The operation that lets a model weigh every token against every other to decide what matters in context.
Broad technical consensus on what this term means. Disagreements that invoke it are about the world, not the word.
Attention lets each position in a sequence look at every other position and decide how much each one matters for interpreting it, resolving what "it" refers to, connecting a question to the relevant clause three paragraphs back. The model learns these weighting patterns; nobody specifies them.
Attention is the core operation of the transformer, and its cost grows with how much text is being related to how much text, the fundamental reason context windows are finite and long ones are expensive.