Zero-shot learning is when a model performs a task correctly without having been given any examples of that specific task - it works from a plain instruction alone. For a large language model, that means writing a request such as “translate this sentence into French” and getting a good answer even though no translation examples were supplied in the prompt. The term contrasts with few-shot learning, where a handful of examples are shown, and one-shot learning, where a single example is shown.
A landmark result is the 2021 FLAN paper, “Finetuned Language Models Are Zero-Shot Learners” by Jason Wei, Quoc Le, and colleagues at Google. They showed that instruction tuning - fine-tuning a model on many tasks phrased as natural-language instructions - sharply improves its ability to handle unseen task types, with the resulting model beating GPT-3’s zero-shot performance on 20 of 25 benchmarks. This is why modern assistants can follow a wide range of instructions out of the box rather than needing a custom-trained model for each task.
Why business readers should care: Zero-shot capability is what lets a single general AI model take on a new task immediately from a written instruction, with no examples and no retraining. It is a large part of why deploying AI now often means writing a prompt rather than building a model.