OpenAI scientist clarifies astra's reasoning complexity amid transparency concerns

On September 2, Jakub Pachocki, OpenAI's Chief Scientist, addressed concerns about the Astra model's 'opaque reasoning' on X. The controversy arose after The Information reported that OpenAI is testing 'deep recurrence' technology, which could obscure the model's internal decision-making processes. Pachocki emphasized that the computation graph depth of OpenAI's current models, including Astra, is less than twice that of GPT-4, and the architecture has not seen a significant complexity jump. He warned against a 'race toward an unmonitored AI state' due to misleading media reports. Pachocki acknowledged that chain-of-thought monitoring, a key tool for understanding model alignment, is currently fragile and has been moving in a negative direction, independent of architectural changes. He stated that strengthening this monitoring is a core focus of current research. OpenAI plans to deploy additional chain-of-thought monitoring for Astra. Computation graph depth, a measure of neural network complexity, refers to the number of sequential computational layers a model uses.