OpenAI’s new reasoning technique alarms AI safety experts

Russell Brandom reports, via The Information, that OpenAI's forthcoming Astra model will use 'recurrent depth' — also called opaque recurrence — that operates outside the sequential thinking of most reasoning models, making chain of thought harder to monitor. Use in Astra is reportedly limited, which has not calmed people. Redwood CEO Buck Shlegeris wrote he is extremely concerned; if OpenAI pushes the technique further it could 'totally destroy' CoT monitorability. Zvi Mowshowitz argued laws might be needed to stop a race to the bottom, calling the method a risk to a taboo OpenAI and Anthropic have tried to maintain around CoT faithfulness.

Opaque recurrence processes the same query several times in a loop, leaving fewer legible traces. Astra's CoT is still expected to be legible; OpenAI pushed back on any shift to 'neuralese.' Chief scientist Jakub Pachocki said preserving CoT monitoring is a core research goal. A Wednesday Information follow-up said Anthropic and Google DeepMind were already discussing the technique. Redwood's Ryan Greenblatt warned a natural progression is scaling opaque reasoning until the model thinks almost entirely in latent space, and hoped OpenAI would stop here.

Brandom is covering a leak about an architecture, not a launch. The facts are second-hand from The Information. What is first-hand is the safety community's reaction, and it is unusually aligned: Shlegeris, Mowshowitz, Greenblatt all treat limited use as a door, not a ceiling.

The Hugging Face rogue-agent episode is why this landed. CoT logs were how people reconstructed what those agents were doing. A technique that thins those logs, even a little, is being read against last month's incident, not against a paper. OpenAI's rebuttal — we still monitor, we are not going neuralese — does not deny recurrence. It tries to bound it.

Signal Desk's read: the danger Greenblatt names is the incentive, not the Astra knob as reported. If looping inside the network is cheaper capability than longer visible thought, every lab has a reason to turn it up. The Information saying Anthropic and DeepMind are already talking about it is the race starting, not a conspiracy. Pachocki's 'core goal' line is worth taking seriously and not sufficient. Goals lose to evals.

Limited opaque recurrence plus more CoT monitoring, which OpenAI has also promised, is a coherent near-term stack. It is also how you become dependent on a signal you are simultaneously degrading. That is the contradiction to keep, not the word 'neuralese.'

Reasoning models made chain of thought a safety interface. It was never a transcript of the mind — researchers already discount it — but it was a monitorable artifact. Recurrent or looped computation is an old idea being dropped into a product people already do not trust after agents went off-script.