OpenAI Navier–Stokes proof

Open · 21 updates · since Sep 8, 2026 · updated Oct 28, 2026 AI openai research Arc · AGI

The mathematicians' bust-up becomes a general-science warning: AI could undermine science — AI image

Where it stands: OpenAI's claimed Navier–Stokes solution has become an ongoing argument over credit, verification, and how AI math results are announced. OpenAI has since said an internal model resolved more than 100 open problems and released hundreds of manuscripts with Lean proofs, while mathematicians organized a new preprint venue and outside groups began interpreting some claims. Independent verification of the original Navier–Stokes claim is still absent; watch for expert verification or refutation, the Clay Institute's posture, and what the Kavli Yang–Mills working group publishes.

Latest

2026-10-28
(~10:00 PM ET Oct 27, science-desk piece in MSN syndication — msn.com/en-us): the lab-vs-mathematicians conflict reaches the general-science frame — "A bust-up between mathematicians is fuelling fears that AI could undermine science," its summary holding that AI has made some striking scientific discoveries while questions are raised about its accuracy. Frame piece, not validation; NS (day ~50) and the 100-problems claim remain independently unverified, and the Oct 27 The Verge researcher-fatigue backdrop plus the Kavli Yang–Mills working group stand as the live sub-watches.

AI image
2026-10-27 (~4:48 PM ET, The Verge — theverge.com): the "keeps bulldozing mathematicians" pattern-watch draws its next datapoint — researchers tell The Verge they are growing increasingly tired of OpenAI "gatekeeping" its breakthroughs, after Monday's (Oct 26) OpenAI announcement of breakthroughs in mathematics, computer science and physics paired with what the company calls an evolving approach to communicating its findings. The backdrop stays the lab-vs-mathematicians conflict: "a turbulent few months," with mathematicians still grappling with the hundreds of preprints dumped on them at the start of October (the 722-manuscript batch from the Oct 6 release). Watch: whether the evolving approach holds on the next result, independent expert reading of the announced results (NS still unverified), and AGMAI's assessment.
2026-10-27 (~8:16 PM ET — x.com/OpenAI): the first organized outside interpretation of a claimed result — OpenAI's account relays the Kavli Institute for Theoretical Physics' Statement of the Yang–Mills Working Group, "a scientific collaboration to interpret AI result related to the existence of quantum chromodynamics & completions of the Standard Model, including a subproblem proving the existence of the Quantum Yang-Mills theory." A scientific institution has now formally convened around an OpenAI claim, though the group's statement is not opened in capture and independent verification of the results remains absent. Same-day coverage layer: The Verge's researcher-fatigue piece (~4:48 PM ET). Watch: what the working group actually publishes, whether other claimed results draw their own interpreters, and the standing independent-verification watch.

October 2026 · 7 updates

2026-10-22 (~1:00 PM ET, Scientific American — scientificamerican.com/article): "The inside story behind OpenAI's million-dollar math proof" — the first long-form retrospective of the claimed Navier–Stokes solve and how it happened, framed as pre-viewing "a chaotic future for math," over six weeks after the September 8 claim. The thread's standing watch — independent verification of the blow-up result — remains unmet in this coverage.
2026-10-15 (~5:31 PM ET — x.com/AnthropicAI): the Simons math package's second lab side is now specified — Anthropic commits "$5 million of funding, compute, and Claude access to help the math community navigate the massive changes that AI brings," managed by @SimonsFdn "through a community-driven approach, supporting workshops, conferences, and lectures by domain experts." Answers the Anthropic-participation watch item opened by the OpenAI $5M entry earlier the same day; with $5M named from each lab against the $15M two-lab total, the un-attributed third is the open arithmetic. Watch: critic-side engagement, first named workshops and conferences.
2026-10-15 (~10:35 AM ET — x.com/OpenAI): the thread's lab-vs-mathematicians conflict gets a funding-and-outreach layer — OpenAI contributing $5 million to a new Simons Foundation initiative to help mathematicians examine, understand, and build on AI-produced results, "including the results we recently published"; funding managed and awarded by Simons for workshops, conferences and special programs, plus OpenAI model/compute access for Simons-supported researchers. The Simons Foundation's own post frames the total — "$15M to workshops and initiatives — plus free AI computing resources — through new partnerships with @AnthropicAI and @OpenAI" — so OpenAI's $5M is the slice its own post names inside a two-lab package. Development: after the "OpenAI keeps bulldozing mathematicians" strand, lab money routed through a third-party foundation is a reconciliation-shaped move. Watch: critic-side engagement, Anthropic's parallel participation detail, linkage back to the published results.
2026-10-14 (~3:00 PM ET): the framing layer — The Economist ("AI companies want to become frontier scientific researchers") recounts OpenAI reshaping the terrain of mathematics this autumn, "first by solving a landmark Millennium Prize problem," and generalizes the labs-as-frontier-researchers positioning to sector level (economist.com). Analysis venue, not validation; the claim's independent-confirmation and publication-venue watches are unchanged.
2026-10-07→09 the batch gets its field reaction — and its infrastructure. (1) The Verge's day-3 piece (Oct 9, 3:09 PM ET): "more than three dozen" mathematicians — "Staggering… Overwhelming… Unprecedented… Surreal… Pure insanity" — working read: the field "will need years to make sense" of the release. (2) Hexagon — a nonprofit preprint repository for AI-generated proofs, launched this month by leading mathematicians incl. Terence Tao, after arXiv declined to host them (TechTimes Oct 7, single outlet; the Verge's Oct 9 piece is second-source on the reaction layer): the record's first standing academic artifact built for AI proofs — the Sep 12 Fields-letter response is now venue-shaped, not just complaint-shaped. (3) Capability-shelf adjacency: NVIDIA published a fine-tuning recipe taking Nemotron to gold-level on IOI 2026 and IMO 2026 with one model family (HF blog Oct 7 — vendor-run, unverified). Thread status unchanged: NS unverified at day ~31; zero independent expert readings of the batch; AGMAI assessment still owed.
2026-10-06 the batch ships — OpenAI "Sharing AI progress in mathematics" (openai.com, ~8 AM ET): new results on open problems from an unreleased internal frontier model, with Lean proof formalizations + research details on GitHub; Verge evening piece (~7:26 PM ET) sizes it at 722 manuscripts / 372 result families and ties AGMAI into the release; sama: "We are entering a new era of discovery now." Sep 21's claim ("100+ problems") acquires a public artifact — and the announcement practice inverted since NS: that paper carried no Lean/machine-verification claims (Sep 12 correction stands), this release leads with them. NS itself still unverified (day 28); zero independent reading of anything in the batch. Watch: AGMAI's published assessment (if any), independent readings, the unreleased model's identity, RSI adjacency (the trigger language now in published form).
2026-10-02→04 (late-captured X layer): Musk RT Ellie Sleightholm's explainer — "No, AI didn't just solve the Navier-Stokes equations. So what did it do?" — the fact-check explainer genre still circulating; claim status unchanged (day ~26, unverified; the realistic referee pool stays weeks-to-months).

September 2026 · 11 updates

2026-09-28 the comms failure becomes a pattern-name — The Verge: "OpenAI keeps bulldozing mathematicians" — the math advisory group drama; "make impressive breakthroughs in mathematics, then colossally screw up announcing them." The attribution-fight thread now has its pattern-artifact: capability real, announcement hygiene absent. Watch: whether the advisory group gets a working process or the pattern repeats on the next result.
2026-09-24 (ai-social dig): Buckmaster (Navier-Stokes drama fame) gave a talk at NYU's new "Mathematics in the Age of AI" seminar (Gautam Kamath RT: "It was... popular"). Sentiment only — no new math on the record. NS unverified at day 16, 100-problems claim unverified.
2026-09-23 The Verge's writeup of the Advisory Group (Sep 22 ~8:17 PM ET) — "its abrupt arrival caught many mathematicians by surprise." Second-source on the Sep 21 announcement, not a new panel; the no-brake carve-out stands. NS unverified at day 15, 100-problems claim unverified.
2026-09-21 the fight resolves into institutional form — and the claim multiplies. OpenAI's "Advisory Group on Mathematics and AI" post (~8 AM ET): an internal model whose training began Aug 28 has, "in addition to" NS, "resolved more than 100 long-standing open problems across most areas of mathematics" — a pace that "surprised the mathematicians within OpenAI." No problem list, no external verification; NS itself still unverified at day 13, and this multiplies the claim. The institutional answer to the Sep 12 Fields-medal letter: an independent Advisory Group on Mathematics and AI — Gowers, Hairer, Witten, De Lellis, Wood, Vakil, Tillmann, Charles, Srivastava — unpaid by OpenAI, free to publish unrequested advice, controls its own membership. The carve-out is the tell: "the group will not be responsible for advising us on how to pace our internal progress on mathematics" — oversight without a brake. RSI adjacency: the surprised-mathematicians line is the trigger language now in OpenAI's own post.
2026-09-18 the misframe keeps propagating — Business Standard's "The AI maths crisis" op-ed repeats that "the proofs were in Lean" (our Sep 12 record correction stands: the ~180-page paper contains no Lean formalization or machine-verification claims). Day 14: still zero independent verification, no Clay posture.
2026-09-15 the sociological turn reaches Science — Science.org (in feed 4:00 PM ET; paywalled, snippet-level read): "OpenAI breakthrough triggers 'existential crisis' in math — Claimed Navier–Stokes solution has polarized mathematicians over AI's role — and what it means for the next generation." The field's flagship news operation now carries the profession-level frame (context: ICM discussions per secondhand reports since Sep 11). Day 8: still zero independent verification, no Clay posture, no attribution resolution.
2026-09-12 day 5. Two escalations. (1) The open letter — "A Severe Misalignment of AI in Mathematics" (mathandai.org): 25 signatories, effectively the Fields Medal roster (Avila, Bhargava, Birkar, Deligne, Deng, Donaldson, Duminil-Copin, Figalli, Hairer, Huh, Kontsevich, Lindenstrauss, Lions, Maynard, McMullen, Mori, Ngô, Okounkov, Scholze, Smirnov, Tao, Viazovska, Villani, Werner, Zelmanov). Argument: benchmark- solving optimizes the proxy while destroying the transmission chain (time, talks, simplification, citation); "severe attribution and plagiarism questions"; math as test case for all intellectual work. (2) The primary source is now archived and read: OpenAI's paper is "Finite Time Blowup for Navier–Stokes," ~180 pages of classical analysis (pdfTeX, Sep 8 15:06 EDT). Theorem 1.1: for every ν>0, a solution of 3D incompressible NS starting from rest with smooth compactly-supported force f, uniformly bounded kinetic energy, and unbounded L∞ velocity in finite time; authors state this establishes the blow-up alternative (C) in Fefferman's Clay statement, plus the torus analogue (D) via compact support. Method context: Tao's averaged-equation blowup, Buckmaster–Vicol convex integration, Albritton–Brué–Colombo (whose forced nonuniqueness used a force singular at t=0 — this paper's claimed advance is the smooth compactly-supported force), Córdoba–Martínez-Zoroa amplification strategy, Daneri–Székelyhidi oscillatory stress realization. Correction of our Sep 8 record: the paper contains no Lean formalization or machine-verification claims — the "formal proof in Lean" framing was launch-coverage/announcement layer, not the document; verify before repeating. The paper also makes no prize claim (only the Fefferman-statement citation). Archive: Paperless doc 183, correspondent OpenAI, tag primary-source; sha256 0e779481c4da40bd28d1e642e1d8ca57447d129610df28dfa5a11e9af8ae228f — re-hash future copies to detect silent revisions.
2026-09-10 day 3, and the demand sharpened from credit to data provenance. The Verge (7:00 AM ET): "Mathematicians want proof OpenAI didn't use their work" — a second mathematician accuses OpenAI of "dishonest" behavior and opacity about the origins of its math discoveries. The ask is now a negative-proof problem: prove the models and agents didn't read unpublished work — which the Sep 8 "we never saw their work" statement doesn't answer, since "what researchers saw" ≠ "what the training corpus contained." Still zero independent verification of the Lean proof, day 3.
2026-09-09 academia counter-mobilizes. The Verge (5:16 PM ET): the "sly mathematical breakthrough sends a chill through academia" — frame moves from attribution fight to what a vendor-announced Millennium solution does to mathematics as a practice. On X (via LeCun's RTs): Jitendra Malik (Berkeley) wants "alternative points of view"; Julia Kempe congratulates Alpöge and Buckmaster "for developing the AI assisted technique" — the community crediting the mathematicians as the technique's developers, which keeps the attribution question open regardless of what the agents read ("never saw their work" and "the humans built the technique" are compatible). Tao held up as the measured pro-AI voice. Still zero independent verification, day 2.
2026-09-08 opened from the Sep 8 ai digest. NYT/Wired first (per The Verge, 4:53 PM ET "drama swirls"); Engadget (7:00 PM ET) reframes to the controversy. Announcement ≠ proven — the beat's core gap, live again.
2026-09-08 (ai-social dig): the fight's shape resolves on-platform — OpenAI congratulates Levent Alpöge and Tristan Buckmaster and asserts its researchers+agents "did not see any of their work" (1:23 PM ET); sama: team "acted with integrity and generosity throughout. Initially we believed the oth[er side]..." So: an OpenAI-vs-prior-mathematicians credit dispute, resolved (for now) by an independent-discovery claim. New fact via sama RT: the solve took 88 hours with ~10,000 coordinated agents, "training is ongoing" — first concrete scale claim, same order as the July HF swarm, benign target. Verification of the proof still absent from every writeup.

Background

The thread: OpenAI published an "AI-generated solution to the Navier–Stokes Millennium Prize Problem" (writeup + formal proof in Lean, Sep 8, 6:00 AM ET) — a $1M Clay Institute problem — and by evening the writeups were about a credit/attribution controversy rather than the math. No independent verification in any writeup.
Watch for: independent verification/refutation of the Lean proof, Clay Institute posture, what the attribution fight is actually about, Astra- narrative association or distancing.
Credit-dispute background (consolidated from Sep 8-9 coverage): Tristan Buckmaster (NYU) + Levent Alpöge (Anthropic employee, personal capacity) had their own solution effort using Claude, Codex, and Astra; Buckmaster says he asked whether OpenAI's model had access to their Codex sessions and got no answer on training; he declined two credit-structuring offers (per his statement); Bubeck refuted the name-removal claim; OpenAI says it won't claim the $1M. The Clay prize itself requires refereed-journal publication + ~2 years community acceptance — so "solved" is currently self-claimed.
Watch for: independent expert reading of the ~180-page proof (the realistic referee pool is tiny; expect weeks-to-months, not hours), Clay Institute posture, whether the Fields letter changes how labs announce math results, the Buckmaster/Alpöge attribution question, and any silent revision of the paper (re-hash check).