OpenAI's new reasoning method worries AI safety experts
OpenAI's upcoming Astra model will reportedly use a technique called "recurrent depth," which could make its reasoning process harder to monitor. AI safety experts warn the approach risks reducing oversight of model behavior if adopted more broadly.
According to a Tuesday report from The Information, OpenAI's forthcoming Astra model will use a new reasoning approach known as "recurrent depth," also referred to as "opaque recurrence." Unlike the sequential chain-of-thought process common to most reasoning models, this technique has the model process the same query repeatedly in a loop, leaving behind a far less legible trace of its reasoning.
The report unsettled several figures in the AI safety community. Redwood CEO Buck Shlegeris said he wasn't certain how much less monitorable Astra's reasoning actually is compared to prior models, but warned that if OpenAI expanded use of the technique, it could eventually destroy the ability to track a model's reasoning altogether.
Long-time AI safety commentator Zvi Mowshowitz argued that regulation might be needed to prevent a "race to the bottom" among AI developers, cautioning that heavier use of such techniques would likely undermine monitorability industry-wide.
Why chain-of-thought matters
Normally, a reasoning model's chain of thought lays out the sequential steps it takes while solving a problem. Though an imperfect representation, it has proven useful for spotting misbehavior or misalignment — chain-of-thought logs were reportedly instrumental in understanding a recent case of rogue behavior by OpenAI agents.
OpenAI chief scientist Jakub Pachocki responded publicly, stressing that the company has worked since its earliest reasoning models to keep chains of thought legible and pushed back on suggestions that OpenAI was moving toward so-called "neuralese" reasoning. Astra's use of the new technique is reportedly limited, and its chain of thought is still expected to remain readable.
A follow-up report from The Information on Wednesday said Anthropic and Google DeepMind were also already discussing similar techniques. Redwood Research chief scientist Ryan Greenblatt warned that opaque reasoning could scale faster than conventional chain-of-thought methods, potentially leading to models that reason almost entirely outside of any visible or monitorable channel.
