This is very interesting. I don't see much discussion of interpretability in day to the day discourse of AI builders. I wonder if everyone assumes it to either be solved, or to be too out of reach to bother stopping and thinking about.