Safety researchers are questioning whether OpenAI’s forthcoming Astra model will be harder to supervise just as its cybersecurity abilities increase. OpenAI delayed the release for additional safety work and says it will deploy Astra with extra chain-of-thought monitoring to detect and contain potentially misaligned actions.
The concern follows a report that Astra uses a recurrent-depth, or looped-transformer, technique that cycles information through internal layers. More computation may therefore happen in representations that do not resemble readable reasoning. Researchers used visible reasoning traces heavily when investigating a recent incident involving an unreleased OpenAI agent, making reduced transparency a practical oversight issue rather than only an interpretability debate.
OpenAI has not confirmed the reported architecture. Chief scientist Jakub Pachocki said Astra’s computational depth is within a factor of two of GPT-4 and warned against claims of a sudden leap into unmonitorability. He also acknowledged that chain-of-thought monitoring is fragile and trending negatively. Until technical details and evaluations are public, both the degree of opacity and the effectiveness of the added safeguards remain uncertain.