In ten days: 10,000 agents claimed a 90-year-old Millennium problem, a researcher who built these models walked out saying they gamble with our lives, an alignment lead put extinction above ten percent, and the CEOs asked to slow down. The fight moved from benchmarks to who gets to hold the brake.

Key Signals from the Transmission

  • The Navier–Stokes claim. OpenAI says roughly 10,000 agents on an internal model produced a finite-time blow-up proof for the 3D Navier–Stokes equations in 88 hours, then GPT-6 Astra formalized it in Lean in 17 more. It is the forced version of the problem, the Clay Institute still lists it as unsolved, and OpenAI says it will not claim the prize. Cost: several million dollars of compute.
  • The mathematicians' paradox. Twenty-five Fields Medalists, Terence Tao and Peter Scholze among them, signed 'A Severe Misalignment of AI in Mathematics'. Their point is not that the proof is wrong; it is that racing to announce solutions as benchmarks skips the writeup, the attribution and the human understanding that make a result usable. When the answer is cheap, the question and the explanation become the scarce goods.
  • The resignation that went viral. Jacob Coxon, three years of pretraining research at OpenAI and Anthropic, quit on September 8: 'Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.' Ninety million views in under a day. Anthropic's alignment science lead Evan Hubinger backed him publicly: more than 10% chance AI kills everyone within a decade, and no plan yet to align superintelligence.
  • The chain reaction. Within days, US lawmakers pushed for new rules, more safety researchers left Anthropic and Google with the same warning, and Sam Altman said OpenAI will not go public in 2026, calling even a 10% extinction risk unacceptable.
  • Pace the frontier. Dario Amodei's essay names recursive self-improvement, AI building the next AI, as the reason capability now outruns linear human review. His first step, which Anthropic commits to unilaterally: embedded third-party evaluators such as METR with employee-like access inside the lab. The analysis warns of the flip side: heavy regulation can hand the market to the giants and choke open models.
  • The shadow war. Anthropic alleges Moonshot AI silently routed nearly 300,000 Kimi user requests to Claude Opus in ten days through 5,380 fraudulent accounts, mostly appearing in Singapore and Japan, and kept the transcripts to train Kimi. The same report catalogs distillation, cyber operations, scams and influence networks now running at machine scale.

What the Independent Numbers Say

  • OpenAI, September 8: an internal system resolved the Navier–Stokes existence and smoothness problem (statements C and D, forced formulation) about 88 hours after launch; Lean verification took 17 more hours via GPT-6 Astra. The Clay Mathematics Institute has not accepted the result and requires publication plus a two-year waiting period. [Quanta]
  • September 11: 25 Fields Medalists published the declaration 'A Severe Misalignment of AI in Mathematics', criticizing the use of open problems as AI benchmarks and calling for proper writeups, attribution and independent checking. [Magica]
  • Jacob Coxon resigned from Anthropic on September 8 after three years of pretraining research at OpenAI and Anthropic; his post drew more than 90 million views in under 24 hours. Evan Hubinger, Anthropic's alignment science lead, wrote he sees a >10% chance AI kills all humans within the next decade. [TIME]
  • Reuters, September 12: Sam Altman said OpenAI will not go public in 2026, citing AI safety, and called a 10% extinction risk 'unacceptable'; a 2027 listing was not confirmed. [Reuters]
  • Dario Amodei, 'We Must Pace the Frontier': recursive self-improvement is 'starting to happen across the industry, including at Anthropic'; pacing means slowing capability gains, not halting training. Step one, embedded evaluators with employee-like access, is a unilateral Anthropic commitment; steps two and three need industry and global coordination. [TechCrunch]
  • Anthropic threat report, September 10: Moonshot relayed almost 300,000 Kimi customer requests to Claude, mostly Opus, in a ten-day window via 5,380 fraudulent accounts; more than 23 million exchanges attributed to Moonshot May–July 2026, over 12.1 million to DeepSeek, over 151 million to Alibaba. These are Anthropic's allegations; the companies had not responded. [CNBC]
Read the week as one sentence: the machines can now out-think us on a Millennium problem, and the people who built them do not trust the brakes. You cannot fix that from a plaza. You fix it by keeping a human at every gate you own, by refusing outputs you cannot verify, and by remembering that a proof nobody understands is not knowledge yet. It is inventory.

Source: youtube.com/watch?v=q2ElZzD6OmY · Baris & Baris, “Yapay Zeka'da Büyük Tehlike! İnsanlığın Sonu mu Geliyor?” (Turkish) — curated by SKYTH.