates/transformation/ — Quanta updates, 6 and 9 Oct 2026. Commelin, Dafermos, AHM 800+.
12. https://www.ahmath.org/statements — Association for Human Mathematics, 7 Oct 2026.
13. https://agmai.org/ and https://agmai.org/general-sep29/ — AGMAI members and guidelines, Sep-Oct 2026.
14. https://techcrunch.com/2026/09/21/openai-forms-math-advisory-group-as-its-ai-resolves-more-than-100-open-problems/ — TechCrunch, 21 Sep 2026. "more than 100", 25 Fields Medalists.
15. https://techcrunch.com/2026/10/08/openais-math-solutions-arent-meeting-the-fields-standards-yet/ — TechCrunch, 8 Oct 2026. 719, 10 CoT, 42%, Tao and Wood.
16. https://www.science.org/content/article/openai-just-changed-math-forever-exactly-how-no-one-yet-knows — Science, Oct 2026. Kogler, 300 of 719. Not fetched (403).
17. https://fortune.com/2026/09/08/openai-says-it-cracked-navier-stokes-math-grand-challenge-buckmaster-accusation-cheating-intimidation-tao-lament/ — Kahn, Fortune, 8 Sep 2026. Navier–Stokes, Buckmaster, Tao "strip-mining".
18. https://sherwood.news/tech/openai-claimed-a-math-breakthrough-this-weekend-only-to-be-smacked-down/ — Sherwood News, 20 Oct 2025. Weil's post, Hassabis, LeCun, X links.
19. https://gigazine.net/gsc_news/en/20251020-openai-researcher-announced-gpt-5-math-breakthrough/ — Gigazine, 20 Oct 2025. Bubeck posts, Bloom's three sentences, X links.
20. https://arxiv.org/abs/2511.16072 — "Early science acceleration experiments with GPT-5", OpenAI et al., 20 Nov 2025. Sawhney/Sellke section.
21. https://www.scientificamerican.com/article/ai-uncovers-solutions-to-erdos-problems-moving-closer-to-transforming-math/ — Howlett, Scientific American, 12 Feb 2026. "about 100", "souped-up literature search".
22. https://github.com/teorth/erdosproblems/wiki/AI-contributions-to-Erd%C5%91s-problems — Tao et al. wiki, to 30 Jun 2026. Category definitions, problem lists.
23. https://papex.org/blog/how-many-erdos-problems-has-ai-solved — Papex, 21 Oct 2026. Wiki tallies (69), 586 solved on 5 Oct, erdosproblems.ai counts.
24. https://www.erdosproblems.ai/ — Plasma AI, fetched 1 Nov 2026. 494 / 110 / 617.
25. https://www.quantamagazine.org/why-the-legendary-erdos-problems-are-falling-to-ai-20260803/ — Kakaes, Quanta, 3 Aug 2026. 565/652, 111 changes, #333.
26. https://epoch.ai/latest/announcing-frontiermath-erdos — Epoch AI, 1 Sep 2026. 68 problems, 1,217/652, scores, budget, Bloom's 3-5.
27. https://arxiv.org/abs/2609.25050 — Adamczewski and Bloom, "FrontierMath Erdős", 6 Sep 2026.
28. https://arxiv.org/abs/2601.22401 — Feng et al. (Google DeepMind), 29 Jan 2026 (v3 5 Feb). 700 / 13 / 5 / 8.
29. https://the-decoder.com/terence-tao-says-gpt-5-2-pro-cracked-an-erdos-problem-but-warns-the-win-says-more-about-speed-than-difficulty/ — The Decoder, 16 Jan 2026; Tao primaries mathstodon.xyz/@tao/115855845720777387 and /115891257393270694 (14 Jan 2026, "1-2%").
30. https://mathstodon.xyz/@tao/115591487350860999 — Tao, 22 Nov 2025. "becoming routine".
31. https://techcrunch.com/2026/05/20/openai-claims-it-solved-an-80-year-old-math-problem-for-real-this-time/ — Bellan, TechCrunch, 20 May 2026. #90, Bloom "cathedral".
32. https://www.sciencenews.org/article/ai-guardrails-erdos-math-problem — Hulick, Science News, 8 Jun 2026. Bubeck 50%, Leiden 1,590, Bloom quote.
33. https://physicsworld.com/a/ai-led-solutions-of-erdos-problems-spark-debate-over-the-future-of-mathematics/ — Physics World, 29 May 2026. #1196.
34. https://epoch.ai/benchmarks/frontiermath and https://epoch.ai/frontiermath/tiers-1-4/about — Epoch AI. 338 problems, OpenAI funding/access, v2 errors 42%.
35. https://deepmind.google/discover/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/ — Google DeepMind, 21 Jul 2025. 35/42 graded; 2024 28/42.
36. https://www.maginative.com/gemini-with-deep-think-achieves-gold-at-international-math-olympiad/ — Maginative, Jul 2025. Luong on OpenAI's uncertified IMO claim (secondary).
37. https://xenaproject.wordpress.com/2026/10/01/to-grieve-or-not-to-grieve/ — Buzzard, 1 Oct 2026.
38. https://arxiv.org/abs/2608.24961 — Jin, Ke, Sui, "The Gold Rush in AI4Math", 25 Aug 2026. arXiv AI-use shares.