In-context learning

Few-shot prompting feels like teaching. You show the model a few examples and it suddenly does the thing. But it isn't learning in the way the word implies, and understanding the difference makes you much better at steering it.

When you put examples in a prompt, the model does not update its weights. Nothing about it changes. What the examples do is activate latent pathways that were already there. The model already knew how to do the thing; your examples redirect it toward that existing capability instead of some other one. You're not writing to its memory. You're pointing at a part of what it already contains.

So why does showing examples work so well?

Because the model's behavior is enormously sensitive to what's in front of it, and examples are the sharpest way to say "this specific flavor, not the thousand others you could produce." Two or three good examples of the exact output you want will beat a paragraph of description, because the examples are the target rendered concretely instead of described abstractly.

The way I understand it is training the new guys at the mill. You don't rewrite anyone's brain. You put them in a few situations where the right move is obvious, and they pattern off it. The capability was there. The examples just made which capability you wanted unmistakable.

The session-length version of the same thing

This is also why long working sessions produce better output than cold one-shot prompts. Every message you add sculpts the probability landscape a little more toward what you're actually after. The context accumulates into a sharper and sharper picture of the target. A model twenty messages deep into a real working session with you is not smarter than it was at message one, but it is far better aimed.

The takeaway: examples and context don't teach the model new things, they aim it at what it already knows. So show the output you want, and build context deliberately, because aiming is most of the game.