TechCrunch/The Information: Astra uses recurrent-depth reasoning; safety experts warn CoT monitoring gets harder

Per The Information, reported by TechCrunch and The Verge, OpenAI's upcoming Astra uses “recurrent depth” (also called opaque recurrence / a looped transformer): cycling computation through internal layers before output, leaving fewer legible natural-language chain-of-thought traces. That can boost performance while making misbehavior harder to monitor. Redwood's Ryan Greenblatt called it potentially “the single worst development for AI security/safety to date”; Zvi Mowshowitz and others warned of a race to the bottom on monitorability. OpenAI says use is limited and CoT remains legible; chief scientist Jakub Pachocki stressed CoT monitoring and said Astra's compute depth is within about a factor of two of GPT-4. TechCrunch says a Wednesday-morning Information follow-up reported Anthropic and Google DeepMind are already discussing the technique.