Self-supervised learning creates training targets from the structure of unlabeled data, allowing models to learn representations at enormous scale without manual labels for every example. This guide explains the mechanism, trade-offs, evaluation, and controls that matter in practice.