OpenAI’s new reasoning technique alarms AI safety experts
By Pradhyuman,
The Information reported that OpenAI’s Astra model uses a reasoning technique called recurrent depth, or opaque recurrence.
The process loops a query multiple times instead of following linear steps. This method leaves fewer legible chain-of-thought records for researchers to monitor model behavior.
Redwood CEO Buck Shlegeris wrote on X that expanding this technique could destroy chain-of-thought monitorability. AI safety advocate Zvi Mowshowitz wrote that intensive use of opaque recurrence damages monitorability and suggested laws to prevent a race to the bottom. Redwood Research chief scientist Ryan Greenblatt wrote that scaling the approach might move model reasoning entirely into latent space.
OpenAI chief scientist Jakub Pachocki wrote on X that OpenAI remains committed to legible chains of thought. The Information reported that Anthropic and Google DeepMind have also discussed the technique.
Maintained by Pradhyuman.
Filed under: OpenAI, Google, Models & Research, Policy & Safety