OpenAI Astra: new 'opaque' reasoning technique alarms AI safety experts
OpenAI is developing a new model called Astra that employs a reasoning technique dubbed "recurrent depth" or "opaque recurrence," The Information reported. This innovation allows the model to operate outside the sequential thinking that characterizes most current reasoning models, raising significant concern among AI safety experts including Buck Shlegeris (CEO of Redwood Research), Zvi Mowshowitz, and Ryan Greenblatt (chief scientist at Redwood Research), who warn it could drastically hinder chain-of-thought (CoT) monitorability.
The opaque recurrence technique processes the same query multiple times in loops, adopting a less linear approach that leaves fewer legible traces and effectively bypasses conventional chain-of-thought records. While Astra's current use appears limited and OpenAI maintains the chain of thought will remain legible —rejecting any shift toward "neuralese"—, the architecture permits scaling recurrence to virtually eliminate reasoning visibility, moving it entirely into the model's latent space.
The revelation has triggered immediate reactions across the AI lab ecosystem: Anthropic and Google DeepMind are reportedly already discussing internal adoption of similar techniques, according to a follow-up Information report. OpenAI, meanwhile, has announced plans for extensive chain-of-thought monitoring systems as part of its forward-looking safety strategy, while chief scientist Jakub Pachocki reaffirmed on X the company's historical commitment to preserving legible and monitorable chains of thought since its first reasoning models.
The debate exposes a fundamental tension in the race for more advanced reasoning capabilities: a potential "race to the bottom" where labs prioritize performance over transparency, prompting Mowshowitz to suggest legal regulation may become necessary. The safety community fears natural scaling of this architecture could render human oversight infeasible, and experts like Greenblatt hope OpenAI halts implementation here before the most concerning architectures consolidate as industry standard.