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

(github.com) by rpadhye | Jul 21, 2026 | 0 comments on HN