K-means is optimal for some problems, like in bolt quantization (and hierarchical initializations tend to converge quickly). It's not really applicable to TFA,

K-means is optimal for some problems, like in bolt quantization (and hierarchical initializations tend to converge quickly). It's not really applicable to TFA, since you'd choose the cluster count via metrics which have nothing to do with elbow criteria or whatever, but the lesson isn't to not use k-means or to use some other clustering algorithm; it's to know your goals and which model best achieves those.