TLDR Tech / InfoSec
OpenAI's new reasoning technique alarms AI safety experts
OpenAI's Astra model uses "recurrent depth," letting it loop through the same query multiple times instead of reasoning step by step, which makes chain-of-thought logs harder to read. Scaling the technique could push reasoning fully into latent space, killing visibility into model behavior. OpenAI's Jakub Pachocki said Astra's chain of thought stays legible and that monitoring remains a research priority.
Read the full story