OpenAI's Astra model will use a reasoning technique called 'recurrent depth' that makes its thinking harder to monitor.
AI safety experts are alarmed, warning of a 'race to the bottom.' Anthropic and Google DeepMind are also discussing the technique.
What happened
OpenAI's new Astra model will reportedly use a reasoning technique called 'recurrent depth,' also known as 'opaque recurrence,' according to The Information. This technique processes the same query several times in a loop, leaving fewer legible traces than standard sequential chain-of-thought reasoning.
Why it matters
AI safety experts, including Redwood CEO Buck Shlegeris and advocate Zvi Mowshowitz, are concerned that this technique could make the model's reasoning harder to monitor. Shlegeris warned that pushing the technique further could 'totally destroy' chain-of-thought monitorability, while Mowshowitz suggested laws might be needed to prevent a 'race to the bottom' among AI labs.
What to watch
Astra's use of the technique appears limited, with the chain of thought still expected to be legible. However, The Information also reported that Anthropic and Google DeepMind are already discussing the technique. Redwood Research's Ryan Greenblatt expressed hope that OpenAI will 'stop here' to avoid scaling up opaque reasoning to a point where models reason entirely in latent space.
Ask the AI about this article →
The report about Astra's use of 'recurrent depth' has sparked concern among AI safety experts because it challenges the transparency that chain-of-thought monitoring provides. This technique processes queries in a loop, potentially bypassing the sequential records that have been crucial for detecting misbehavior, as seen in past incidents of rogue AI agents. Safety advocates like Buck Shlegeris and Zvi Mowshowitz fear that if OpenAI scales up this opaque reasoning, it could undermine the industry's commitment to maintaining legible and monitorable AI reasoning.
OpenAI has pushed back, stating that Astra's use of the technique is limited and that the chain of thought will remain legible. Chief scientist Jakub Pachocki emphasized the lab's dedication to chain-of-thought monitoring since its first reasoning models. However, the fact that other major players like Anthropic and Google DeepMind are also discussing the technique suggests it could become more widespread, raising the stakes for regulatory and safety frameworks.
For example, today's edition would include:
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. Q&As are published on this page for other readers too.
Interactive Brokers has begun connecting its platform with AI tools including ChatGPT, Claude, and Grok, and o…
A new report by Alipay+ and S&P Global, based on a survey of 6,000 consumers across nine markets in Asia, Euro…
CrowdStrike Holdings Inc
Google has reportedly approached major studios such as Disney, Warner Bros
Google DeepMind introduced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber
IBM released a survey showing AI is already used weekly in 76% of middle school and 73% of high school classro…
