Given unlabeled points, group them so similar points end up together and different points end up apart. No y to learn — just structure to find.
Why k-means specifically: a natural way to define "similar" is distance. Assume clusters are round blobs — find a few center points so every point is close to its nearest center.