How does supervised learning work?
Through labeled history. Each training example is an input plus the answer a person or a system recorded, its ground truth. The model adjusts until its predictions match those labels, then is scored on a held-out set it never saw. Two task types cover most uses: classification predicts a category, approve or escalate, and regression predicts a number, such as next month’s demand. The model can only be as good as its labels, which is why data annotation and label review are often the real budget line.
Supervised vs unsupervised vs self-supervised learning: what is the difference?
The difference is where the answers come from.
| Supervised | Unsupervised | Self-supervised | |
|---|---|---|---|
| Labels come from | People or recorded outcomes | Nowhere | The data itself, by hiding part of it |
| Goal | Predict a known field | Find structure | Learn general representations |
| Typical task | Classification, regression | Clustering, anomaly detection | Next-token prediction |
| Business example | Churn score per customer | Customer segments | Pretraining a language model |
| Main cost | Labeling | Interpreting the results | Data and compute at scale |
Unsupervised learning has no answer key at all. Self-supervised learning manufactures its own labels, so it trains like supervised learning on data nobody annotated.
Are LLMs supervised or unsupervised?
Both, in stages. Pretraining is self-supervised: the model predicts the next token across vast amounts of text, and the text supplies every label. The step that makes a model follow instructions is supervised. The InstructGPT paper describes collecting “a dataset of labeler demonstrations of the desired model behavior” and fine-tuning GPT-3 on it with supervised learning, before reinforcement learning from human feedback (Ouyang et al., arXiv:2203.02155, 4 March 2022, read 5 November 2026). That stage is now called supervised fine-tuning, the first part of post-training. When a company adapts a model on its own examples, that fine-tuning is supervised learning too.
Where does supervised learning pay off in a business?
On repeated decisions whose outcomes are already recorded. Fraud flags learn from confirmed chargebacks, churn models from customers who left, demand forecasts from past sales, document routing from tickets people already sorted. The labels are often already sitting in your system of record, which makes a first model cheaper than it looks. If nobody records the outcome, there is nothing to supervise. A language model can also label examples zero-shot to bootstrap a dataset, after which a small trained classifier handles the volume at a lower cost per prediction.