OpenAI Chief Scientist Clarifies Model Opacity Concerns and Computation Depth
OpenAI Chief Scientist Jakub Pachocki responded to concerns about model opacity, clarifying that frontier models like Astra have not experienced an exponential increase in architectural complexity. He stated that Astra's computation-graph depth is less than twice that of GPT-4 and affirmed that Chain-of-Thought (CoT) monitoring remains a primary research priority. As frontier AI models adopt architectures like recurrent depth, fears have grown that internal reasoning could become hidden within complex neural states rather than readable text. Pachocki's statement reassures the AI safety community that OpenAI is actively prioritizing interpretability and alignment monitoring as models evolve. Media reports had suggested that recurrent depth—where models reuse network layers to update internal representations—might obscure a model's Chain-of-Thought. Pachocki acknowledged that CoT monitorability is currently fragile due to non-architectural factors, but emphasized that strengthening CoT monitoring is a core objective of OpenAI's current research.
## BACKGROUND
Standard Transformer architectures process information through a linear series of neural network layers, often outputting step-by-step reasoning in natural language, known as a Chain of Thought (CoT). Recurrent depth allows models to repeatedly reuse the same network layers to refine internal representations, which can leave more reasoning in hidden activation states rather than explicit text. Monitoring CoT outputs is a key technique in AI safety, allowing researchers to evaluate an AI system's intent and reasoning before it takes action.