This is amazing. Feels like it will be the start of a new era for interpretability research