The file
Sourced AI briefs in publish order — newest first. Citeable originals live on the homepage under From Signal Desk.
After injunctions in Brazil and Norway, a UK CMA investigation and bipartisan Senate letters, Meta shelved a rewrite requiring Business account holders to allow AI crawling of linked sites, even behind paywalls, with no opt-out short of leaving. The pause leaves the consent question open.
A year after open-sourcing Aerial, Nvidia says adoption of AI Aerial and Sionna grew 10x, with 8,000+ institutions. King’s College, SoftBank, Leeds and CTTC results show what AI-native 6G could do; independent reproduction and live-traffic tests are still ahead.
Ramp’s corporate card data shows token spend up just 1.5% across August and September after a 42% July jump, with a partial October rebound. Price cuts, discounts and model routing are capping revenue just as Anthropic’s reported valuation falls toward $1.2–$1.3 trillion.
After the New York Times reported US intelligence believes Huang was targeted through a Megaspeed executive, the chairs of the House China, foreign affairs and intelligence committees asked him to appear on November 18. Nvidia rejects the claims and won’t say if he will attend.
The first of Ai2’s FoundationWorlds pairs a simulated arm and lettered blocks with 85,336 auto-generated demonstrations. Early results: VLA models still generalize poorly, frontier agents clear about 63% of hard search-and-spell tasks, and world-model data lifts one policy from 19% to 78%.
Startup Underdog released Woof, a roughly 31.7MB, 31-million-parameter text-to-speech model with downloadable weights, inside a retro game that jabs at cloud AI. It won’t match big cloud voices, but it may be good enough for local assistants and private audio.
Dutch group Geef Tegengas scaled a crane at a site planned for three 85-meter towers, hanging banners against AI and Microsoft’s Israel ties. The project proceeds through loopholes in Amsterdam’s ban, and the protest joins a widening European and US backlash.
MIT Technology Review’s Grace Huckins booked a real cleaning. The cute, four-and-a-half-foot robot is mostly remotely piloted, staggered on stairs, quit early in heat and knocked over olive oil, while collecting training data on every wipe. Tau has done only a couple hundred cleanings.
OpenAI and Hacktron will jointly pick high-priority test areas focused on how OpenAI systems interact with third-party services, and Hacktron gets expanded frontier-model access via Daybreak Red. With Cathedral joining a sprint and bounties rising, remediation speed is the real scoreboard.
OpenAI will no longer use NDAs with officials on data center projects beyond narrow security and trade-secret protections, and says it will pay full electricity and infrastructure costs, publish hiring reports and attach local compacts to every owned campus. The commitments are progress, not a pledge to build less.
Built with biotech Ligo, the benchmark scores agents on 100 tasks from real experiments, including failed runs, mislabeled plates and data too weak for a conclusion. Astra beats GPT-5.6 Sol’s 43.8% but still makes ordinary scientific errors, and better lab skill is also a biosecurity signal.
Perplexity’s new Teammates are persistent agents that take on several roles, keep memory and report into Slack. They join OpenAI’s Dots and Anthropic’s Claude Duet, at a much higher price, and with a CEO vision of hundreds of agents per employee.