According to GoogleDeepMind on X, Neel Nanda outlines mechanistic interpretability, chain of thought monitoring, and safety auditing in a new podcast.