Skip to main content

Context Is Not Training

This video presents the same text shown beside it, spoken and on screen. It adds nothing the text does not say.

State

During a conversation, a language model's knobs do not move — what changes is the context, the words attention reads.

Show

A correction shapes replies for the rest of the session because it joined the passage — and vanishes when the session ends.

Watch for

Cross-session memory, where real, is engineering around the model — stored notes fed back into context, not learning inside it.