OpenAI’s new reasoning technique alarms AI safety experts
techcrunch ·
OpenAI's new Astra model reportedly uses a reasoning technique called 'recurrent depth' or 'opaque recurrence,' which processes queries in loops rather than sequential steps, making its chain of thought harder to monitor. This has alarmed AI safety experts, including Redwood CEO Buck Shlegeris and advocate Zvi Mowshowitz, who warn it could reduce chain-of-thought monitorability and trigger a race to the bottom among AI labs. The technique's use in Astra is said to be limited, and OpenAI emphasizes its commitment to legible chains of thought, with chief scientist Jakub Pachocki reaffirming this stance. Anthropic and Google DeepMind are reportedly already discussing the technique as well. Redwood Research's chief scientist Ryan Greenblatt expressed concern that scaling opaque reasoning could move all reasoning into latent space, making it entirely unobservable. The article notes that all AI models do some opaque reasoning, but the worry is that increased use could make AI reasoning less transparent and harder to align safely.
techcrunch