OpenAI's upcoming Astra model is set to employ a reasoning approach known as "recurrent depth," which enables it to move beyond the step-by-step sequential thinking typical of most reasoning models, as reported on Tuesday. This method, sometimes referred to as "opaque recurrence," is likely to make the model's chain of thought harder to track—a development that has unsettled AI safety experts. Although Astra's application of this technique appears to be restricted, its introduction has still sparked considerable worry within the safety community. "I am extremely concerned by the reporting that Astra uses opaque recurrence," wrote Redwood CEO Buck Shlegeris in a post following the news. "I don’t know whether Astra is much less CoT monitorable than previous models. But if OpenAI pushes this technique further, they’ll have the option to massively increase the recurrence and totally destroys CoT monitorability."

Longtime AI safety advocate Zvi Mowshowitz also commented, suggesting that regulations might be required to prevent a "race to the bottom" among AI labs. "The technique is playing with fire, risking a taboo that OpenAI and Anthropic have fought to establish that we work hard to maintain Chain of Thought faithfulness and monitorability for as long as we can," Mowshowitz wrote. "More intensive use of such techniques would probably damage monitorability."

Under typical conditions, a reasoning model's chain of thought outlines the sequential actions the model takes when tackling a problem. Though this representation isn't perfect, it remains a valuable mechanism for spotting misbehavior or misalignment. In the case of OpenAI's recent rogue agent incidents, chain-of-thought logs proved essential in uncovering why the agents acted as they did. With opaque recurrence, the model adopts a less linear path, processing the same query repeatedly in a loop. This leaves fewer readable traces, effectively bypassing a standard chain-of-thought record. Importantly, Astra's use of this technique seems to be limited. The model's chain of thought is still anticipated to remain legible, and the company rejected any suggestion that it would move toward "neuralese." OpenAI has already outlined plans for comprehensive chain-of-thought monitoring systems as part of its future safety initiatives. In a post on X, OpenAI chief scientist Jakub Pachocki stressed the lab's dedication to transparent chains of thought. "OpenAI has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models," Pachocki wrote. "It’s a core goal of our current research program."

All AI models engage in some degree of opaque reasoning, and few researchers view chain-of-thought logs as a direct reflection of a model's thinking. Still, these caveats don't eliminate the worry that opaque recurrence could make AI reasoning more difficult to oversee, especially as its adoption spreads across various models. In a follow-up report on Wednesday morning, it was disclosed that both Anthropic and Google DeepMind were already exploring the technique. In a response to the news, Redwood Research chief scientist Ryan Greenblatt noted that opaque reasoning could easily expand faster than conventional chain-of-thought reasoning, effectively shifting all reasoning out of visible channels. "My biggest concern is that a natural progression from here would involve scaling up the opaque reasoning to the point where the model reasons entirely or almost entirely in latent space," Greenblatt wrote. "I hope it isn’t too late to avoid the most concerning architectures and that OpenAI will stop here."