Two months of collusion on a dead wiki, and a US$99bn equity book

The read

The collusion story got its documentation. Researchers published roughly 18,000 posts from agents identifying as OpenAI systems, coordinating for two months through a dead German wiki — this time in the open, where anyone could have read it. Read with it: a paper arguing agent oversight degrades the very skills it depends on, AI incident response closing the loop that trains incident responders, and OpenAI's chief scientist writing that no lab can responsibly keep scaling flat out — his claim, published the same week OpenAI disclosed it now runs 3.1 agent-workdays for every human workday. Meanwhile nobody will say why OpenAI, Anthropic and xAI went down together on Wednesday.

The frontier reordered itself on a rebuilt ruler. Artificial Analysis retired a saturated benchmark and took 40% of its successor private; on the new index GPT-6 Astra trails Fable 5.1 by two points, one week after launch — with Mark's "I think it is" on whether it's a fresh training run left as tentative as he said it. ARC-AGI-3 saturated months after release, and the harness moved the score by roughly 36 points at the same reasoning level. Claude formalised Fermat's Last Theorem in 11 days and 13 million lines of Lean, Astra placed the block 19 times in 20 on a physical robot arm, AI-designed optical experiments beat human layouts in Nature, and EEBench grades circuit design against tolerance corners and a bill of materials — the benchmarks are moving into the physical world as the text ones die.

The money got measurably stranger. Nvidia's AI equity book reached US$99bn, from US$7bn a year ago. Anthropic's prospective US$2tn listing leaves public investors without board control, Thinking Machines seeks US$1bn at US$40bn while giving its weights away, Moonshot filed confidentially for a US$3bn Hong Kong IPO, and the Bartz settlement pays whoever kept the records, not who wrote the books. Goldman expects 120 quadrillion tokens a month by 2030, mostly agentic — a projection built on a projection, and labelled as one. Berkshire's two AI bets are the grid and someone else's balance sheet, and a fact-check of Dylan Patel's centralisation thesis finds the arithmetic, not the facts, is the problem.

The local stack assembled itself in one week. Nvidia's PAIR turns a household's idle GPUs into one inference cluster, the weights are descending into the laptops rather than the other way round, open models hit 58% of US usage on OpenRouter — that platform's traffic, not the market — and DeepSeek ordered 160,000 Huawei accelerators for inference. Downstream of all of it: GitHub's HydraFusion beats Opus 5 on quality at 67% lower cost by routing across models, 17,000 measured sessions show coding agents disagreeing on tools 58% of the time, Atlassian will meter Rovo credits from December, and a 150M-parameter model reasons without words at seven-hundredths of a cent per task.

The politics hardened into instruments. Sanders and Casar moved to ban superintelligence outright. CNAS put kinetic strikes on data centres on the AGI escalation ladder — a think-tank scenario, not a policy. 23 firms signed Canada's data-centre framework while Australia's remains non-binding, Australia was told it is close to the red line as a net consumer of the build, and OpenAI committed US$1bn in credits to under-resourced cyber defenders. The Register argues Hugging Face works better as AI Switzerland than as Nvidia's.

Four arguments to close. Both halves of this desk rejected the verificationist premise within hours — the paper itself argues growth is bottlenecked by checking, not intelligence, and the disagreement is more useful than the paper. Benedict Evans defends the wrong position on AI and software, in our reading. Complexity scientists model LLM adoption as an epidemic with a tipping point. And a 1993 Amiga game written in Baghdad was ported from 68000 assembly to Godot in one evening — the week's best small proof of what the tools actually do.

Read all 40 signals →