[slyfocks]

Apparently it has only trained on 30k songs thus far. This would explain the bias towards pop, as 30k is far too small of a training set to accurately capture the long tail of preferences.