principles.fyi · the brain · concept

zero-shot

Asking the model to do a task with just an instruction and no examples.

instruction + input -> answer (0 examples shown)

Zero-shot means you describe the task in plain language and the model attempts it cold, with zero worked examples in the prompt — for instance just 'Translate this to French:' followed by the sentence. It works because instruction-tuned models have effectively seen the shapes of countless tasks during training, so a clear description is enough to point them at the right behavior. It contrasts with few-shot, where you also paste in a handful of examples.