Types of learning: supervised, unsupervised, self-supervised, RL
The categories differ in what supervision signal the model gets — where the "correct answer" comes from during training.
The four families
- Supervised — learn from
(input, label)pairs. The label is the answer. Classification and regression. Powerful but bottlenecked by labeling cost. - Unsupervised — no labels; find structure. Clustering, dimensionality reduction, density estimation. Answers "what's in this data?"
- Self-supervised — labels are generated from the data itself. Hide part of the input and predict it. No human annotation needed.
- Reinforcement learning (RL) — learn from a reward signal by acting in an environment. No labeled answer, just better/worse outcomes over time.
Why self-supervision changed everything
LLMs are trained self-supervised: the "label" for each token is simply the next token in the text. That turns the entire internet into training data without a single human annotation, which is why pretraining scaled the way it did. The model learns language, facts, and reasoning patterns as a side effect of getting good at "predict the next token."
Self-supervision = unsupervised data, supervised-style training. It is the bridge that made foundation models economically possible.
Where each shows up in modern AI
| Stage | Learning type |
|---|---|
| LLM pretraining | self-supervised (next-token) |
| Instruction tuning / SFT | supervised (prompt → ideal response) |
| Preference alignment (RLHF/DPO) | reinforcement / preference learning |
| Embeddings, clustering | unsupervised / self-supervised |
A frontier model is a stack of these, not one. See fine-tuning & alignment for how the later stages work.
Connects to: LLMs · how learning works · fine-tuning & alignment