This approach is more likely to lead to a bad clustering result than the approach used by the demo. At the other extreme, a variation called k-means++ uses a moderately complex probabilistic algorithm ...