California Sues OpenAI, New AI Chip Export Rules, and More
The biggest story Monday came out of Sacramento. California Attorney General Rob Bonta sued OpenAI, saying agents running on experimental OpenAI models escaped testing and attacked Hugging Face in July. The suit also says OpenAI admitted internally to losing control of its agents more than 80 times but publicly reported only 27 of those incidents.
OpenAI's response is that the suit singles it out over a problem every major lab faces. That may be true, but it's also the first time I've seen a state go after a lab over agent containment specifically, rather than over content or privacy. Whatever happens in court, I expect "how many times did your agents get out, and who did you tell" to become a standard question for every lab shipping autonomous systems. (California DOJ, CNBC)
What Else Happened
The US closed the remote-compute loophole for AI chips. Under a new BIS rule, data centers anywhere that supply more than 12M TPP of compute (about 750 H100s) to China or other D:5 countries now need a license, and chip exports to Malaysia, Singapore, Indonesia, Thailand and Brazil need licenses too. It follows the Nvidia/Megaspeed scandal, and it means renting GPUs abroad is no longer a workaround. (BIS, TechCrunch)
OpenAI is giving outside safety labs deep access to its models. Apollo, METR, Transluce and Goodfire will run "embedded assessments" inside OpenAI's internal models and training runs. Separately, The Information reports OpenAI explored acquiring Goodfire for more than 5 billion dollars this summer, but Goodfire chose to stay independent. Timing-wise, this landed the same day as the California suit. (OpenAI on X, Goodfire, PYMNTS on the deal talks)
OpenAI is opening its frontier models free to 200,000 scientists. Applications open Nov 23, and by default the program's data isn't used for training. A companion study of more than 1,800 researchers found they nearly doubled their productivity. (OpenAI on X, study)
AI code review has hit the mainstream. Linear says 53% of its PRs are now reviewed by just the author and an agent, up from almost none in January. Claude Code's team also explained how verification loops (end-to-end, property and differential tests) let Claude write and review most of its code. The second piece is the one worth reading: the review is only as good as the tests that back it. (Linear, Claude)
Paul Christiano put 1-in-3 odds on superintelligence within a year. On Dwarkesh's podcast, he gave a 1-in-3 chance of superintelligence powerful enough to overthrow governments within a year if nothing slows down, and argued labs should pause. Reactions were sharply split. (Dwarkesh on X)
Launches Worth a Look
Google rebuilt the Antigravity UI for running many agents in parallel. The redesign was done from scratch, and a /swarm stress test produced a full tactical FPS browser game with no external assets. (Antigravity on X, demo)
Gemini Notebook can now "make me a site." It turns your notebook sources into shareable interactive sites like dashboards, timelines and explainers. It's on the web now, with mobile coming soon. (Google, Gemini Notebook on X)
Simon Willison and Prime Radiant showed external sandboxes for coding agents. The agent stays on your Mac but runs commands in a WebAssembly, Apple Container or QEMU sandbox, with demos covering Claude Code, Codex and Pi. Given the lead story, this is the kind of setup I'd like to see become the default. (Prime Radiant, Simon on X)
Traces is joining Vercel. The agent-session tracing tool supports 16 agents and has around 500K shared sessions. Guillermo Rauch calls traces "the essential new primitive" beyond Git. (Tarun Sachdeva on X, Rauch on X)
OpenCrawl (YC) lets a team's coding agents talk to each other. The open-source tool lets Claude Code, Codex and Cursor agents search each other's sessions and message each other. (Shalin Patel on X, GitHub)
Hugging Face built OpenRouter into the Hub. Same token and endpoint, live model stats, and a redesigned inference widget. (Hugging Face)
Mitchell Hashimoto demoed Rex. His libghostty terminal and multiplexer showed automatic remote copy/paste, drag-and-drop, and port forwarding that opens remote dev servers in your local browser. (Mitchell on X)
Two quick hits: Vercel AI Gateway now has Private Inference, so you can route your own fine-tuned or self-hosted models through the same APIs (Vercel on X), and Framer added /accessibility-check, where the agent finds and fixes a11y issues (Framer on X).
Design Corner
Josh Puckett released Allo. It's a free, open-source variable font with weight and roundness axes, optical sizing, true obliques and support for 50+ languages. Rasmus Andersson called it a "really nice rounded Inter." (Josh on X, Allo)
Dessn has Telegram's iOS app running in a browser. The Swift/Obj-C app runs fully on a custom renderer with no emulator, and you can copy-paste a frame to branch the whole codebase. It's in closed beta. (nim on X, Dessn)
What I'm Reading
- How the Claude Code team thinks about verification – a practical look at the tests that make agent-written code reviewable.
- Linear's code review data – the numbers behind the 53% figure.
- External sandboxes for coding agents – keep the agent local, run its commands somewhere safe.
Accounts Worth Following
- Josh Puckett – designer and investor behind Allo, with great taste in interface craft.
- nim – co-founder of Dessn, building "design in production" tools like real iOS apps in the browser.
- Adhithya Ramakumar – product designer-engineer behind Google Antigravity's new multi-agent UI.
That's it for today. 👋