Module IX
Talking to the Model
Chapter I
The Context Window
The last module left us with a model whose knowledge is frozen and which keeps no memory between conversations. That sounds like a dead end. It is the opposite. If the model itself cannot change, then everything you actually steer happens in the text you hand it. That text has a name, and a shape, and edges.
A conversation with a language model can feel like talking to someone who is following along. The model is doing nothing of the kind. Each time it replies, it starts cold and reads one block of text from the top: the window. Whatever is in that window is its entire world for that moment. Whatever is not in it does not exist.
So the window is the one lever you hold. Fill it well and you get sharp, useful answers. Let it clog with noise, or let the part that mattered slide off the edge, and the model works with what is left, never knowing anything is missing.
This chapter is about that window: what is in it, why it has a hard edge, and what happens when it overflows.