Place K points randomly in your data space. These are the initial *centroids* — the centers of your clusters. Think of them as flags dropped blindly onto a map.