This neuron lights up on the single most “surprising” or content-heavy word in each chunk—i.e. rare or highly informative tokens (names, unusual terms, key verbs/prepositions).