Hacker News new | ask | show | jobs
by rocho 569 days ago
That's correct. I saw a paper recently that showed how LLMs performance collapses when they are trained on synthetic data.