Not what it's about to say — what it's thinking about while it decides.
Same sentence, different landmark — watch which lights change and which don't.
Tap a layer to freeze it, tap a concept to inspect it
Think of it as a shared noticeboard every part of the model reads from and writes to. Early layers post rough notes, middle layers post concepts, late layers read the board and commit to an answer. The logit lens only ever shows you the final word; this shows you the reasoning that produced it — including the options it held and dropped.
Illustrative concepts. Real feature readouts are pulled with sparse autoencoders and browsed in tools like Neuronpedia.