Few-shot learning is a way of using a large model in which you show it only a handful of examples and it completes a new task following their format and standard. There is no retraining involved. You simply include two or three pairs of input and desired output in your prompt, the model infers the pattern behind them, and then applies the same treatment to new material. It works because the model already learned language and task understanding during pretraining; the examples act less like a lecture and more like an on-the-spot demonstration.
Zero-shot, one-shot and few-shot
The three terms differ only in how many examples you give. Zero-shot means stating the requirement with no example at all, so the model answers purely from its understanding of the task, which suits familiar jobs such as translation or summarising. One-shot provides a single complete demonstration so the model can see what the output should look like. Few-shot usually means two to five examples, and the more unusual the task or the stricter the format, the more they pay off. The rule of thumb is simple: when a first answer drifts off target, the problem is usually not ability but that the model does not know the shape you want, and adding examples fixes that faster than rewording rules.
Why examples beat explanations
People often write long lists of rules into a prompt and still get off-target answers, because rules are abstract while examples are concrete. For the same classification task, ten sentences of criteria teach less than three already-labelled samples, since the examples simultaneously fix the input format, the output length, the tone and the edge cases, giving the model something to check item by item. Quality matters more than quantity here: if the examples contradict each other or only cover easy cases, the model will learn the wrong pattern just as faithfully. Pick examples that cover the most confusable situations and keep their output format perfectly consistent.
It is not fine-tuning
Few-shot learning is often mistaken for a form of training, yet it never changes the model's parameters. The examples live only inside that one conversation's context, and once the conversation ends or the context is cleared, the demonstration effect is gone. Fine-tuning is different: it genuinely updates model weights with a dataset, so the effect persists, but it demands data preparation, compute and evaluation. The trade-off is clear. For temporary tasks, shifting rules or a quick trial, use few-shot. For a fixed, long-term, high-volume task where stability matters, fine-tuning earns its cost. One more misconception deserves correcting: more examples are not always better. Examples consume context, and piling in too many squeezes the material you actually need processed. Three to five good examples are usually enough.
