AI news for Tuesday, July 21, 2026
The Daily AI Espresso — the links the most-followed people in AI actually shared, curated every morning. Edited by Alexis · Live updates →
July 21 · today the experts are reading a final judgment, a safety confession, and Beijing's newest open weights.
Judge Araceli Martinez-Olguin granted final approval in San Francisco on Monday: about 500,000 books, roughly $3,000 per work, and authors can now start getting paid. The deal stems from last year's ruling that training on the books was fair use but downloading them from pirate libraries was not. It settles one case, not the question: the suits against Google, Meta, Midjourney and OpenAI are untouched by it. techcrunch.com
The search engine that bragged about sending you away as fast as possible has become a destination: AI Overviews and AI Mode answer on the page, and people stay on Google. The number the media crowd keeps quoting is Cloudflare's, cited in the piece: human traffic to the most heavily crawled publisher categories is down as much as 40% in under a year. nytimes.com
OpenAI says a model trained for long-running tasks produced novel failures during limited internal use, failures its pre-deployment evaluations never caught, so it pulled access. It went back with trajectory-level monitoring, new evals and more user visibility before restoring limited use. The stated lesson: no fixed test suite anticipates every behavior, so testing has to come with monitoring and the ability to roll back. openai.com
Nathan Lambert's teardown of Moonshot AI's announcement: 2.8 trillion parameters, a mixture-of-experts that activates 16 of 896 experts, roughly 2.5 times the scaling efficiency of Kimi K2, and #2 and #3 spots on the big capability indexes. The weights go public July 27. His uncomfortable conclusion is that China's frontier labs are getting this close to the American ones with far, far fewer resources. interconnects.ai
The Commission's AI Office published findings from more than 100 experts last week, and they're now making the rounds: frontier development is concentrated almost entirely outside the EU, compute and growth-stage capital lag badly, and the fix list starts with infrastructure and energy within two years. Yoshua Bengio is among those passing it on. ec.europa.eu
The Information reports a server chip called Frozen v2 that embeds parts of Gemini's architecture directly in silicon, cutting how much data the chip has to move. It could run 6 to 10 times more power-efficient than Google's latest TPUs and might deploy as soon as 2028, as a complement to the TPU line rather than a replacement. Alphabet rose about 3% on the report. cnbc.com
Two things that flew under the radar today.
From earlier this month, and suddenly everywhere among the experts we track: Bruce Schneier and legal scholar Jon Penney argue AI takes surveillance from expensive and targeted to cheap and total, and that the self-censorship it breeds chills the countercultures progress depends on. Their example: same-sex marriage and marijuana reform both began as behavior society watched and punished. theguardian.com
MIT FutureTech and the University of Queensland had the experts rank the risks by severity. Even assuming pragmatic mitigation, five domains still carry at least a 10% chance of catastrophic outcomes, defined as a million deaths or $100 billion in losses. Dangerous capabilities and AI-enabled weapons top the list. mitsloan.mit.edu
What's trending in AI, from the app charts to the community feeds.
Anthropic ended three weeks of deadline extensions late last week: since Monday, Fable 5 is a permanent part of Max and Team Premium plans at half limits, while Pro users get a one-time $100 credit and then pay per token. Simon Willison's read is that GPT-5.6 made subscriptions without your best model commercially unviable. YouTube explainers on what changed are pulling views at over 1,200 an hour in our tracking. simonwillison.net
Microsoft 365 Copilot has spent the week rising through the US Productivity chart and now sits in the mid-20s in our App Store tracking. People are installing their work AI on their own phones. apps.apple.com
That's the shot. — Alexis · AI Weekly