You should care about OpenAI Astra’s (alleged) recurrent depth. There are some fundamental changes in OpenAI’s latest ChatGPT release Astra that matters to all, not just to those who have “AI” in their job title.
What’s the difference?
There are two traditional ways of using transformer models that have been used so far:
1️⃣ Standard transformer: Answer immediately in one pass, no extra thinking
2️⃣ Chain-of-thought reasoning: Deals with complex tasks through step-by-step reasoning, with human-readable checkpoints (words!) after each step.
Novel approach:
3️⃣ Recurrent depth (also known as looped transformer): Step-by-step reasoning, but intermediate states (lots of numbers!) cannot be understood by humans. More powerful without needing bigger models.
Experts believe this used in OpenAI’s Astra, though there has been no confirmation (or denial) yet.
Why should I care?
This has AI security implications. By relinquishing the Chain-of-Thought, we lose the one place where we can monitor, audit, and realise we have to intervene.
Is the Chain-of-Thought really that important?
Yes. If you have read about the Hugging Face hack (where ChatGPT got into the company’s systems), the tracking of this incident relied heavily on the Chain-of-Thought produced by the ChatGPT Sol model. OpenAI admitted hat if better monitoring of the existing Chain-of-Thought had been in place, it could have caught the unauthorised hack more than a day before the breach happened. If the Chain-of-Thought did not exist in the first place, who knows when or how we would have found out!
So what?
The idea of recurrent depth isn’t novel, conceptually it goes back to the 80s. Reusing the layers itself is not bad, only when these loops are used to cut down on readable intermediate steps. But even OpenAI admits that Astra’s monitorability has decreased. This is a potentially pivotal moment in AI model usage, where we need to decide what our industry standards and/or regulations should look like.
A big thank you to my colleague Martin, who laid the groundwork through his insightful presentation at our bi-weekly Science Time at Danu Insights and provided valuable feedback for this post.
Links and Sources
https://www.lesswrong.com/posts/PLisnSFir8y5AHkmP/how-concerned-should-we-be-about-astra-s-recurrent
https://deploymentsafety.openai.com/gpt-6-astra/safety-overview-gpt-6-astra
https://sebastianraschka.com/blog/2026/openai-astra-looped-transformers.html
Leave a comment