Good morning. It is Monday, October 19, and the weekend brought a run of stories about who holds power over frontier AI: the companies that build it, the outsiders who test it and the regulators who may soon police it.
The big three
Anthropic dominated the weekend. It announced a Public Benefit Fund that will receive 5% of its shares, it brought outside evaluators from METR inside the company, and its red team warned that small models could someday copy themselves across ordinary smart devices. Each story is covered in full above, and each carries caveats worth reading before drawing conclusions: the fund’s governance is still undefined, METR’s findings have not been published, and the worm warning is a forecast from experiments.
Money and infrastructure
Google’s Gemini 4 Argon is now rolling out, but only for paying users, and its low API price has an expiry date. Microsoft’s exit from a UK datacentre project is a reminder that electricity, not chips, is increasingly the limiting factor. NVIDIA’s decision to open-source Run:ai gives cloud providers a shared tool for squeezing more out of expensive GPUs.
Rules, rights and deals
An unconfirmed report says the FTC is investigating Anthropic and OpenAI. Meta backed away from terms that publishers feared would open their sites to AI training. Disney is picking up Character.AI’s people and possibly its technology in a bankruptcy sale.
Research worth your time
Today’s Paper of the Day asks whether the brain could assemble its own learning machinery, and Ai2’s new emulator shows how AI can generate extreme-weather data far more cheaply than physics models.
Also today
Our evergreen pieces cover how to write a system prompt, why the same prompt gives different answers, and cloud APIs versus local models.