For anyone who has used a chatbot and wondered how it knows what “it” means. The weights are invented to show the shape of the thing, not measured. The colour washes stand for each word’s value (a vector, not the word), and the mix is added to “it”, not swapped for it; the word keeps what it had. The arc from “it” to itself is left out for legibility. This is one head of many: in a real model the work is spread over many heads and layers, and none of them is the pronoun-resolver: the best single coreference head Clark et al. (2019) found in BERT picks the right antecedent 65% of the time, losing to a rule-based sieve. The arc running forward from it to hungry is legal only because this is encoder-style attention, reading in both directions; a causal decoder masks it. And real attention maps put a large share of every row on a delimiter or first token as a no-op, which would swamp a drawing this size.