Attention
The core mechanism in transformer models — each token looks at all others and decides how relevant they are to the context.
When the model generates the next word, it asks every previous token to contribute information, weighted by relevance. "Bank" in "I sat by the bank and fished" gets more attention from "fished" than from "sat".
The paper "Attention Is All You Need" (2017) kicked off the entire LLM era.