OpenAI's Astra Model Uses 'Opaque Recurrence,' Alarming Safety Experts
Original: OpenAI’s new reasoning technique alarms AI safety experts
Why This Matters
Reduced AI reasoning transparency could weaken safety oversight at a critical stage of advanced model deployment.
OpenAI's upcoming Astra model employs a reasoning technique called 'recurrent depth' or 'opaque recurrence,' which processes queries in loops rather than linear steps, reducing the legibility of chain-of-thought records and raising significant concerns among AI safety researchers.
OpenAI's new Astra model will use a reasoning technique called 'recurrent depth' — also known as 'opaque recurrence' — that allows it to process queries in iterative loops rather than the sequential steps typical of most reasoning models, according to a report by The Information. Unlike conventional chain-of-thought (CoT) reasoning, which produces legible sequential traces used to monitor model behavior and detect misalignment, opaque recurrence leaves fewer readable logs, potentially undermining oversight mechanisms.
Redwood Research CEO Buck Shlegeris stated he was 'extremely concerned' by the reports, warning that if OpenAI scales the technique further, it could 'totally destroy CoT monitorability.' AI safety commentator Zvi Mowshowitz suggested that legislation may be needed to prevent a competitive 'race to the bottom' among AI labs, calling the technique 'playing with fire.'
OpenAI pushed back against the alarm. Chief Scientist Jakub Pachocki posted on X that the company has been committed to chain-of-thought monitoring since its first reasoning models, calling it 'a core goal of our current research program.' The company also clarified that Astra's use of the technique is limited and that the model's CoT remains largely legible. OpenAI has previously announced plans for extensive CoT monitoring systems.
A follow-up report from The Information indicated that both Anthropic and Google DeepMind are already internally discussing opaque recurrence, suggesting the technique may spread across the industry.