Naomi Clark (@metasynthie) 暗悪直美 @halfrobot.com · Apr 19 The state of understanding how these machines work—the ones that people make wild epoch-defining claims about?!—is distressingly thin. This guy invented Task Manager and has 30+ years of programming. He’s not a ML person but neither am I, and I know how token probability sampling works at temp>0!
27 likes 2 replies
Replies
am I reading that post right, that he's explaining "LLMs add randomness to their output instead of always giving the same answer to the same question" as a _defense_ of LLMs??
Sad thing is, it could be EASY to grasp via UI design — there ARE inference UIs that show off how stochastic probabilities work in LLMs, token by token. You can see that a LLM works on “40% chance the next word is blue, 60% red” or whatever. Instead, owners want to pretend it’s a magic thinkbox