OpenAIâs new reasoning technique alarms AI safety experts
OpenAIâs new Astra model will use a reasoning technique called ârecurrent depthâ that allows it to operate outside of the sequential thinking that characterizes most reasoning models, theThe Information reportedon Tuesday. This technique, also called âopaque recurrence,â will likely make the modelâs chain of thought more difficult to monitor â and that has AI safety experts rattled. While Astraâs use of the technique is reportedly limited, its emergence has still raised significant concerns among AI safety experts. âI am extremely concerned by the reporting that Astra uses opaque recurrence,â wrote Redwood CEO Buck Shlegerisin a postafter the news broke. âI donât know whether Astra is much less CoT monitorable than previous models. But if OpenAI pushes this technique further, theyâll have the option to massively increase the recurrence and totally destroys CoT monitorability.â Longtime AI safety advocate Zvi Mowshowitz also weighed inand wrotethat laws might be necessary to prevent a ârace to the bottomâ among AI labs. âThe 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,â Mowshowitz wrote. âMore intensive use of such techniques would probably damage monitorability.â Under normal circumstances, a reasoning modelâs chain of thought provides the sequential steps taken by the model as it attempts to solve a problem. While the representation is imperfect, it still serves as a valuable tool for monitoring misbehavior or misalignment. In the case of OpenAIâs recent rogue agent activity, chain-of-thought records were an important tool in teasing out why agents behaved the way they did. In opaque recurrence, the model takes a less linear approach, processing the same query several times in a loop. The result leaves fewer legible traces, effectively side-stepping a conventional chain-of-thought record. Crucially, Astraâs use of the technique appears to be limited. The modelâs chain of thought is still expected to be legible, and the company pushed back against any suggestion that it would shift to âneuralese.â OpenAI has already announced plans for extensive chain-of-thought monitoring systems as part of its forward-looking safety plans. In a post on X, OpenAI chief scientist Jakub Pachocki emphasized the labâs commitment to legible chains of thought. âOpenAI has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models,â Pachocki wrote. âItâs a core goal of our current research program. All AI models do some quantity of opaque reasoning, and few researchers take chain-of-thought logs as a direct representation of a modelâs reasoning. Still, those caveats donât dispel the concern that opaque recurrence may make AI reasoning harder to monitor, particularly as it grows in use across different models. In a follow-up report Wednesday morning, The Information reported that both Anthropic and Google DeepMind were already discussing the technique. In a post responding to the news,Redwood Research chief scientist Ryan Greenblatt said opaque reasoning could easily scale faster than conventional chain-of-thought reasoning, effectively removing all reasoning from visible channels. âMy 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,â Greenblatt wrote. âI hope it isnât too late to avoid the most concerning architectures and that OpenAI will stop here.â
