OpenAI's new Astra model will use a reasoning technique called "recurrent depth," also known as "opaque recurrence," which allows it to operate outside the sequential thinking typical of most reasoning models, according to a report by The Information. This technique makes the model's chain of thought harder to monitor, raising concerns among AI safety experts.

Redwood CEO Buck Shlegeris expressed extreme concern, warning that if OpenAI pushes the technique further, it could "massively increase the recurrence and totally destroy Co T monitorability." Longtime AI safety advocate Zvi Mowshowitz called the technique "playing with fire" and suggested laws might be needed to prevent a "race to the bottom" among AI labs.

Under normal circumstances, a reasoning model's chain of thought provides sequential steps for monitoring misbehavior. In opaque recurrence, the model processes queries in a loop, leaving fewer legible traces. However, Astra's use of the technique appears limited, and OpenAI has pushed back against suggestions it would shift to "neuralese," emphasizing its commitment to legible chains of thought through chief scientist Jakub Pachocki.

The Information later reported that both Anthropic and Google Deep Mind are already discussing the technique. Redwood Research chief scientist Ryan Greenblatt warned that opaque reasoning could scale faster than conventional methods, potentially removing all reasoning from visible channels.