I trained a 30M-param LLM from scratch and the scaling "floor" was a mirage(github.com)
3 pointsby rpadhyean hour ago1 comment