Robust Feature-Level Adversaries Are Interpretability Tools

AI Safety Fundamentals - Robust Feature-Level Adversaries Are Interpretability Tools