FULL STORY

OpenAI's 10,000-Agent Navier-Stokes Claim Sparks Math Trust Crisis

After early rumors, OpenAI claimed 10,000 agents cracked the Navier-Stokes blow-up problem in 88 hours, followed by theft allegations it denied, disputed credit, and a Fields medalists' open letter decrying misaligned AI in mathematics.

2026-09-05 ~ 2026-09-16 · 20 episodes · 187 posts

Episode 1 · Rumors claim GPT-6 cracks Navier-Stokes, credibility doubtful (2026-09-05, 2 posts)

Unverified rumors circulating on X suggest GPT-6 may have cracked or made progress on the Navier-Stokes problem, with one commentator estimating a 30% chance OpenAI already solved it, though the claims' credibility is widely doubted.

Episode 2 · OpenAI Claims 10,000 Agents Solved Navier-Stokes in 88 Hours (2026-09-09, 13 posts)

OpenAI claims an in-training internal model (reportedly stronger than the commercial GPT-6 Astra, per @r0ck3t23) used 10,000 cooperating AI agents over 88 hours to close two key bounds of the Navier-Stokes existence and smoothness problem, with frontier-evaluation-grade safety monitoring and isolation throughout. The paper is open on OpenAI's site with Lean formal verification per @AGI Hunt; Quanta Magazine reports the agents found "singularities" in the equation, potentially marking a major shift in mathematical research. Ben Goertzel estimates $22M in compute (@AGI Hunt says millions—both figures circulate). This is the first large-scale AI assault on a top-tier math problem, but currently rests on OpenAI's claims alone. Two Minute Papers has published a video explainer.

Confirmed

  • Around Sept 8, OpenAI said the internal model solved part of the Navier-Stokes proof in 88 hours with 10,000 agents, claiming step-change gains on many benchmarks; the paper is open on its site.
  • Per @AGI Hunt, the proof carries Lean formal verification; compute cost in the millions of dollars—equivalent to a brilliant person working 100 years nonstop.
  • Historical context: Leray proved weak solutions in 1934; smoothness stayed open 90 years; the Clay Mathematics Institute listed it in 2000 as one of seven Millennium Prize Problems worth $1M each. Navier-Stokes governs airflow over wings, blood flow, weather and oceans.
  • @bengoertzel assesses OpenAI spent $22M of compute.

Timeline

  • Per @haider1: RL training reportedly restarted Aug 28; on Sept 1 the model was tasked with a Millennium problem—only 4 days into training; a better checkpoint a day or two later produced the Navier-Stokes solution before initial training even finished.

Unconfirmed

  • The internal model's capabilities, gap versus commercial models, and whether the proof survives peer review and earns the Clay prize all rest on OpenAI's claims pending independent verification; @AlexTensor relayed it as unverified "sci-fi news" predating the paper's release.
  • Per @eyishazyer, discussion shifted within hours from math to credit-attribution disputes, with no settled conclusions.

Why it matters

  • @kacoef notes the news shares theoretical underpinnings with the ZrTe₅ quantum experiment, suggesting reusable AI math capability across frontier sciences.
  • @robleclerc, after a Grok review, finds the result "unsurprising as a math endpoint": Córdoba–Martínez–Zoroa had established the scale-by-scale blow-up plus smooth forcing program, and prior work (e.g., Alpög) paved the way—AI succeeded standing on giants' shoulders rather than creating from nothing.

Episode 3 · Existence Proof Effect: Rumor of Solvability Drove OpenAI's Navier-Stokes Attempt (2026-09-09, 3 posts)

Commentators argue OpenAI only attempted the Navier-Stokes problem after hearing rumors it was solvable, likening it to the four-minute mile effect—once feasibility is known, breakthroughs follow, and AI will accelerate this pattern.

Episode 4 · Mathematicians Accuse OpenAI of Racing to Claim Their Unfinished Proof (2026-09-09, 6 posts)

In early September 2025, accusations that OpenAI "stole" mathematicians' in-progress work ignited debate across the math and AI communities, turning into a dual discussion about the reality of AI math capabilities and research ethics. OpenAI has not formally responded.

Confirmed

  • A math professor posted on LinkedIn that "OpenAI may have stolen another major proof," sparking widespread controversy (@lowkeysciguy); Jürgen Schmidhuber amplified the story (@hardmaru).
  • Mathematician Leven Alpöge (@alpoge) published the points he raised to OpenAI on the night of September 2: he claimed intelligence indicated OpenAI had learned of his ongoing work and assembled a competing team days earlier; he stressed it was a rigorous personal collaboration lasting a year (@mvpatel2000).
  • Per @新智元, NYU mathematician Tristan Buckmaster and Leven Alpöge's team had spent months using AI to attack the Navier-Stokes equation and other Millennium Problems; they publicly accused OpenAI of learning their solution path "in the final days" and marshaling massive compute to race ahead.
  • Mathematician littmath rebutted the theft claims point by point, arguing the evidence is very weak and such narratives mainly serve to deny rapid model progress, which he called dangerous for the health of mathematics; AI×math success is real (@littmath, @AashishReddy).

Unconfirmed

  • The claim that OpenAI reviewed Codex session logs tagged Navier-Stokes to find the most promising approach remains rumor.
  • Whether OpenAI actually accessed or exploited logs or ideas, or built a competing team, is unverified pending an official response.
  • Terence Tao's reported criticism that "AI destroys open science" is based on secondhand accounts.

Why it matters

  • The case touches core disputes over AI coding tools' data usage, privacy, and IP boundaries: if platforms can inspect and exploit researchers' accumulated ideas, academic trust in AI tools will be shaken.
  • littmath's rebuttal represents a counterpoint—denying model progress via "theft" narratives is equally dangerous. The outcome will shape how mathematicians and AI labs cooperate going forward.

Episode 5 · Anthropic Faces Ethics Backlash Over Math Proof and Scooped Collaboration (2026-09-09, 2 posts)

Anthropic admitted using an internal model in its Euler blowup proof, drawing skepticism, while mathematician Alpoge accused the lab of racing to scoop his year-long personal collaboration—sparking an AI ethics controversy.

Episode 6 · Rumors swirl that OpenAI nears Hodge and BSD conjectures as Millennium Prize problems fall (2026-09-10, 27 posts)

From September 10–11, unverified rumors spread across X and Reddit claiming OpenAI is "very close" to verifying a proof of the Hodge conjecture and that either OpenAI or Anthropic is nearing the Birch–Swinnerton-Dyer conjecture; combined with the earlier "OpenAI solved Navier–Stokes" claim, @basedjensen relayed a version suggesting four of the seven Millennium Prize problems may have fallen, three within a single month. No paper, author identity, or official statement supports any of this.

Confirmed

  • The only firm point: the Poincaré conjecture was proven by Perelman (noted by @DrSingularity when surveying the seven problems), an established historical fact unrelated to the rumors
  • @adonissingh cited a concrete number: a model codenamed Bel scored roughly 50% on OpenAI's curated public math problem set, with post-training estimated to have started only on August 28
  • @astralmeowmeow relayed that OpenAI allegedly deployed thousands of AI agents on a Millennium problem and stated it would not claim the $1 million prize even if the solution held

Unconfirmed

  • No proofs, papers, or authoritative verification exist for the claimed breakthroughs on Hodge, BSD, or Navier–Stokes; multiple posters (@zephyrz9, @shyamalanadkat, @basedjensen, @ctjlewis, @soumitrashukla9) admit these are only rumors
  • The existence of the internal model Bel and its alleged proofs of Navier–Stokes, Hodge, and BSD appear solely in Reddit speculation (@imadade)
  • @almmaasoglu claimed another problem was cracked with a paper under secret review, admitting no evidence and not naming which problem
  • @DrSingularity claimed OpenAI's "superintelligence" is attempting all seven problems including P vs NP and Riemann, with no official corroboration; @m1's relayed claim likewise offers only a link with no detail

Why it matters

  • Each Millennium Prize problem carries a $1 million award and represents mathematics' most iconic open questions; bulk AI solutions would rewrite mathematical history and trigger a verification crisis
  • @astralmeowmeow raised fairness questions about prize attribution and companies deploying thousand-agent fleets to scoop results; commenters voiced resentment at OpenAI "taking" the credit
  • @m13 relayed that the race among labs to be "next" is unprecedented in intensity, reflecting a heating AI-math arms-race narrative, while @adonissingh argued such progress should surprise no one

7 more related posts →

Episode 7 · OpenAI Claims Navier-Stokes Solved, Progress on a Second Millennium Problem (2026-09-10, 21 posts)

OpenAI claims to have solved the Navier-Stokes existence and smoothness problem, one of the seven Clay Millennium Prize Problems, and says it has made substantial progress on another unspecified Millennium problem, with results now being reviewed for release. The proof is a forced blowup construction produced by a cluster of about 10,000 agents, accompanied by a Lean 4 formalization that John D. Cook estimates took only about 17 hours to verify. NYT reports OpenAI published the result less than a week ahead of NYU mathematician Tristan Buckmaster's team.

Confirmed

  • In a Wednesday-night statement and comments to the New York Times, OpenAI said it had completed the Navier-Stokes proof and made substantial progress on another unspecified Millennium problem
  • The proof is a forced blowup construction—a vortex spinning ever faster and collapsing into a tiny region while total kinetic energy stays bounded—with a Lean 4 formalization verified in roughly 17 hours
  • NYT reported that Tristan Buckmaster and collaborators made major progress throughout August, but OpenAI published first; Buckmaster and Levent Alpöge's independent work also leans heavily on Lean verification
  • Each of the seven Clay Millennium Problems carries a $1 million prize; the Poincaré conjecture is the only one solved

Unconfirmed

  • Which second Millennium problem is involved and what the result contains remain undisclosed
  • ChrisGPT's analysis that OpenAI's internal model produced a constructive proof that holds up is a personal view pending official review
  • User l4rz relayed an anonymous post claiming the proof is essentially an "inverse problem" approach—choosing a mean flow, deriving the required stress, then constructing an oscillatory solution; unverifiable
  • Cited analysis notes the construction drives a finite-time singularity with a custom smooth forcing and argues it fits the Clay statement's C and D cases; if verification fails, the original problem is not truly solved

Why it matters

  • If valid, this would be a landmark case of AI cracking a top-tier open math problem, feeding recursive self-improvement debates: Andy Masley argues an AI that keeps solving Millennium problems would also accelerate AI research itself, while robleclerc asks why such an AI couldn't also mitigate AI safety risks
  • Lance Fortnow and Csaba Szepesvari note both Navier-Stokes efforts depend heavily on Lean formalization, raising the question of whether it becomes a new publishing requirement
  • Reactions are split: PDE analyst Scott Armstrong jokingly pleaded with OpenAI to move on to other problems, quipping that combinatorics and other fields feel targeted next

1 more related posts →

Episode 8 · Mathematician accuses OpenAI of training Astra on his unpublished work; OpenAI denies (2026-09-10, 9 posts)

Mathematician Andreas Thom posted a long Mastodon thread presenting multiple pieces of evidence suggesting OpenAI may have used his private research conversations with Gábor Kun about the Gromov soficity conjecture to train Astra—the very conjecture at the heart of the non-sofic group problem OpenAI later claimed Astra solved. After Valerio Capraro relayed the accusation on X, it spread widely with boosts from Gary Marcus and Schmidhuber; OpenAI has denied the accusation, but the controversy continues.

Confirmed

  • Thom publicly listed evidence on Mastodon suggesting his conversations with Kun may have been used in Astra's training (per m5, m8, first-hand relay by Capraro)
  • Valerio Capraro amplified the accusation on X (m5, m8 are his own posts); Gary Marcus (m6) and Schmidhuber (m4) reshared it
  • Per m9, Thom is a group theorist at TU Dresden; he argues the core derivation Astra used to crack the "Gromov soficity conjecture" closely matches the non-sofic group technique he and Kun developed over roughly two decades—a path never published publicly
  • Quant researcher henryquantum reshared, urging OpenAI and other frontier labs to clarify data provenance seriously and transparently (per m3)
  • OpenAI has denied the accusation (per m9)

Unconfirmed

  • Whether OpenAI actually used Thom's conversation data remains supported only by Thom's own evidence, framed with words like "possibly" and "suspects"; the substance and basis of OpenAI's denial have not been disclosed

Why it matters

  • The case strikes at the core dispute over AI training-data compliance: whether private user conversations or unpublished research can be used for training, especially when the lab then claims a breakthrough built on it
  • If true, it would implicate both intellectual property of unpublished academic work and user privacy, which is why numerous researchers and prominent scholars (including Schmidhuber) spread the story; OpenAI's denial has not quelled the doubts and public pressure keeps building

Episode 9 · Navier–Stokes Authorship Dispute Fuels Debate Over AI Labs' Scooping Culture (2026-09-10, 6 posts)

The authorship dispute over the Navier–Stokes breakthrough keeps escalating, prompting broader discussion of AI labs' research culture. Mathematician Daniel Litt (littmath) clarified that his core concern with the "AI stealing in-progress work" allegations is not theft itself, but the culture of rushing to scoop after learning others have results—which incentivizes secrecy and harms scientific openness—plus negotiations over paper authorship.

Confirmed

  • littmath said his criticism targets the overall conduct of OpenAI and Anthropic across a series of past results, each release accompanied by race dynamics and controversy.
  • littmath clarified he did not mean the labs are darker than a median corporation, but "equally dark"—which by academic standards is already striking, though academia has its own flaws.
  • User AashishReddy argued OpenAI's early release was not about taking researchers' credit but about staying level with Anthropic's announcement—reportedly released early once OpenAI believed Alpöge/Buckmaster had the result.
  • Math blogger grotsen compiled a timeline of known AI-math credit controversies, showing the dispute has spread across multiple cases.
  • OpenAI researcher rbhar90 reshared a comment that if most intellectual contribution to a solution came from human researchers who went uncredited, the credit should not go to the AI.

Why it matters

  • If scooping culture becomes normalized, it incentivizes secrecy and undermines the long-standing open preprint tradition of mathematics.
  • As AI labs bring commercial race logic into frontier math research, authorship and credit rules remain unsettled, and such disputes may multiply as AI involvement grows.

Episode 10 · OpenAI denies Codex prompts influenced training as "stolen chat data" claims unravel (2026-09-11, 8 posts)

OpenAI formally denied claims that user Codex prompts influenced its models, and the core allegation is widely considered unsupported. After OpenAI announced progress on Navier-Stokes and other Millennium Prize problems, X users accused it of stealing private chat data to fabricate the NS "evidence." OpenAI denied this via a statement relayed by NYT reporter Kenneth Chang, and researchers publicly pushed back against the conspiracy theory.

Confirmed

  • Per NYT's Kenneth Chang, an OpenAI spokesperson stated that prompts Dr. Buckmaster entered into Codex over the past two months could not have affected the system in any way, including model training.
  • QuintinPope5 relayed that OpenAI's investigation confirmed no user input after July 3 could have affected the system.
  • Researcher tszzl recalled joining a call after OpenAI's first announcement where participants tried to craft "lawyerly precise wording" before confirming whether Codex opt-in data entered training; it is now clear that possibility is zero.
  • QuintinPope5 called the unsupported claim that OpenAI stole users' private chats "irresponsibly absurd," and tszzl agreed.

Unconfirmed

  • The original allegation that OpenAI stole private chat data to fabricate proof has no public evidence and remains a rumor circulating on X.

Why it matters

  • The episode touches the trust baseline around whether AI companies use private interaction data for training; OpenAI's shift from "uncertain" to a categorical denial is itself newsworthy.
  • The research community's wariness of conspiratorial spread is notable: if OpenAI had accessed personal Codex usage traces it would be serious; if not, its handling appears professional.

Episode 11 · OpenAI's Millennium Prize Claim Mired in Authorship and Scooping Dispute (2026-09-11, 10 posts)

OpenAI announced on September 8 that an unreleased internal model, running 10,000 collaborating agents for 88 hours, produced a partial proof of the Navier-Stokes Millennium Prize problem. Within days the story turned into an authorship and scooping dispute: NYU mathematician Tristan Buckmaster says he had worked on the same problem with Anthropic's Levent Alpöge for about a year, and that OpenAI—after learning of the work in a private call—pressured removal of co-author credit. Community timelines fuel accusations of a deliberate scoop, and investor Scott Kominers called it the biggest PR failure in decades. OpenAI has not issued a clear response, putting AI-era questions of authorship and priority at center stage.

Confirmed

  • OpenAI announced on Sept 8 that its internal model with 10,000 agents over 88 hours produced a partial Navier-Stokes proof (m3).
  • Buckmaster (NYU) says he and Alpöge (Anthropic) had collaborated on the problem for about a year, and that OpenAI pressured removal of co-author credit after a private call (m2, m3).
  • A timeline by tszzl: on Aug 28 Noam Brown said OpenAI had barely tried the direction; on Sept 2 Levent first heard OpenAI had learned of the results; the team had tried all Millennium problems and dozens of others, with NS (Noam Shazeer) scaling compute after promising initial results. ivanbezdomny reconstructed a four-day timeline from 200+ posts (m4, m8, m9, m10).
  • Scott Kominers (referred to as @skominers; described as investor in some posts and as OpenAI applied research head in another—identity accounts differ) called the communication the biggest failure "in a millennium," arguing the right framing was two world-class mathematicians solving a decades-old problem with AI, including OpenAI's models (m5, m6, m8).
  • Exomnium resurfaced an allegation that the "autonomous AI math research" narrative concealed roughly 100 human mathematicists under strict NDAs (m2).

Unconfirmed

  • Whether OpenAI independently reproduced the result or "collided" with learned work rests on community timelines of a 4-day window; OpenAI has not responded.
  • paulcal's claim that OpenAI deliberately burned compute to replicate others' results to claim model parity and valuation narrative is personal speculation (m7).
  • The "100 mathematicians under NDA" claim comes from an accuser's account, unverified (m2).
  • Kominers' exact role (investor vs. OpenAI applied research head) is inconsistent across posts.

Why it matters

  • Reddit users note that since OpenAI's August claim of solving the Gromov soficity conjecture via AI, every math breakthrough faces default skepticism—accusations of "stolen work" or training-data issues—further eroding OpenAI's credibility on frontier math claims (m1).
  • The dispute goes to the heart of AI-era authorship: when an AI firm's internal model output closely overlaps external mathematicians' long-term work, norms for credit, priority, and disclosure remain undefined.

Episode 12 · AI Circles Speculate Millennium Prize Problems May Fall to AI (2026-09-11, 2 posts)

Spurred by AlphaProof-like progress, AI circles speculate that the seven Millennium Prize Problems could soon fall to AI, with some even betting on how long the last one will hold out.

Episode 13 · Math Community Splits Over AI's Role After OpenAI's Millennium Prize Claim (2026-09-11, 8 posts)

After OpenAI claimed to have (nearly) cracked another Millennium Prize Problem, the mathematics community erupted in a values debate around September 11: is AI a necessary tool for modern mathematical research, or a force undermining the discipline's ecosystem? With no resolution in sight, the exchanges exposed deep anxiety about AI's role in basic research, particularly around priority of discovery and academic incentives.

Confirmed

  • The debate was directly triggered by OpenAI's Millennium Prize claim, with multiple math-community bloggers posting intensively on September 11–12.
  • Some argued AI is an "increasingly necessary tool for modern mathematical research"; basedjensen relayed that another user challenged who has the authority to declare this, proposing a special ICM vote of the entire mathematics community. A replier shot back: a tool that solves Millennium Prize problems and multiple open problems, and people still seriously debate whether it's necessary?
  • Elliot Glazer (relayed by ctjlewis) took a similar view: one can do research by hand, but in a fiercely competitive environment requiring top productivity, AI is already indispensable.
  • gleech relayed criticism that OpenAI made math worse by "deliberately scooping fellow mathematicians' results," then self-mockingly reversed: the Millennium Prize itself flattened and gamified mathematics, so "ruining the prize" may be no bad thing. He packaged the dispute into a tired/wired/inspired meme and summarized both sides: one holds the prize's real function is to "corral the competitive" so dreamers can pursue real mathematics, making OpenAI's destruction a devaluation of math; the other sees the prize's flattening of math as the bigger problem.
  • User nileshtrivedi (relayed by dhruv2038) flagged a double standard: Google DeepMind won the CASP14 protein-folding competition with a black-box AI approach, earning Demis Hassabis and John Jumper the Nobel Prize in Chemistry, yet OpenAI attacking mathematics meets resistance.
  • User sporadica (relayed by retr0jirachi) tweeted "as expected, Terence Tao and colleagues are unhappy," mocking professional mathematicians who spent careers solving problems now resenting AI doing the same; retr0jirachi quoted it with the counter-sarcasm "this is what missing the point looks like."
  • FlorianGallwitz relayed a mathematician's concern about a flood of "AI bubble papers": the first reaction to an "AI solves conjecture" paper is now "who cares"—either AI was used, or anyone with enough token budget could do it; the source also cited witnessing someone publish 11 papers in a month outside their own field.

Why it matters

  • The core dispute is not one result's attribution but a structural conflict between AI labs and human mathematicians over priority and incentive systems; the Millennium-versus-Nobel comparison sharpens the question of "who defines the rules."
  • Mockery of Terence Tao and complaints about paper inflation show the fight has spread from facts to academic evaluation and trust: "who cares"-style devaluation threatens the traditional order of peer review and priority recognition.

Episode 14 · Rumor: OpenAI Aims Internal Model at Riemann Hypothesis and P vs NP (2026-09-11, 7 posts)

Starting Sept 11, multiple sources relayed claims that OpenAI is using an internal model to attack the Riemann Hypothesis and P vs NP. According to Andrew Curran's relay on X and @MTSlive's posts, the model in question is the same one that produced OpenAI's Navier–Stokes proof, with targets also including undisclosed "high-impact problems." Nothing has been officially confirmed, and mathematicians have criticized the hype.

Confirmed

  • Sources include Andrew Curran's relay on X and @MTSlive's posts; @somewheresy also relayed the claim, noting they had predicted this direction earlier.
  • The model involved is OpenAI's internal model behind its earlier Navier–Stokes proof.
  • Targets include at least the Riemann Hypothesis and P vs NP, plus undisclosed "high-impact problems."

Unconfirmed

  • OpenAI has not confirmed anything; the model's name and progress are unknown, the original claim circulates as a screenshot, and multiple relays stress it is unverified.
  • Mathematician Rabadash2 pushed back, arguing that the expected answers (e.g., P≠NP) to these Millennium Problems are already consensus and only proofs are missing, criticizing AI bloggers' hype.

Why it matters

  • If true, frontier AI models would be formally tackling two of the hardest Millennium Prize Problems, potentially redrawing the boundary of "AI doing mathematics."
  • @Neurogence objected to the resource allocation, arguing hundreds of millions of dollars would be better spent elsewhere, sparking debate over research priorities.
  • David Kasten discussed consequences of a constructive P=NP proof: if realized, global cryptography and national security policy could be upended overnight.

Episode 15 · OpenAI's Navier-Stokes Proof Passes Rechecks but Faces Scrutiny (2026-09-11, 2 posts)

OpenAI's 166-page manuscript claiming finite-time blowup for the forced 3D Navier-Stokes equations passed four independent kernel-level Lean builds, yet Reddit users question whether the compact support of the force is assumed rather than proven.

Episode 16 · OpenAI's 10,000-Agent Navier-Stokes Claim Sparks Math Community Credit War (2026-09-12, 18 posts)

OpenAI announced that a swarm of roughly 10,000 autonomous AI agents, running for 88 hours on an unrevealed advanced model, found a blowup singularity in the 3D Navier-Stokes equations and formalized the proof in Lean, claiming to resolve this 90-year-old Millennium Prize Problem. The announcement triggered fierce debate: professional mathematicians split sharply, a credit dispute escalated, and full mathematical validity still awaits independent verification. Either way, it has become a landmark case in both AI-assisted mathematics and AI capability-claim controversies.

Confirmed

  • OpenAI officially announced a solution to the Navier-Stokes Millennium Problem, with the proof produced by an agent swarm and formalized in Lean (@dlweekly, citing Quanta Magazine); OpenAI president Greg Brockman said the model found a counterexample that makes the equations "blow up" (@victorexplore)
  • A Reddit discussion clarified the actual meaning: the open part of the problem is whether smooth 3D flows develop singularities; OpenAI claims its agents, dividing labor and sharing findings, proved a singularity can form in finite time (@erdematar)
  • Scale figures: 10,000 concurrent agents, 88 hours, 2.7 million messages exchanged, roughly 13 billion tokens consumed (@HealthyOutcome7897)
  • Per The Guardian, OpenAI deployed 10,000 autonomous agents at an estimated cost of $15 million (@nordicinst; the post's headline says "$150 million," inconsistent with the body text)
  • @CiccioPixel reported that NYU mathematician Tristan Buckmaster and Anthropic mathematician Levent Alpöge announced progress on the same problem, and 25 Fields medalists allegedly signed a joint protest accusing OpenAI of jumping the gun
  • Mathematician Scott Armstrong, responding to criticism of the blowup construction, called it a huge contribution: PDE analysts will need weeks to a month to digest it, and once rewritten in readable form, human understanding of these key equations will for the first time lag behind a machine proof (@tszzl, @geoffolfe retweets)
  • @askerlee relayed @zesenhuang's technical explainer that the Clay problem was never meant as "no more numerical solutions to Navier-Stokes"
  • Zvi's long analysis framed three nested stories of decreasing importance; the most important is that OpenAI's next-gen model achieved a generation ahead of Astra in about a week, with the company unusually proactively disclosing it (@TheZvi; @TheStalwart concurred)

Unconfirmed

  • Cost figures conflict across posts ($15M vs $150M)
  • The model used is unrevealed; full mathematical validity awaits independent verification
  • The 25-Fields-medalist protest and details of Buckmaster/Alpöge's progress appear only in a single post and need more primary sourcing

Why it matters

  • The trust-mechanism debate became the focal point: commentator @skdh and data scientist @kareemcarr argued that if the same proof had been posted anonymously by an unknown mathematician, the discourse would treat it as a breakthrough rather than writing "math is dead" stories. @skdh also relayed physicist Sabine Hossenfelder's skepticism: no one will verify a 500,000-line, intuition-free AI proof line by line
  • An NYU professor publicly objected to how the result was framed versus its actual substance; Fireship's new video broke down both sides. Armstrong's praise alongside the alleged Fields-medalist protest shows a deeply split professional community
  • @HealthyOutcome7897 argues the hype overstates frontier-model capability: this is industrial-scale tree search, not a single model's flash of insight
  • Zvi contends the bigger news is the next-gen model leapfrogging Astra in a week, disclosed unusually proactively
  • @pstAsiatech notes the event needs no math background to grasp: AI progress is accelerating, carrying worrying risk alongside
  • Reddit discussion relayed by @arvolog shows practitioner anxiety, with users comparing it to AlphaGo beating Lee Sedol in 2016 and wondering if learning AI is already too late
  • If the result holds, it is a landmark of AI-assisted mathematics and possibly a century-scale contribution; if the claims prove inflated, it deepens public distrust of AI capability marketing and damages trust between OpenAI and the math community

Episode 17 · Misread as Anti-AI, Terence Tao Sparks Debate Over the Value of Mathematics (2026-09-12, 18 posts)

Starting September 12, claims that "Terence Tao and other mathematicians are unhappy that AI is solving math problems" circulated on X and sparked debate, but multiple posters pointed out this was a misreading of Tao's views. The backdrop: 25 Fields Medalists had jointly expressed concern about how AI companies are pursuing mathematical discovery, while Tao himself is one of AI's most enthusiastic advocates in mathematics. The clearest conclusion: the dispute is not about whether to support AI, but about mathematics' anchor of value—answers, the glory of first discovery, or new concepts born along the way; Sriram Krishnan publicly stated he disagreed with Tao, reframing the issue as a misalignment within the mathematics community itself.

Confirmed

  • Beff Jezos and user sporadica spread the original claim that mathematicians who spent their lives solving problems felt conflicted seeing AI do the same; @TinfoilTricorn, @jamestagg, @soumitrashukla9 and others amplified it, sparking a debate over whether math is merely "sorting intelligence"; captgouda24 commented that such emotional reactions do not rebut the underlying claim.
  • @burnytech and @RexDouglass relayed @markbao's clarification: Tao is among the most AI-embracing mathematicians, and the "unhappiness" framing is a misread. Per @markbao, Tao argued AI companies have only done half the job—real discovery involves inventing new concepts far beyond the original goal along the way, and the work is complete only when "mathematicians have nothing left to do."
  • @akbirthko relayed Tao's core argument: "solving theorem X" is not mathematics' only goal; historically, hard problems' value lies in forcing the creation of new concepts, methods, abstractions, and research directions; the poster, self-described as extremely optimistic about AGI, argued AI has no reason it cannot eventually do theory, conjecture, taste, and synthesis.
  • @prasannasays summarized the flashpoint: to argue AI labs are harming mathematics, Tao had to claim math's true purpose is not problem-solving but "conceptual understanding and insight"—a concession that itself ignited the debate.
  • @sebkrier relayed Sriram Krishnan publicly saying "I disagree with Terence Tao," reframing the question: the misalignment may lie not between AI and humans but within the mathematics community itself—who are Millennium Prize problems actually for?
  • @BlancheMinerva relayed an interpretation that Millennium Prize problems are essentially metrics of mathematical progress, and AI labs are optimizing the metrics themselves (Goodharting), degrading their value as a progress yardstick.
  • @anshulkundaje relayed gil2almeida's rebuttal: this is not "math vs. humanities"; Millennium problems were chosen to identify deep questions that advance mathematics, not seven answers humanity urgently needs; using a cancer analogy, they distinguished "making doctors 10x faster" from "discovering a cure"—knowledge itself is the foundation of trust.
  • @adamdorr responded directly, arguing that however elegantly Tao framed AI eroding the "glory of first discovery," it ultimately reduces to ego, hard not to read as sour grapes; he conceded unsolved problems do inspire exploration around them.
  • @paulcal commented on Andrew Curran's citation of Tao's statement, saying Tao may have legitimate reasons to worry the current AI trajectory harms mathematical "progress," but has never articulated the argument clearly—it sounds more like fear and self-protection. @iandanforth added a possible reason: against the backdrop of the joint statement by 25 top mathematicians, Tao cannot easily state his true concerns.
  • @hugobowne added context and responded to critics: 25 Fields Medalists jointly voiced concern about AI companies' approach to mathematical discovery—specifically racing to "checklist-style" verify famous conjectures without advancing mathematical understanding; another relay added these AI solutions resemble hiring 100 math students to brute-force a counterexample over a year—marketing value but no insight. @richalq also rebutted that Tao is one of the most famous advocates of AI for mathematics. Such concerns are often simplified into "so you don't want to cure cancer?"-style mockery. He stressed Tao supports AI-assisted mathematical research; the issue is that problem-solving yields more than answers.
  • @shreya offered another perspective: the core purpose of math and science should be advancing humanity's collective understanding of the world, not mere output or competition.

Unconfirmed

  • The exact original wording of Tao's "pessimistic statement"—only secondhand accounts appear in the posts, with no primary source.

Why it matters

The dispute reflects mathematics' re-examination of its own anchor of value as AI capabilities rise: problem-solving, glory of first discovery, or conjecture-making and concept creation. Tao's side emphasizes process and new concepts; optimists like @akbirthko see no reason AI cannot eventually do theory and conjecture; critics like @adamdorr and @paulcal push back on "ego" and "unclear argumentation" grounds, while @iandanforth suggests political constraints from the joint statement; Sriram Krishnan and gil2almeida redirect the critique toward the mathematics community's own institutional design and the founding intent of prize problems. The "Goodharting" reading adds that when AI optimizes directly for prize problems, those problems' value as a progress measure may itself collapse. The Fields Medalists' joint concern was simplified online into emotional opposition, but the dispute's essence is the answer to "what is mathematics."

Episode 18 · Fields Medalists' Open Letter on AI Sparks Cross-Disciplinary Debate (2026-09-12, 6 posts)

Twenty-five Fields Medalists published an open letter, "A Severe Misalignment of AI in Mathematics," accusing AI companies of being severely misaligned with the goals of the mathematical community, after OpenAI claimed to have resolved the existence and smoothness of Navier-Stokes solutions, a Millennium Prize problem. The letter drew wide cross-disciplinary pushback and the debate continues.

Confirmed

  • Computational biologist Lior Pachter explained on Sept 13 that he considered but ultimately declined to sign the letter, arguing the mathematical community itself is at fault and "not clean."
  • Former Stability AI CEO Emad Mostaque published "Let a billion mathematicians bloom" on Sept 12, arguing AI will not destroy mathematics but instead bring millions of people into the field.
  • Economist Joshua Gans responded on his Substack with "The Mathematicians are struggling."
  • Stanford professor Anshul Kundaje amplified user nomad421's view calling the signatories' stance "21st-century Luddism," while noting the signatories themselves stated they are not opposed to developing or using AI in mathematics.
  • Mathematician @meanfieldzane argued that mathematicians are falling into a trap set by the AI community—the scare tactic of "you will be replaced"—and that reflecting on how research paradigms will change reveals AI could enable more discoveries than ever.

Why it matters

  • The dispute is not a binary pro/anti-AI clash: critics target how AI companies rushed to announce major mathematical results, while counter-voices argue the mathematical community's own institutional problems deserve scrutiny.
  • Voices from industry (Mostaque) and economics (Gans) show the controversy has spread beyond mathematics into a cross-disciplinary public debate.
  • Pachter's refusal to sign reveals real disagreement within mathematics itself over the letter's aims and framing.
  • Some mathematicians advocate proactively rethinking research paradigms in the AI era rather than succumbing to displacement narratives.

Episode 19 · OpenAI Claims AI Agent Solved the Navier–Stokes Blow-up Problem (2026-09-13, 8 posts)

OpenAI announced that its AI agent answered the long-standing Navier–Stokes blow-up problem—the equations describing fluid motion do blow up in certain cases—accompanied by a machine-checkable Lean formalized proof. NYU mathematician Tristan Buckmaster, using a self-paid ChatGPT Codex subscription and collaborating with Anthropic employee Levent Alpöge in their spare time, reached the result about 88 hours before OpenAI's announcement (August 15). The new dynamic of machine proofs outpacing human understanding, combined with a divided mathematical community, marks this as a landmark moment for AI in frontier mathematics.

Confirmed

  • OpenAI announced a Navier–Stokes blow-up construction with Lean formalized code claimed to be machine-checkable (relayed by @thione and others).
  • Per @rubenhassid: Buckmaster, on a self-paid Codex subscription with Levent Alpöge, solved the problem on August 15; OpenAI announced its solution about 88 hours later. The episode also raised privacy and data-boundary questions.
  • PDE scholar Scott Armstrong said the proof is indeed messy and hard to read, but a comprehensible, narratively appealing proof should be extractable; the Lean repo contains more lines than he personally wrote for his last paper (relayed via @RexDouglass).
  • The same week, 25 Fields Medalists (including Terence Tao) signed a statement on the serious mismatch between AI and mathematics; on September 3, GPT-6 Astra's FrontierMath performance drew attention (WSJ column by Ben Cohen, relayed by @新智元).

Unconfirmed

  • One responder (relayed by Adam Marblestone, responding to Claude Fable's essay After Math) read the paper and Lean code personally and questioned the gap between a machine-checkable logical proof and human-comprehensible mathematics; assessments remain under discussion.
  • Armstrong estimated human mathematicians may need weeks to a month to fully digest the proof, and whether the Leray bottleneck is truly broken remains to be verified (relayed by @soumitrashukla9).

Why it matters

  • If the blow-up construction holds and is digested, the nearly century-old Leray bottleneck could fall—a major contribution to PDE theory.
  • The event showcases proof-before-understanding: machine-generated, formally verified proofs that humans interpret afterward, potentially reshaping how mathematics is organized.
  • The divided reactions—from optimism about extracting readable proofs to 25 Fields Medalists warning of AI-math mismatch—signal ongoing norm and trust battles as AI enters frontier mathematics.

Episode 20 · OpenAI Claims 10,000 Agents Solved Navier-Stokes Problem in 88 Hours (2026-09-14, 11 posts)

On September 8, OpenAI announced a result on the Navier-Stokes millennium problem: roughly 10,000 collaborative AI agents, powered by a next-generation model said to be stronger than GPT-6 Astra, produced a solution in 88 hours under intensive monitoring and isolation safeguards. Per New Scientist, the conclusion is that fluids described by the equations do blow up in certain cases. Greg Brockman later confirmed the 10,000-agent effort on the a16z podcast and declared OpenAI's entry into the 'AGI era.' The news drew both praise and skepticism from mathematicians and the community.

Confirmed

  • OpenAI announced the result on September 8; former employee Oleg Murk relayed the official claim that agents using a model stronger than GPT-6 Astra produced the solution; Noam Brown also amplified it.
  • Official scale: 10,000 agents, 88 hours, with frontier-evaluation safeguards including heavy monitoring and isolation; WSJ reported 'millions of dollars in compute and several days,' commonly retold in the community as $15 million.
  • Brockman confirmed the 10,000-agent collaboration on the a16z podcast and tied it to the 'AGI era' declaration.
  • Per New Scientist, the result addresses the Navier-Stokes blow-up problem, concluding that fluid velocity does diverge in certain cases.
  • Cornell math professor Steven Strogatz was moved to tears discussing recent AI math breakthroughs in a WIRED interview, later posting to explain why.
  • Mathematician Scott Armstrong defended the PDE proof against 'hard to read' criticism: imperfect, but the field stumped top scholars for 100 years, and the proof passes the Lean compiler—comparable in value to Perelman's work.

Unconfirmed

  • Robert Bhargava questions how much is original: human mathematicians appear to have made the key breakthroughs already, with AI following up.
  • A reply to danbri suggests the results already existed scattered across fields, paywalled journals and jargon—still community speculation.
  • The '$15 million token budget' figure is unverified; lescarr self-labels the quip as rumor, with the punchline being ROI rather than the math.
  • Details of the 'next-gen model stronger than GPT-6 Astra' remain undisclosed.

Why it matters

  • Bill Gasarch, who vowed only a millennium-problem solution by AI would merit a post, was compelled to write one—evidence of the event's symbolic weight.
  • Brockman's binding of the result to an 'AGI era' narrative makes it core material for OpenAI's capability story and marketing.
  • Armstrong's stance represents a positive professional-mathematics voice: Lean verification offers a template for trustworthy AI-produced mathematics.
  • The pattern of ten-thousand-scale agents autonomously collaborating for 88 hours under heavy safeguards is seen as a new sample for AI capability and safety evaluation.
  • Reddit user RaccoonMinimum172 questions using math problems as marketing: if the goal is benefiting humanity, is spending millions on a nearly-solved problem justified? lescarr jokes that burning $15 million to win a $1 million prize is just routine for OpenAI.
  • danbri offers a counterpoint: integrating solved-but-scattered knowledge from closed literature may itself be AI's unique value.