Thanks for sharing the gif, it really helps!

Follow up question to this; we wouldn't be doing this with each and every single point in our training data right since that would not help reduce number of comparisons at inference.

We will need to select some initial points to be the centroids and build the partitions around them, isn't it? If yes, how do we select those initial points, if no, could you please correct me where I am thinking wrongly?