Comment by hxypqr

Predicting the next character alone cannot achieve this kind of compression, because the probability distribution obtained from the training results is related to the corpus, and multi-scale compression and alignment cannot be fully learned by the backpropagation of this model