OpenAI Responds to Astra Transparency Concerns: Computation Depth Similar to GPT-4

OpenAI's chief scientist Jakub Pachocki recently responded to concerns about the transparency of the AI model Astra. Earlier reports suggested that OpenAI's testing of 'recurrent depth' technology could make model reasoning harder to monitor. Pachocki emphasized that the computation graph depth of frontier models, including Astra, is less than twice that of GPT-4, and the architecture hasn't seen a sudden increase in complexity. He stated that OpenAI has always valued chain-of-thought monitoring and plans to strengthen it for Astra, which is a core research goal. The 'recurrent depth' technique involves repeatedly using the same set of network layers, unlike traditional linear structures, potentially keeping more of the reasoning process in internal states, making external monitoring more difficult. OpenAI's response aims to clarify the technical developments and reaffirm its commitment to model safety and transparency.