The ability of a model to learn from zero, one, or a few examples, respectively. In practice this "learning" ability happens at deployment time -- after the model has actually learned (i.e. been trained). Models generally need to be trained in a particular setup for {zero, one, few}-shot learning deployment, but can sometimes display such capability as an "emergent" property. Examples: [zero-shot learning] a model learns to classify an animal as a zebra without having seen any, by knowing the concepts of horse and stripes; [one-shot learning] a model learns to tell if two pictures are from the same person even if it hasn't seen that person, by learning a measure of similarity between pictures; [few-shot learning] a model can respond to a new query after seeing a few examples of query-response pairs, by learning with high accuracy how to predict the next word in a sentence (this is an emergent property of large language models).