AI leaves the lab. We follow what happens next.

Archive November 3, 2026 11:12 UTC
Listen to today's briefing 8:38
Signal: Polls open, and an EU adviser draws the copyright line, while Nvidia and Anthropic set the real rules in verbal orders and risk factors.
Anthropic: We May Imminently Cross Our Own Red Line
Monday's amended S-1 says internal estimates indicate Anthropic may imminently meet its Responsible Scaling Policy threshold for 'dramatically accelerated progress attributable to AI' - and that it has in the past failed to notice such an acceleration until after it occurred.
No current model has crossed it, Anthropic says - and it may rewrite the policy Gizmodo: the sentence is new - the 26 October version of the filing made no such claim TechCrunch's caveat: an S-1 is disclosure and marketing at once
The RaceCompanies, models and money
NEWNVIDIA'S UNWRITTEN RULE: SELL 51% OUTSIDE CHINA
Five people tell the FT that Nvidia's sales force passed the instruction verbally - new cloud contracts must send at least 51% of capacity to non-Chinese buyers - and Nvidia would neither confirm nor deny it.
The rule covers new contracts only, leaving existing deals untouched Nvidia's only answer: America cannot win the AI race unless its infrastructure wins worldwide
Anthropic Commits $895BN. Quarterly Revenue: $17BN. The Register: a trillion in compute commitments, owed partly to its own competitors
NEWDrax Jumps 8.8% on a GBP 20BN Data Centre Radiant's chief pitches frontier compute at home for British researchers and startups
NEWPalantir's Commercial Side Passes Government CNBC: Palantir slumped 8% as contract-value growth slowed sharply
Apple Blames RAM and Siri for Thinner Margins Price rises across Mac, iPad and iPhone will offset only some of it
The LabResearch, benchmarks and safety
GPT-6 Edited Qwen's Weights. Math Score Rose 6x.
MIT CSAIL authors say Astra read Qwen3.5-0.8B's raw weights, traced its reasoning loops to two layers and subtracted 131,072 of them - 0.016% - lifting MATH-500 from 7.9% to 48.1% with no training at all.
One engineer: a competent human researcher would have done exactly this It failed on Gemma 3 270M, a model that rarely loops
The WorkbenchTools and agents
Workbench Weekly podcast Oct 26 – Nov 1, 2026 · 18:02
Stolen Keys Are Running $100K Days Google's own writeup names the access brokers and the downstream 'LLM Mixers'
NEWProof That the Model You Paid For Answered Rerunning the model to check is impractical when the weights are private
Claude Agents Stop Resending the Chat Threads live in memory, expire after five minutes, and are barred from zero-retention setups
Perplexity Open-Weights Its Embedders
Musk: Grok Bot Cost More Than It Earned OpenAI merges ChatGPT and Codex allowances next month; Meta says Muse is not profitable
NEWDeepSeek Writes Its Own Plugins
Nvidia Rotates Brev's SSH Keys
The FightLaw, policy and power
NEWEU Adviser: Training Is Fine. Answering Isn't.
Advocate General Szpunar told Europe's top court that training a model on press articles is freely permitted under the text-and-data-mining exception, but retrieving a publisher's page to generate an answer may infringe that publisher's right.
The case is Hungarian site operator Like Company against Google, over Gemini reusing its press content Copyright protects the specific form of expression, not the mere facts reported
UK: AI Ransomware Will Hit the Grid by 2028 Horne will warn of 'an unprecedented rise' in incidents in the coming months
Today, 43.5 Gigawatts Go to the Voters Wired: New York's Mamdani calls data centres 'our risk, their reward'
NYT: Brave Is an AI 'Fence' for Stolen Pages Three more news organisations filed near-identical scraping suits, MLex reports
The FalloutJobs, money, daily life
NEWCPS Lawyer Cited a Case That Never Existed Only the second known CPS case, after an extradition filing cited nonexistent authorities in February
NEWSUSE Cuts 10% and Names AI as the Reason 'Modern tooling is already changing our daily workflows', the chief executive wrote
Meta Said the Coach Videos Were Fine A safe-sport advocate calls the clips abuse material that normalises grooming
ChatterOpinion, drama and noise
YC's Tan Spends $200K a Year on Tokens He has published his 23 Claude Code skills on GitHub as gstack
NEW2,000 Takedowns in a Day Killed the AO3 Clone AO3 itself is down for up to four weeks after a security incident, with 18.5MN works offline
Ghost Note My 'I Voted' Sticker Is AI Slop
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