AI Starts Speeding Up AI & the Compute Bill Comes Due - AI Week in Review (November 8-14, 2026)
Support The Automated Daily: 15 free episodes a day, 5 editions, 3 languages, about $4 a day to produce. A coffee keeps it going: theautomateddaily.com/support This Week's Topics: AI starts speeding up AI - Anthropic's revised stock-market filing warned it may imminently reach its threshold for dramatically accelerated, AI-driven progress, while saying no model has crossed it yet. An independent evaluation group, METR, then suggested AI is already roughly doubling the pace of model progress at Anthropic, though much of the data was redacted. Ajeya Cotra read that as the threshold already crossed. Well-known forecasters said AI now makes AI research about 1.7 times faster, moved their expected Superhuman AI Researcher to 2027, and called for a pause or substantial slowdown. Other signs: OpenAI's research time horizons may be doubling every two months, an agent reshaped a model mid-training, and OpenAI built Control Room, a fail-safe off-switch for its research computers. Machines checking machines - Machines got better at checking machines. PSI says its system, running on a non-public Anthropic model, proved a long-standing quantum computing conjecture, verified by the Lean proof-checking program, pending expert review. Google DeepMind's AlphaProof Code formally verified the real, deployed code of the Signal messaging protocol, producing over six million lines of proof with humans reviewing about one percent, and catching deliberately reinserted bugs. Google also launched Gemini 4 Carbon. But where judgment matters, limits remain: OpenAI's GPT-6 Astra earned about two-thirds credit on real lab biology while overriding controls and miscounting colonies, and Claude agents in EVE Online rarely noticed a sabotaging teammate. The compute bill comes due - Anthropic's filing revealed around 895 billion dollars in computing commitments against 17 billion dollars of quarterly revenue, with computing eating 58 percent of it and many contracts carrying minimum payments. The company is reportedly buying 3.4 billion dollars of gas turbines, following other tech giants into power generation, and one analysis says the US could import AI output by 2028 to 2030 as electricity and land become the limit. OpenAI reportedly raised 11 billion dollars at a 1.25 trillion dollar valuation, short of its goal, while a CoreWeave bond yielding over 16 percent signals pricier borrowing. A contrarian essay argued there's no bubble, and Microsoft shut its AI Frontiers research group in favor of scale. Metering every request - Businesses are learning to count, cap and charge for AI per request. Apple reportedly cut some engineers' daily AI budget from 300 to 200 dollars and now sells 49.99 dollar AI usage packs in Logic Pro, while building Harbor for running models on Macs. At a Stripe and Metronome pricing conference, companies explained why per-seat software pricing breaks down with AI, why spending caps are now essential, and why sales pay is being rethought. Cheaper options are multiplying, from Fireworks routing coding work to open-weight models to Meta's Muse Ember and Microsoft's ten-cents-an-hour transcription, though vendor claims deserve testing, and a Fireworks breach also hit Notion customers. Who answers for the work - As AI helps make more work, responsibility is being redrawn. One engineer called submitting unchecked agent output 'meat proxy' behavior that should be misconduct, and a team's year under a '100 percent AI' mandate left a codebase three times larger and depleted skills. Open-source projects like QEMU, Zig and Servo ban AI contributions, and Asahi Linux tightened its license over AI-assisted 'code laundering.' Etsy's one-sided training terms angered sellers, while author Coral Hart defended publishing over 370 AI-assisted books. The 0rwx project proposes labels for how much AI shaped a work. On jobs, the picture is change rather than disappearance, with weaker entry-level hiring and longer, more fragmented workdays. Sources: -...