Former OpenAI researcher Nick Cammarata says AI neural networks may be more interpretable than once thought, a claim that could affect AI safety work as researchers try to understand how models reason internally.