The fun part is comes with future training via scraping the internet. At this point, the vast majority of new articles are LLM-infused at minimum. I'm sure smart AI scientists are trying to curate the training data, but I have no idea how they're avoiding incestuously training LLMs on themselves. Otherwise the future it seems like we'll get is a bland, grey Hapsburg training set (read: internet) and even worse LLMs