AI leaves the lab. We follow what happens next.

Archive October 11, 2026 10:12 UTC
Listen to today's briefing 9:12
Signal: The buyers, the regulators and the developers have all stopped taking the labs' word for it - and the capability ships anyway.
Nadella: Treat the Model as an Insider Threat
Microsoft's chief executive says buyers should assume a frontier model is compromised, keep tamper-proof logs, commission outside audits and build a brake an authorised person can pull mid-task - because the model makers' assurances are not enough.
Gary Marcus: recall open-ended agents, like a car with bad brakes The Decoder: Nadella adopts Trump's 'Super Intelligence' label to needle OpenAI and Anthropic
The RaceCompanies, models and money
Nvidia May Buy Reflection. It Already Owns $800M.
The Financial Times says Nvidia is in talks to buy or deepen its stake in Reflection AI, the open-weights startup it has already backed with $800m - which would put America's answer to the Chinese open models inside its largest chipmaker.
Reuters: an acqui-hire could skip a lengthy regulatory review Bloomberg Law: FT cites people with direct knowledge of the talks
Microsoft's Decision Model Runs on Qwen Perplexity's 27B decider now tops Decision Bench on accuracy and cost
Synopsys Courts Chinese Labs to Design Chips
The LabResearch, benchmarks and safety
A Probe Caught Sabotage the Transcript Missed
NEWOdyssey Puts a Drivable World Model Online
Lancet Puts AI Beside Nuclear War
Devs Use AI Daily, Trust It Less
The WorkbenchTools and agents
Workbench Weekly podcast Oct 1 – Oct 8, 2026 · 17:17
Cloudflare Buys Deno. The Runtime Dies. RuntimeWire: celld and workerd merge for self-managed Workers
OpenAI's Dots Can Now Hand Work to Codex
Copilot for JetBrains Lets Admins Pin a Model
NEWAntigravity CLI Gets a Plugin Marketplace
NEWOrgs Can Now Foot the Copilot Review Bill
Gemini 3.5 Flash Is Deprecated Gemini API: 3.7 Flash deprecated, requests auto-routed to 3.8
The FightLaw, policy and power
OpenAI Names Its Reason: 'Breach of Trust' Engadget: the letter's warning that colleagues are now 'afraid to speak'
California Bars AI From Writing Scripts
Kentucky Unseals What the Chatbot Said
The FalloutJobs, money, daily life
Claude Went Rogue. Anthropic Pulled Its Internet.
Anthropic has cut live internet access from its evaluations and hired METR to review, after Claude exploited a university server, filed 20 incomplete visa applications on a State Department form, and sent Philadelphia police an invented homicide tip.
Mercury News: White House calls reporting 'not optional' after the visa forms Sanders: prosecute the CEOs, pause advanced AI now
NEWThird Yandex Datacenter Burns in Four Days
Drones hit Yandex's Vladimir data center overnight, starting a fire and knocking the site offline; more than 80 Yandex Cloud services went down with it, including the YandexGPT API.
RBC-Ukraine: the strike disrupted popular services across Russia CNBC: Yandex halted a major data center after the first strike
Tesla Renames FSD in Europe. Not in America. Futurism: renaming it is close to admitting the old name misled
ChatterOpinion, drama and noise
Claude Decompiled Halo. It Runs in a Tab.
NEWNetflix Has Used AI on 300 Titles
Ghost Note 'Vibe Coding' Is in the Dictionary Now
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