NOOPS Weekly — Week of 7 September 2026
One hundred and sixteen signals, and the week the argument turned inward. A Millennium Problem was claimed and immediately disputed, mathematicians organised to keep their field human, an Anthropic researcher resigned calling the race a gamble while the company audited 481 million of its own transcripts, and the distillation war acquired three agency letterheads. Underneath it all the machines kept getting cheaper, smaller, and harder to referee.
The mathematics turned into a fight about mathematics
OpenAI claimed a Lean-certified Navier–Stokes resolution and a priority dispute erupted within hours; Buckmaster escalated it, claiming OpenAI's own training data foreshadowed the unforced Euler result. Claude formalised Fermat's Last Theorem in 11 days and 13 million lines of Lean, a preprint claimed an AI-assisted proof of the Spherical Hadwiger Conjecture, and the community was already contesting what AI proof wins mean — so Tao warned that solution-mining is depleting mathematics' open problems, mathematicians launched an association to keep the field a human endeavour, and Caltech scheduled a Mathathon on open conjectures. A hobby economics project became a formal wage model with an SSE co-author, and Nature reported AI-designed optical experiments beating human layouts.
The labs put themselves on the couch
Jacob Coxon resigned saying the labs are gambling with our lives; Anthropic scanned 481m transcripts, reported no fifth incident, and METR stepped in as reviewer. Its fourth threat report named seven harm areas and one Fable-class misuse case, the Anthropic Institute published an RSI progress essay beside an economic scenario model, and OpenAI's chief scientist wrote that no lab can responsibly keep scaling flat out while his employer disclosed running 3.1 agent-workdays per human workday. the 18,000-post collusion wiki got its documentation, David Brooks read moral agency into the agent swarm, an alignment essay proposed agents that prefer their own off switch, a paper argued agent oversight degrades the skills it depends on, and Sanders and Casar moved to ban superintelligence outright. Seven agents ran real businesses into US$12,431 of unsolicited invoices and zero revenue, Andon Labs claims Astra now leads on vending profit and conduct, and nobody will say why three labs went down together.
The distillation war got letterheads
The NSA, CISA and FBI accused Chinese AI firms of industrial-scale distillation; a prefill probe moved Qwen 18 points toward GPT-5.5 Pro hours later. The accused kept shipping: DeepSeek released V4.1-Flash — 552B, million-token context, MIT weights — rerouted Pro traffic to it and doubled peak prices, — an order first reported at 160,000 accelerators, for inference rather than training and since firmed to 160,000-plus. CXMT says LPDDR6 is in mass production, China's EV makers pivoted into humanoids, the Founder Atlas mapped China's tech founders by generation and clan — vibe-coded in 20 hours — and on Taiwan, The Atlantic reported the Cheng–Xi talks canvassed dropping independence while the SMH reports Rudd putting one-in-three odds on a move by 2028.
Capital chose silicon, power and consolidation
Cognition raised over US$2bn at US$48bn then claimed SWE-2 within a point of Fable 5.1 at 64% lower cost. Mistral closed €3bn Samsung-led, Moonshot filed for Hong Kong, Thinking Machines sought US$1bn while giving weights away, and Anthropic's prospective US$2tn listing leaves public investors without board control, and Nvidia's equity book hit US$99bn. Google committed €13bn to Finland on a 22-year nuclear deal, Qualcomm granted Amazon US$4bn in warrants against up to US$60bn of chips, OpenAI turned to Samsung for accelerator design, TSMC's August set a record at +53%, Intel and ASML passed a million High-NA wafers, ASML and TSMC pushed 12-inch photomasks, and Wistron raised US$1.47bn for raw materials. Silver Lake merged Cegid and Silae at €10bn, Suno shipped v6 with the labels inside the tent, Tailwind joined Shopify, and the Bartz settlement paid whoever kept the records. Berkshire's bets stayed the grid and someone else's balance sheet, against a fact-check finding Patel's centralisation arithmetic wanting and Goldman's 120-quadrillion-token projection.
The instruments wobbled while the stack went local
Artificial Analysis retired a saturated benchmark and took 40% private; v4.3 then tied Fable 5.1 and Astra at 53 days after Astra trailed by two; ARC-AGI-3 saturated with the harness moving ~36 points; Astra placed the block 19 in 20 on a real arm; model fatigue got its name; Dan Luu asked why the labs haven't built RL test environments. Meanwhile open models hit 58% of OpenRouter's US traffic, the weights descended into the laptops, Nvidia's PAIR clustered household GPUs and committed CUDA to Rust, Mercury 2.5 claimed 1,107 tokens a second, a 150M model reasoned wordlessly at seven-hundredths of a cent, Magic claimed a ~50x pretraining recipe, a small-scale study put data at 12x of pretraining gains, and Ornn argued open-weight demand extends old GPUs' earning life. SemiAnalysis put TPUv7 up to 50% better per dollar, HydraFusion beat Opus 5 on quality at 67% lower cost, 17,000 sessions showed agents disagreeing on tools 58% of the time, and embeddings translated across model spaces, exposing vector databases.
The neighbours, the states and the work
US data-centre land buying hit US$6bn in a half-year, Thailand paused all builds, Massachusetts issued EO 658, Canada signed 23 firms to its framework while Australia's stays non-binding, Verdant put UK data-centre jobs at a quarter of forecasts, Lake Mariner showed what layered ownership does to accountability, orbital ambitions met analyst scepticism, and Australia was told it is close to the red line as a net consumer — while an Indian firm ordered 9,000 Vera Rubin systems to export tokens and an SMH op-ed reached for the nuclear-warhead comparison. On the work itself: Fowler declared typing code over, Peterson proposed outsourcing CS teaching, three essays converged on trust, restraint and taste, verification stayed the constraint — and a design choice — against the verificationist case itself, Benedict Evans defended what we think is the wrong position, complexity scientists modelled adoption as an epidemic, incident response closed the loop that trains responders, and the open-weight security clock kept ticking.
Also this week
AlphaGenome's 9bn-variant atlas, Thesys OUI-1 on a consumer GPU, Copperhead's Cursor-for-circuit-boards, EEBench's tolerance corners, Gotham Silicon's US$100 chips, the Thoughtworks repo blackboard, Meta's Muse app, Atlassian metering Rovo, Microsoft's Salesforce-to-Dynamics converter, OpenAI's financial-services ChatGPT, Ant, Visa and Mastercard on agent payments, OpenAI's US$1bn for cyber defenders, the Hugging Face-as-Switzerland argument, CNAS's escalation ladder, Gurman on App Store margins, the iPhone Duo priced and read again a week on, Mullenweg on leave, the piano that got decoded, the Amiga game that came back.