OpenAI Chief Scientist Responds to Astra Controversy: New Architecture Did Not Suddenly Make the Model a Black Box

Dynamic Beating AI Newsletter, OpenAI's Chief Scientist Jakub Pachocki has refuted some of the external interpretations of the Astra recurrent architecture. He stated that, including Astra, OpenAI's current cutting-edge models, computation graph depth, and GPT-4 differ at most by a factor of 2, rather than suddenly increasing by several orders of magnitude.

This is mainly in response to a previous report by The Information. The report revealed that Astra uses recurrent depth, allowing the same set of Transformer layers to be repeatedly computed. As a result, some inferences can occur within the model without having to be entirely written out as a chain of reasoning. This has also raised concerns about whether the model will become increasingly difficult to monitor.

Pachocki acknowledged that CoT monitoring is indeed fragile and is evolving towards being more challenging to monitor. However, he clearly stated that this is not due to changes in the model architecture. OpenAI has been preserving and utilizing CoT monitoring since the first generation of reasoning models and still considers enhancing this capability a core research goal.