

LLM discoveries are everywhere. You find them by poking the model and trying stuff. My tweet below is the first prior example cited in @repligate
’s “Mysteries of mode collapse” on LW — I had no idea what it was; I just thought it was weird. https://
lesswrong.com/posts/t9svvNPN
mFf5Qa3TA/mysteries-of-mode-collapse
…

