adjective · AI technique
Zero-shot
also called zero-shot prompting
Asking the AI to do something cold, with no examples. You just describe the task and see what happens.
Prompting a model to perform a task from instructions alone, without any worked examples in the prompt.
The prompt states the task and nothing else: no worked examples. The model relies on patterns learned in training to infer the format and the answer, which works well for common tasks and poorly for unusual ones.
Real-life example
You ask a model to label customer messages as complaint, question or praise. With one clear sentence of instructions it gets most of them right. Then it labels a sarcastic message as praise. Zero examples worked for the easy cases and missed the odd one.
Where the analogy breaks
-
Zero-shot does not mean no instructions. You still have to describe the task clearly, only the examples are missing.
-
It often works better than people expect. For common tasks the model has already seen many similar ones during training.
-
When the format or the edge cases matter, add examples. That is the next step up, called few-shot.