Kiru Lab / Mechanistic Interpretability
Causal Methods
Patching, ablation, and probing — the difference between watching and cutting.
In this module you will
- Run activation patching and read the resulting map
- Choose an ablation baseline that does not create an artifact
- State what a probe does and does not establish
Activation Patching
The workhorse causal method: transplant an activation from one run into another and watch what moves.
Hands-On Lab · 55m
Ablation and Lesion Studies
Remove a component and measure the damage. The oldest method in neuroscience, applied to weights.
Hands-On Lab · 40m
Probing: What a Classifier on Activations Can and Cannot Tell You
Training a probe shows information is present. It does not show the model uses it.
Concept · 35m