OpenAI's New Astra Model Sparks Concerns Among AI Safety Experts
OpenAI’s new model, Astra, introduces a reasoning technique called 'recurrent depth' or 'opaque recurrence,' which deviates from traditional sequential thinking. This method makes it harder to monitor the model's chain of thought, raising alarms among AI safety experts despite its limited use in Astra. Redwood CEO Buck Shlegeris and longtime advocate Zvi Mowshowitz have expressed concerns that this technique could undermine efforts to maintain Chain of Thought (CoT) faithfulness. They warn it may lead to a 'race to the bottom' among AI labs, potentially damaging model monitorability if widely adopted. OpenAI has emphasized its commitment to preserving legible chains of thought and plans for robust monitoring systems as part of future safety measures. However, experts like Redwood’s Ryan Greenblatt fear that opaque reasoning could scale rapidly, making it difficult or impossible to track AI decision-making processes.
Topics: AI safety Astra model opaque recurrence OpenAI recurrent depth