The neuron appears to light up on the single most “unexpected” or low‐predictability token in each snippet—i.e. it flags whichever word in the local context is least anticipated.