Dave Orr comments on AI #33: Cool New Interpretability Paper