Few-shot learning is the ability to perform a new task after seeing only a few examples of it, rather than the thousands typically needed to train a model from scratch. The idea entered the mainstream with the 2020 paper “Language Models are Few-Shot Learners,” which introduced GPT-3. The paper showed that a sufficiently large language model can do a task when the task and a few demonstrations are written directly into the prompt, “without any gradient updates or fine-tuning.” The model is not retrained; it simply conditions on the examples it is shown and produces the right kind of answer.
The terms attached to this vary by how many examples are given: zero-shot means a plain instruction with no examples, one-shot means a single example, and few-shot means a small handful. This in-context behavior was a surprising emergent property of scale, and it is why prompting - rather than building a custom model - became the default way to get useful work out of large language models.
Why business readers should care: Few-shot learning is the reason teams can adapt a general AI model to a specific task by writing a prompt with a few examples, instead of running an expensive training project. It dramatically lowers the cost and time of putting AI to work on a new problem.