I trained a 30M-param LLM from scratch and the scaling "floor" was a mirage

(github.com)

5 points | by rpadhye 12 hours ago ago

1 comments