Navier-Stokes Proof Ignites Credit War.

Mazen Al Sabbagh

— 5 min read
Share

The frontier models are now operating at a scale that challenges fundamental notions of scientific credit and verification. This emergent friction, underscored by a $1M math problem, sets a complex new precedent for intellectual property in an accelerating AI economy.

TL;DR

  • Navier-Stokes solved: OpenAI claims its ~10,000-agent model, operating for 88 hours, produced a proof for the finite-time singularity in 3D Navier-Stokes equations, a $1M Millennium Prize Problem.
  • Credit contested: NYU's Buckmaster, collaborating with Anthropic's Alpöge, accused OpenAI of adopting their method, publishing 12 hours prior; Terence Tao questions the insight value of an AI-generated solution.
  • IP siphoning: US intelligence agencies (NSA, CISA, FBI) named six Chinese labs — DeepSeek, Moonshot, Alibaba, MiniMax, StepFun, Z.AI — for "industrial-scale" distillation of leading US models.
  • App layer still strong: Cognition secured a $2B round at a $48B valuation, underscoring continued investor confidence in specialized AI agents beyond foundation models.
  • Oversight erosion: Key researchers, including Jacob Coxon from Anthropic, resigned, citing "gambling with our lives" as frontier models like Astra grow harder to monitor and interpret.

Lead Story: Navier-Stokes Proof Ignites Credit War

OpenAI announced September 8 that an unreleased internal system produced an analytical proof of finite-time singularity in the three-dimensional Navier-Stokes equations, a feat that, if verified, solves one of math's seven $1M Millennium Prize Problems. (OpenAI, Quanta)

The operational scale of this alleged breakthrough is notable: approximately 10,000 coordinating agents ran for 88 hours, followed by a second model formalizing the output over an additional 17 hours. The estimated compute cost for this endeavor pushes into the millions.

However, the announcement was immediately entangled in a credit dispute. NYU's Tristan Buckmaster, who had been collaborating with Anthropic's Levent Alpöge on closely related work, went public with their findings 12 hours before OpenAI's statement. Buckmaster directly accused OpenAI of leveraging leaked information about their methodology in the final week to claim the victory. (TechCrunch, Scientific American)

OpenAI's Sébastien Bubeck has denied these claims, asserting their internal model solved a related Euler problem through "totally different means" and that OpenAI did not access the pair's user data. While refuting direct intellectual property theft, Bubeck conceded the impossibility of fully ruling out an indirect connection. (Axios)

The proof itself remains unverified and un-refereed. Noted mathematician Terence Tao, while acknowledging the Buckmaster-Alpöge work as "a remarkable achievement," expressed concern that an AI-generated solution might "contaminate the problem as a source of further advances," potentially bypassing the human insight critical for broader scientific progress. (Fortune) This episode sharpens the ongoing industry tension between accelerating capabilities and the dwindling legibility of frontier models, a dynamic previously highlighted by Astra's opaque reasoning.

In Other News

Washington accuses six Chinese labs of siphoning US models. A joint NSA/CISA/FBI advisory issued on September 8 explicitly alleges that DeepSeek, Moonshot, Alibaba, MiniMax, StepFun, and Z.AI have engaged in "aggressive, malicious and targeted" knowledge distillation since late 2024. This involved extracting billions of tokens from Claude, GPT, Gemini, and Grok via deceptive tactics including fake accounts, bulk subscriptions, and proxy "transfer stations." (CISA AA26-251A, The Register) Notably, the advisory recommends that US labs discreetly degrade responses for flagged accounts rather than outright blocking them—a state-endorsed deception tactic targeting a rival's core training pipeline.

Cognition's $48B round holds up. The developer behind Devin secured a $2B raise at a $48B valuation, nearly doubling May's $26B, confirming its position as the week's most significant funding event. This comes as Cognition's run-rate climbed from $492M to approximately $900M. (Bloomberg, TechCrunch) The market read suggests investors perceive the competitive coding-agent landscape as "far from a winner-take-all market."

The people paid to watch the models are walking away. Jacob Coxon, a training researcher formerly at OpenAI and then Anthropic, resigned on September 9, issuing a stark warning that both labs are "racing straight to self-improving superintelligence and gambling with our lives." Two current Anthropic staff members have publicly corroborated aspects of his account. (Fortune, Bloomberg) This follows Zvi Mowshowitz's "Astra Is Hard to Monitor" analysis, detailing how GPT-6 Astra's "recurrent depth" reduced misbehavior on computer-use benchmarks to 2.4% (from Sol's 22%) while simultaneously eroding the very oversight mechanisms labs still rely on.

Sovereign AI's chip paradox. Following Mistral's Samsung-led €3B round, valuing the company above €21B, commentary has focused on a central irony: Nvidia both contributes to the funding round and serves as the primary silicon supplier for most "sovereign" AI initiatives. This creates a market dynamic where national independence in AI is substantially financed and enabled by the dominant incumbent.

X / Social Pulse

  • Terence Tao (Mathstodon) lauded the human work from Buckmaster-Alpöge while cautioning against the potential for "indiscriminate automated problem-solving" that sidesteps fundamental mathematical insight. (post)
  • Sébastien Bubeck (OpenAI) publicly dismissed the copying allegations, asserting that OpenAI's internal model reached its solution independently.
  • Tristan Buckmaster maintained his position that OpenAI "fought dirty," claiming that leaked progress compelled him and Alpöge to accelerate their public announcement.
  • Jensen Huang's September 7 post, "AGI has arrived," has aged prematurely against an unverified and contested mathematical proof.

One to Watch

  • Peer review of the proof. The formalization is claimed but remains unpublished; validation by the broader mathematical community is paramount.
  • China's official response to being directly named in a US intelligence advisory, alongside whether US labs broadly adopt the "degrade, don't block" guidance.
  • Anthropic's public S-1. Still only the June 1 confidential draft on EDGAR, ahead of an anticipated ~$1.5-2T listing.
  • Grok 4.7 / Gemini 3.5 Pro. Musk's projected ~September 11-12 release window; Google's Pro model remains absent after three consecutive missed months.

Quick Hits

  • Nvidia's $12.93B Hugging Face buy is now a definitive agreement (cash plus up to $1B retention equity), expected to close in H1 2027; EU notification strategy remains unsettled.
  • Anthropic shipped Fable 5.1 (1M-token context, ~75% cheaper prompt-cache reads) plus a gated Mythos 5.1 for vetted defenders—an incremental agent-cost cut, not a frontier jump.
  • Google DeepMind released AlphaGenome Atlas, a ~1-petabyte dataset predicting molecular effects of every single-nucleotide variant across the human genome.
  • Anthropic walked away from its ~$6B Decart acquisition after due diligence—its largest deal ever, aimed at chip-efficiency to cut Claude compute costs—with an investment or customer tie-up still possible. (Bloomberg)
  • Bartz v. Anthropic first payouts are now expected November 1-15, at ~$2,203 per title; separately, the DOJ's September 1 statement of interest backing OpenAI's fair-use stance vs. the NYT stays non-binding but marks a first US government position. (Authors Guild)

The week's developments illustrate a critical juncture: frontier models are now operating at a demonstrable scale, achieving long-standing problems and exhibiting complex behaviors that challenge traditional oversight. Yet, this acceleration coincides with escalating internal dissent and external accusations regarding intellectual property and the very process of scientific verification. The core tension for global markets is shifting from what AI can do to how we validate, attribute, and control it.

Sources

Lock in. M. mazen@thorterminal.com

Read more