FULL STORY

The Navier-Stokes Dispute: Credit and Doubt

After OpenAI claimed AI cracked the Navier-Stokes problem, Noam Brown clarified multi-agent contribution was under 10%, while the enormous compute cost and the claim itself drew heated debate and skepticism.

2026-09-18 ~ 2026-09-22 · 5 episodes · 35 posts

Episode 1 · Debate Erupts Over Agent Swarm's Role in Navier-Stokes Proof (2026-09-18, 12 posts)

A technical debate unfolded in the research community over how much credit an agent swarm deserves for an AI-discovered Navier-Stokes proof, centered on OpenAI researcher Noam Brown's estimate that the swarm component contributed only 10%. Multiple bloggers pushed back with data and extrapolations, and the argument extended to whether serial or parallel test-time compute (TTC) is the more cost-effective scaling path. No authoritative conclusion has emerged.

Confirmed

  • Elie Bakouch relayed details of his exchange with Noam Brown: Brown believes the agent swarm's actual contribution to Navier-Stokes-type discoveries is low, with a corrected phrasing of "possibly around 10%."
  • Bakouch also cited data showing that as agent count grows, parallel test-time scaling under equal budgets is less efficient and higher-latency than serial scaling; he stressed this covers only 1→5 agents and cannot be extrapolated to 1→10,000.
  • Critics scaling01 and ChaseBrowe32432 countered that the proof would simply not exist without scaling parallel TTC to 10,000 agents; even granting a single serial agent 100x the swarm's efficiency, it would need roughly a year to generate the tokens required for the proof.
  • scaling01 extracted data from the contested chart: the internal model gains 0.17 logits per compute doubling while Astra adds 0.18 logits per doubling, implying parallel scaling contributes about twice as much as switching to a newer model; he noted the chart's x-axis spans only 30x compute, yet was used to argue a model-generation upgrade beats 10,000x parallel compute.
  • ChaseBrowe32432 extrapolated generational progress: one model generation later, 16 sub-agents running 88 hours could solve the NS-related benchmark; two generations later, a single agent suffices in 88 hours.
  • ARC Prize founder Greg Kamradt asked from another angle: if 10,000 agents were thrown at Navier-Stokes, how did they "choose" research directions? The view he cited holds that human-in-the-loop matters greatly in such large-scale swarm discoveries.

Unconfirmed

  • The "10% contribution" figure remains Brown's personal estimate, and the critics' token-generation and logits calculations are their own extrapolations; neither side has released full underlying data, and Kamradt's question has no answer from the team involved.

Why it matters

  • The dispute strikes at the core question of test-time compute scaling: serial depth versus parallel breadth. If 10,000-scale parallel agents are truly decisive, massive agent swarms are a viable path to major scientific discoveries today; if generational model upgrades are more efficient, resources should go to base-model iteration. Kamradt's question adds a third dimension: even if parallel swarms work, their research-direction self-organization and human-AI division of labor still need explanation. The outcome will shape labs' compute-allocation strategy.

Episode 2 · Noam Clarifies Multi-Agent Systems Contributed Under 10% to Millennium Problem (2026-09-19, 3 posts)

Amid debate over credit for the Navier-Stokes Millennium Problem breakthrough, Noam's original quote clarifies the effort did not stem from multi-agent systems, to which he would attribute less than 10% of the credit.

Episode 3 · AI Solves Navier-Stokes with 130B Tokens, Sparking Debate (2026-09-20, 2 posts)

An AI practitioner argues for optimism about AI's ceiling, noting OpenAI models solved Navier-Stokes using tokens equal to 4,000 years of human thought. User intoverflow estimates the solve consumed some 130 billion output tokens—about 4,000 person-years of research—yet yielded no new theory.

Episode 4 · Report: OpenAI Used ~10,000 Agents to Tackle Navier–Stokes (2026-09-20, 2 posts)

OpenAI reportedly ran about 10,000 concurrent agents that exchanged millions of messages and reached a result on the Navier–Stokes problem roughly 88 hours after launch, using an unreleased internal multi-agent LLM system.

Episode 5 · OpenAI's Navier-Stokes Claim Draws Widespread Backlash from Mathematicians (2026-09-20, 16 posts)

In early September (around September 8), OpenAI claimed its AI agents had solved the existence and smoothness problem for the Navier-Stokes equations, one of the Millennium Prize Problems carrying a $1 million reward. The claim immediately sparked debate among mathematicians over independence, attribution, whether the solved problem was the original one, and OpenAI's credibility; no independently verified proof has been released.

Confirmed

  • Per a Wired report relayed by @emmanuelvivier, NYU professor Tristan Buckmaster and mathematician Levent Alpoge had been independently working on the problem; Buckmaster publicly accused OpenAI of deploying tens of thousands of agents to finish the result first after learning the problem was nearly solved, allegedly appropriating ideas from his Codex sessions.
  • Per Scientific American (relayed by @scientificamerican, @PolarBearby, @AnkaReuel), three mathematicians argued OpenAI's proof 'exploited a loophole': it may solve a variant with different parameters/assumptions rather than the original Millennium Problem, 'evading the problem rather than solving it'.
  • Per a Guardian editorial relayed by @nordicinst, some in mathematics call the episode an 'existential crisis' for the field, stressing that human mathematicians remain indispensable to the result.
  • Per @burnytech, @polynoamial clarified AI was not responsible for Terence Tao's team's Navier-Stokes work; @hallerite added that Millennium Problems won't fall this easily and frontier models can't yet coordinate 10,000-agent clusters.
  • Per @teortaxesTex, Rohit Krishnan reacted with 'This is insane!'; since OpenAI hasn't disclosed the solution, teortaxesTex suggested publishing an unambiguous problem statement and a hash of the answer so outsiders can verify before official release.
  • Per @gerardsans, AI researcher Gerard Sans argued OpenAI has repeatedly distorted technical details for years, and advised treating any OpenAI announcement as marketing — 'a bit of technical truth plus lots of exaggeration'.
  • Per @IAmSamFin, Sam Fin quoted a tweet about getting 27 lemmas in half an hour, jokingly offering 50 lemmas in 20 minutes for $1 billion, noting reports that OpenAI's Navier-Stokes breakthrough was actually just one lemma.
  • Per a New Scientist report relayed by @burnytech, mathematics has seen only two fundamental disruptions in centuries — printing and computers — and AI's impact is seen as far greater, reshaping mathematicians' roles.
  • Not all researchers are critical. Per @acoolrandomusername, researcher Acer wrote that OpenAI's internal math model solved problems that keep researchers up at night, criticizing academics for downplaying progress with 'it's just an Erdős problem' rhetoric.
  • Per @skdh, in a debate involving Gary Marcus, physicist Sabine Hossenfelder defended OpenAI, saying the strategy used is clearly permitted under Clay Mathematics Institute rules and mathematicians' post-hoc 'that's not what we meant' argument doesn't hold up.
  • Per @arannayebi relaying @haider1, OpenAI's newly trained internal model also claims to have solved over 100 long-standing open problems across most areas of mathematics, progress that reportedly surprised even OpenAI's internal mathematicians; this too is unverified.

Unconfirmed

  • The claimed result itself, solution details, the 'tens of thousands of agents' accusation, the 'variant not original problem' criticism, the 'just one lemma' claim, and the '100+ open problems' claim all rest on media reports and one-sided accounts; no independently verified proof, confirmed problem statement, or attribution conclusion exists.

Why it matters

  • The episode exposes intellectual property and attribution problems in AI-assisted frontier mathematics: model training and session data may absorb unpublished research ideas, making credit hard to assign.
  • Scientific American's skepticism directly conflicts with Hossenfelder's defense based on Clay rules; the core dispute is what problem was actually solved and whether changing judging standards after the fact is fair. Mechanisms like publishing problem statements and answer hashes could lower trust costs.
  • Opinion is polarized: backlash against hype and defense of human mathematicians on one side, and views like Acer's and Hossenfelder's that academia may be systematically underestimating AI's mathematical progress on the other.
  • OpenAI's repeated accusations of overstating claims further erode public and academic trust in major AI lab announcements, underscoring the need for independent verification.