the logit lens is the cheapest interpretability trick and nobody uses it

you can decode intermediate layers straight to vocab and watch the model change its mind across depth. takes ten lines of code. why is this not the first thing everyone reaches for.