Unsupervised learning encompasses a family of techniques. Clustering algorithms like K-means group similar data points together. Autoencoders learn compressed representations by encoding data to a small bottleneck and then reconstructing it. Dimensionality reduction (PCA, t-SNE, UMAP) projects high-dimensional data into 2D or 3D for visualization. What unites them is the absence of labels — the model defines its own notion of "similar" or "important" based on the data's statistical structure.
LLM pre-training is often called "self-supervised" rather than truly unsupervised, because the training signal comes from the data itself (predict the next token). But the spirit is unsupervised — no human annotator labels each token. The model discovers language structure, factual knowledge, reasoning patterns, and even some world knowledge purely from the statistical patterns in text. This is why pre-training requires such massive datasets: without labels to guide it, the model needs enormous amounts of data to discover meaningful patterns on its own.