Mechanistic interpretability Language models are evaluated from the outside.
We study them from the inside:
the computation itself. From that, guarantees about behavior,
and better models.