Not Everything in Mech Interp Has to Be Causal — Sida Liu

A recurring criticism in mechanistic interpretability (mech interp) discussions is that a finding is "not causal." A study might identify an interesting pattern in a model's internal representations, only to face the objection that it has not demonstrated a causal effect on the model's behavior.

saved by

• Ishaan Panigrahi