Attention lets every token decide which other tokens to read from. A token’s query is compared with every key by a dot product; softmax turns the scores into weights that add up to 1; the output is the weighted average of the values. Pronouns find their nouns this way, and words find their descriptions.