05 · Getting Models to Act
ICL
In-Context Learning
The phenomenon behind few-shot: the model “learns” a task on the fly from prompt examples, without any weight updates. Not learning in the training sense — pattern-matching from context — but it behaves remarkably like it.
Concrete example
Give it a few examples of a made-up coding style and it can follow that style while those examples remain in context. Whether a later session recalls the preference depends on the product's memory and rules features.