Few-Shot vs. Zero-Shot
Same model, same question. Paste two examples and the answer flips. Here's the ladder.
Grounded in the research on few-shot learning
1. Start zero-shot
Zero-shot is you just describing the task. No examples, only an instruction: "Classify this review as positive or negative." The model has to infer the format and intent from words alone. It's the cleanest starting point. The term blew up with Brown et al.'s 2020 GPT-3 paper, literally titled "Language Models are Few-Shot Learners." Always try zero-shot first. If it nails it, you're done and you saved tokens.
2. Add one example
One-shot is the next rung: paste exactly one worked example before the real input. "Review: great food -> positive. Now: terrible service -> ?" That single demo pins the output format and shows what "done" looks like. GPT-3's authors tested this rung separately because one example often jumps accuracy hard versus zero. If zero-shot gets the idea but botches the formatting, one shot usually fixes it for nearly free.