Am I missing something about training time? Does adding per token noise cause training to slow significantly? Cool paper though!