Self-supervised learning is training where the data itself provides the target label.
For language models, the most common objective is next-token prediction: given previous tokens, predict the next token.
The model sees text like:
the capital of France is Paris
and training examples are created automatically:
input: the capital of France is
target: Paris
No human has to label every example. The sequence already contains the answer because the next token is known.
Self-supervised learning is what makes foundation-scale pretraining possible. It turns massive raw corpora into training data.
The limitation is that next-token prediction teaches language patterns, not necessarily helpfulness, safety, correctness, or task-following. That is why pretraining is usually followed by SFT, preference tuning, or RL-style post-training.
Related: llm, tokenization, fine tuning.