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

5 pointsposted 7 hours ago
by rpadhye
(github.com)

1 Comments