the problem to the scaling thing is that we don't nearly have as many samples per artist as we do have "text pointed to by links on reddit", as used in GPT2 (openAi's text generation thing). I would expect that they can't go as crazy as they could