logoalt Hacker News

MagicMoonlighttoday at 11:45 AM0 repliesview on HN

It is trained on its own slop. They haven’t trained these models on books for three years at this point. Only on generated slop. (And RL slop upvotes/downvotes from users)