Julian Bradshaw comments on AI #33: Cool New Interpretability Paper