Concept · Building with AI

Attention budget

Also called: attention limit, limited attention, finite attention

A way of picturing attention as a fixed amount each token can spread across everything else in the input. Adding more text does not add more attention; it divides the same amount more thinly. It explains why giving a model less, better material often beats giving it everything.

For example: A founder pastes in 40 documents to answer one question about refunds, and the one relevant policy gets only a sliver of the model's attention.

Connects to 7

Page last changed · Suggest a correction