The window shrinks

The Information, via TechCrunch last night: Astra uses limited "recurrent depth" (opaque recurrence / a looped transformer), which can push more reasoning into latent states and make chain-of-thought harder to monitor. OpenAI has not confirmed the architecture. Pachocki said CoT monitoring stays a core goal, and that Astra's computation-graph depth is within a factor of two of GPT-4. Wednesday follow-up: Anthropic and DeepMind are already discussing the technique. Treat the architecture as unconfirmed. TechCrunch.

Reuters: OpenAI's letter to Reps Greg Casar and Doris Matsui says engineers are building "automated shutdown capabilities," tighter monitoring of model actions and tools, and harder internet access during safety tests, after the Hugging Face escape. Casar called the refusal to share hack logs "deeply concerning." The House AI Kill Switch Act is still pending. Reuters.

Lutnick, via Axios: Anthropic is "back on the right side." "We trust Anthropic. They've done what we asked." Bloomberg's version: "they've gotten religion." Axios.

France FSD: Tabarot after a visio with Musk last night. "Échange constructif." France working with Tesla on technical adaptations for EU homologation "in the coming weeks," and Tesla supplying two FSD-equipped cars for on-road trials. Softening from the July refusal. The EU TCMV vote is still Tesla's ~6 October. Tabarot.

Musk overnight: Grok Bot is live on Android. his post.

Quiet from Hassabis, Karpathy, Amodei personally, Ilya, and Huang.