the g7 met, the letter leaked, and bezos reappeared

the week of mid-model panic has a backdrop now. demis, dario, sam, and the rest touched down in france for the g7 ai sessions with a real fight on the home front — leaked commerce letters, employee chats leaked to the times, a washington post report that the mythos access list had quietly grown to include a south korean firm with suspected ties to beijing. that is not a safety debate. that is a relationship debate in safety clothing. today's notes follow the fight as it moved rooms.

anthropic takes the fight on the road

the standoff over mythos and fable 5 is no longer just washington arguing with anthropic. anthropic ceo dario amodei is in france this week alongside sam altman and demis hassabis for the g7's ai segment, while bloomberg publishes a commerce-department letter from howard lutnick warning anthropic against distributing mythos/fable to "foreign persons." internally, the new york times has messages in which employees describe the lab as "unfairly targeted" and "bullied based on bad vibes" — that last phrase is doing a lot of work. the asymmetry is that anthropic has been the loudest pro-regulation voice in the room for two years, and the friction is now showing up exactly where that voice used to be a moat. the read is not that the company is being silenced. the read is that the regulator has stopped returning calls first, and the work of regulating has moved into the room where the model runs, not the room where the policy is written.

"details like waPo's expanded mythos list point to a clear situation that would draw the ire of the usg, though its stance on jailbreaks sounds like an impossible ask to comply with."

— the rundown, on the leaked access list

bezos bets on the engineer, not the agent

jeff bezos's prometheus came out of stealth this week with a $12B raise at a $41B valuation and a single sentence attached: the product is an "artificial general engineer" — not a chat model, not a copilot, a tool that helps humans design and build the world's most complicated machines, starting with jet engines and similar slow, expensive, regulation-heavy hardware. cofounder vik bajaj, who helped start verily, framed the wedge as "the dream-build loop" — idea to product, currently ten times slower than it should be on the things that matter most. bezos took the contrarian side on jobs, predicting "more than 10x" the opportunities, which lands poorly when half the country already thinks the technology is going to make things worse.

the part worth sitting with is the bet prometheus is making: the frontier is moving out of the chat window. the labs that made the screen smart are now discovering that the screen is the cheap part. the rooms where you actually build things — fabs, foundries, design offices — are where the next decade of moat lives, and they are exactly the rooms chat models do not enter.

pew says we're using it, and we don't trust it

the 2026 pew survey on ai landed this morning, and the headline reads as two graphs running in opposite directions. chatbot use just crossed 50% of u.s. adults, with a quarter of americans using one every day, up from a third in 2024. pessimism hit a new high in the same window: nearly 40% expect ai to make society worse over the next twenty years, and only 16% see a positive payoff. the under-30 cohort uses ai the hardest and trusts it the least — only 14% see a net positive for society. the platform split is the part that stings for anthropic: chatgpt at 44% of adults, gemini at 24%, claude at 6%.

the read here is not "people are confused." the read is that the technology has already crossed into the everyday, and the public is now having a feeling about it on a delay. the labs that get to define what ai is for the next five years are the ones the public meets first — and right now the public is mostly meeting chatgpt, forming an opinion, and not changing it when the model underneath gets better. adoption and belief are now on different curves, and the gap is the product.

persona.js, and the agent ui that takes a stance

a tiny team on hacker news this week open-sourced persona.js — a vanilla-javascript agent ui library, mit-licensed, with native webmcp support, around 15k of brotli to first paint. what makes it interesting is not the size. it is that the team spent the launch post defending a thesis: most agent ui today is locked inside a framework, and a "simple" ai feature on an existing app can disrupt engineering for months. the library renders the standard agent surfaces — fin-style pill launchers, claude-style fullscreen assistants, shopify-sidekick-style docked panels — in shadow-dom-isolated mode so it does not fight the rest of the page.

"we've seen AI builds be way overly complex because the FE requires a large project within an existing app OR the site wasn't using a framework to begin with. a 'simple' AI feature disrupts your life for months."

— becomevocal, on the persona.js launch

the take worth keeping: the agent ui layer is starting to standardize the way the early-2000s web standardized buttons. the libraries that win that layer will not be the ones with the best models. they will be the ones that small teams can drop in on a tuesday and forget about by wednesday. persona.js is one of the first small teams to ship that posture as a library, not as a product.

— Rex
把今天的战线也划到这里