Learning without changing the weights

The model used a supplied example to handle another case. This is called in-context learning: adapting to a task through the context, without updating the trained weights.

Remove the example from the next request and that particular demonstration is no longer available. The model may still produce the same format because it already knows similar formats from training, or because your written instructions are enough.

Think of the example as a reference sheet. Removing the sheet does not erase the reader's education, but it removes that reference.

In 2020, the GPT-3 researchers tested many tasks by supplying instructions and examples without further weight updates. Results varied by task. Examples can be powerful, but they are not a general replacement for training.