According to StanfordAILab, self play pretraining from random init shows predictable scaling on text, image, audio, and melodies.