It doesn't really matter if some of the data is generated by LLMs. It's possible for LLMs to improve themselves by training on their own output, there is no st

It doesn't really matter if some of the data is generated by LLMs. It's possible for LLMs to improve themselves by training on their own output, there is no strict requirement for "fresh" content. If it's mixed with human content and gets human feedback through voting that's great training data.