To simply choose the K most further apart entities as initial centroids is rather dangerous. Real-world data sets tend to have outliers, under your approach these would be chosen as initial centroids.
There are many initialization algorithms for K-Means, perhaps you would like to take a look at intelligent K-Means.