AI Tech / news
OpenAI's 'Recurrent Depth' Technique in Astra Model Alarms AI Safety Experts
OpenAI plans to deploy a reasoning method called 'recurrent depth' in its upcoming Astra model, a move that has safety experts worried about the visibility of the AI's chain of thought.
Recurrent depth, also described as opaque recurrence, allows Astra to perform computation outside the sequential thinking used by most reasoning models. This design choice could make the model's chain of thought — the logged reasoning steps safety researchers examine — harder to monitor.
Astra's use of the technique appears limited, but its existence has rattled safety experts. Among them, Buck Shlegeris, CEO of research outfit Redwood, said he was 'extremely concerned' by the reporting on Astra's approach, noting that if OpenAI pushes the technique further, the company could massively increase recurrence and undermine chain-of-thought monitorability.
Zvi Mowshowitz, a longtime AI safety advocate, echoed those worries and suggested that laws may be needed to prevent a 'race to the bottom' among AI labs. He cautioned that the technique risks breaking a norm that OpenAI and Anthropic have sought to establish: keeping chain-of-thought reasoning faithful and monitorable for as long as possible.
The model itself is OpenAI's next-generation system, and the approach has become the latest front in the debate over how much transparency AI developers owe the public as models grow more powerful.