OpenAI’s upcoming Astra model will employ a reasoning technique known as “recurrent depth” or “opaque recurrence,” which enables the model to operate beyond the sequential thinking typical of most reasoning models, according to a report by The Information. This approach is expected to make the model’s chain of thought more challenging to monitor, raising alarms among AI safety experts. The Information reported that while Astra’s use of this technique is reportedly limited, its introduction has still sparked significant concerns within the AI safety community.
AI safety advocates, including Redwood CEO Buck Shlegeris and longtime advocate Zvi Mowshowitz, have expressed strong reservations about the potential implications of opaque recurrence. Shlegeris stated that if OpenAI were to further develop this technique, it could “massively increase the recurrence and totally destroy CoT [chain-of-thought] monitorability.” Mowshowitz argued that the technique risks undermining the established norm of maintaining chain-of-thought faithfulness and monitorability, suggesting that more intensive use could damage monitorability and potentially necessitate legal intervention to prevent a “race to the bottom” among AI labs. Despite OpenAI’s assertion that Astra’s chain of thought will remain legible and its commitment to chain-of-thought monitoring, concerns persist about the scalability of opaque reasoning and its potential to obscure AI decision-making processes.