Training a model is objectively easier than generating the sum total of human creative output prior to 2020. That's why the big labs are doing it. What's the difference here?