Definition
Plain language
Showing a model a handful of examples in the request before asking it to do the real task.
As stated in the literature
In-context learning conditioned on a small number of demonstration examples placed in the prompt; contrasted with zero-shot (no examples) and with fine-tuning (weight updates).
Also called: few shot, few-shot prompting
Why it matters: It matters because a few well-chosen examples can steer a model's behavior without any retraining.
For example, you might paste three sample question-and-answer pairs into a prompt so the model copies that format for the fourth.
Heard on the show
“That horizon depends on a baseline error rate, and that baseline is sensitive to prompt format and few-shot conditioning.”Episode 108 — The Reasoning Cliff: Why Thinking Longer Makes Models Worse at Exact Step-by-Step Tasks