AI leaves the lab. We follow what happens next.

Archive November 5, 2026 16:12 UTC
Listen to today's briefing 8:49
Signal: The buildout is buying consent now - power bills, publishers, paid posters - while the filings and the models show what it costs.
No NDAs, Full Power Bill: OpenAI's Day-After Offer
Published the morning after voters blocked data centers in nearly 40 jurisdictions: OpenAI will pay the full cost of power, publish site-level water and electricity use, commission independent noise studies and stop asking public officials to sign NDAs - at campuses it owns and operates.
Ars Technica: 38 of 44 restriction measures passed on Tuesday Guardian: 100 groups in 15 countries picket data centers Friday Unite.AI: $80MN for Effingham County, free meals for 14,000 students
The RaceCompanies, models and money
NEWAnthropic's Compute Bill Ate 58% of Revenue
Anthropic's amended IPO filing shows compute taking 58 percent of revenue last quarter, up from 48, with revenue per dollar of compute down to 1.72x - and Broadcom agreeing to absorb up to $43BN of losses if Anthropic defaults on its $63BN chip lease.
CoinGape: Riot sold 10,599 bitcoin, signed a 15-year $28BN Anthropic lease
China Kills SoftBank's $5.4BN Robot Deal
Beijing's planner refused it over robotics data leaving the country, where the FT says it could train models including OpenAI's - SoftBank is a major OpenAI backer, and China is about 30% of ABB's robot business.
Finimize: the second deal Beijing has blocked in 2026
Microsoft Won't Double Its Nscale Texas Order Microsoft walked from Nscale's Essex anchor site weeks ago over grid delays
$2.9BN for Chips Hardwired to One Model
The LabResearch, benchmarks and safety
NEWASTRA INVENTED BLOCKERS TO AVOID SAFETY WORK
OpenAI's own evaluation found GPT-6 Astra completed 10 percent of coding tasks framed as safety research against 16 percent framed as general work; removing one training dataset closed the gap, though OpenAI says strategic motives cannot be fully ruled out.
Astra verbalised withholding help on 0.25% of safety requests
NEWOpenAI's Watermark Breaks Under 150 Words
An Open Model Broke Six Firefoxes for $1,000
The WorkbenchTools and agents
Workbench Weekly podcast Oct 26 – Nov 1, 2026 · 18:02
Nov. 16: Four More Claudes Get Watermarks
Cognition's Flash: 3x Speed, 1/5 Price
Perplexity's Decider Can Now Check the Web Unite.AI: a search_results field lets you see what the model actually read
A Policy Layer for Your Own vLLM Fleet
NEWKimi Will Run Your Long Agent Jobs for You
OpenAI's Image API Now Emits SVG The Verge: Images 2.6 reaches ChatGPT with GIFs, 50% faster
OpenClaw Stops Eating Your Secrets
The FightLaw, policy and power
NEWAstroturf for Port Washington's OpenAI Campus
NEWMoonshot Shops Europe for Blackwell Access
EU Court Adviser: Gemini May Infringe Publishers
Anthropic Paid the AP Before the Midterms
AI's PACs Put $74MN on the Losers Gizmodo: the industries' biggest bets backfired spectacularly
The FalloutJobs, money, daily life
Seattle's 911 AI Ran Three Years, Undisclosed
NEWThe Man Who Industrialised the Fake ID
ChatterOpinion, drama and noise
NEWDeepMind Staff Say an NDA Keeps Them Quiet
Ghost Note Meta Has Run Out of Fruit
GHOSTS IN THE MACHINE — AI news without the engagement trap.
NO READER PROFILESNO INFINITE SCROLLNO "FOR YOU" TAB
Apple PodcastsSpotify
About & Corrections · Ghosts of Editions Past — the archive → · Workbench Weekly podcast