In-context learning

In-context learning describes an observed ability of large language models: a model can "learn" a new task from the content of the prompt alone — the instructions and any examples — without any change to its parameters. The adaptation happens solely while the answer is being generated and vanishes once the conversation ends. This is why the very same version of a model can handle hundreds of different tasks depending on how you ask.

In practice, in-context learning is the umbrella concept above zero-shot and few-shot — these differ only in how many worked examples you place in the prompt. It is also why prompt engineering is so effective: well-chosen instructions and examples genuinely change the model's behavior, even though they don't change the model itself.

The key distinction concerns fine-tuning. Fine-tuning permanently modifies the model's weights and bakes in the new skill, but it requires data, time and the cost of training. In-context learning works instantly and bakes in nothing — which is why deployments first squeeze as much as they can out of it, and turn to fine-tuning only when the task is repeatable, well defined, and worth writing into the model for good.