OpenAI Chief Scientist Jakub Pachocki stated that he hopes to avoid a race towards “unmonitorability” of models due to misleading reports. He mentioned that current cutting-edge models, including Astra, have a computational graph depth that differs from GPT-4 by no more than a factor of two. Since the launch of the first inference models, OpenAI has been committed to preserving and utilizing chain-of-thought monitoring, believing that this technology helps observe how the model’s alignment capabilities generalize from the training distribution.

Pachocki also pointed out that chain-of-thought monitoring is relatively fragile, and the trend is not optimistic, but it can still be strengthened through research and has become a core objective of OpenAI’s current research agenda. Previously, The Information reported that Astra uses recurrent depth, allowing the same set of Transformer layers to compute repeatedly. As a result, some inferences can occur internally within the model, without needing to be entirely expressed as a chain of thought, raising concerns about whether the model will become increasingly difficult to monitor.