I have the same worry about LLMs in general - I know that ‘model collapse’ seems to be an unfashionable idea, but when the internet’s just full of garbage (soon?…), what are we going to train these things on?