Why should a model that fits a finite set of training examples do well on data it has never seen? The PAC (Probably Approximately Correct) learning framework turns this question into mathematics: with enough samples, the hypothesis we learn is, with high probability, nearly as good as the best one available.