I trained a 30M-param LLM from scratch and the scaling "floor" was a mirage(github.com)5 points by rpadhye 8 hours ago | 0 comments