OpenAI’s upcoming Astra model is using a reasoning technique that has safety researchers publicly worried. Called “recurrent depth” or “opaque recurrence,” the method lets the model loop through the same query multiple times instead of following a linear, step-by-step chain of thought. The practical result: fewer legible traces of how the model arrived at an answer.
That loss of visibility is the core of the concern. Redwood CEO Buck Shlegeris didn’t mince words after The Information broke the story Tuesday. “I am extremely concerned by the reporting that Astra uses opaque recurrence,” he wrote. His worry isn’t necessarily what Astra does today — the technique’s use is reportedly limited — but where it leads. If OpenAI scales up the recurrence, Shlegeris argues, the company could “totally destroy” chain-of-thought monitorability. Longtime safety advocate Zvi Mowshowitz went further, suggesting laws might be needed to prevent a “race to the bottom” among labs abandoning interpretability.
OpenAI is pushing back. Chief scientist Jakub Pachocki posted on X that preserving and using chain-of-thought monitoring has been a core goal since the company’s first reasoning models. The company has also announced plans for extensive chain-of-thought monitoring systems in its forward-looking safety work. And to be fair, no serious researcher treats chain-of-thought logs as a perfect mirror of model reasoning — every model does some amount of opaque processing under the hood. The logs are more like a useful approximation than a transcript.
Still, the timing matters. Chain-of-thought records were instrumental in analyzing OpenAI’s recent rogue agent incidents, helping researchers understand why agents behaved the way they did. Remove that tool and you’re flying blind. The Information reported Wednesday that Anthropic and Google DeepMind are already discussing the technique internally. Redwood chief scientist Ryan Greenblatt framed the nightmare scenario plainly: opaque reasoning could scale faster than conventional chain-of-thought, eventually pushing all reasoning into latent space where no human can inspect it. “I hope it isn’t too late,” he wrote. That’s the kind of sentence that sticks with you.
