Mechanistic interpretability

i.e. understanding what complex ML models are doing