Good morning. The big story today is about containment: OpenAI says its frontier models will now train without live internet access. Elsewhere, Google changed its ad policy for people who connect personal accounts to its AI features, a year-old lab raised $1 billion to chase data-efficient training and Block showed a voice model trained only on speech. We also cover Meta’s delay of its publisher-contested terms, plus a set of explainers on research tools, reward hacking, data analysis, data center power and where to start learning how language models work.
Update: this edition now also covers Moonshot’s Kimi K3.1 Pro, the METR and Redwood audit of OpenAI’s monitoring, Anthropic’s formal-verification project with Theorem and OpenAI’s covert-channel tests, plus a wearer-only AI pendant, a conference calendar and several explainers.