The neuron is basically triggering on rare or unexpected (low‐frequency) tokens—i.e. it lights up whenever a word or symbol shows up that the model hasn’t seen very often.