OpenAI's Chief Scientist Responds to Astra Controversy: New Architecture Did Not Suddenly Turn the Model into a Black Box

Beating AI Express Insight: OpenAI Chief Scientist Jakub Pachocki has refuted some external interpretations of the Astra recurrent architecture. He stated that for OpenAI’s current frontier models, including Astra, the depth of their computational graphs is at most only twice that of GPT-4, rather than having suddenly increased by several orders of magnitude. This is a direct response to a previous report from The Information, which revealed that Astra uses recurrent depth, allowing the same set of Transformer layers to compute repeatedly. As a result, part of the inference process can occur internally within the model, without needing to be fully written out as Chain-of-Thought (CoT), sparking concerns about whether models will become increasingly difficult to monitor. Pachocki acknowledged that CoT monitoring is indeed fragile and is moving toward a direction that is harder to oversee. However, he clarified that this is not due to changes in model architecture. OpenAI has retained and used CoT monitoring since its first-generation reasoning models, and currently still lists enhancing this capability as a core research goal.