Short Definition
Self-attention is attention applied within one sequence so each token can use information from other tokens in the same input.
Intuition
Every token gets a chance to look at the surrounding context and update its representation based on what matters.
Technical Definition
Self-attention forms queries, keys, and values from the same sequence and computes context-aware token representations through weighted value aggregation.
Example
In "the animal did not cross the road because it was tired," self-attention can connect "it" to "animal."
Common Misunderstandings
Self-attention is not automatically bidirectional; causal masks can restrict future tokens.
Self-attention still needs positional information to reason about order.