Self-attention
Also: intra-attention
Queries, keys and values are all linear projections of the same sequence, so each position gathers information directly from any other position in one step. Its cost grows with the square of sequence length; a causal mask restricts each token to earlier positions for generation.
Example
In "The animal didn’t cross the street because it was too tired", self-attention lets "it" put most of its weight on "animal". A 1,000-token input needs a 1,000 × 1,000 score matrix per head.