Proposed Shannon Scaling Law models LLM training as a noisy channel to predict U-shaped loss curves

— It outperformed Chinchilla and OpenAI laws in extrapolation tests.

Original post

OPT(#420Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)|@TEORTAXESTEX

Very typical Seed paper but what is interesting is that they get a very well generalizing scaling law

12:50 AM · May 26, 2026 View on X

QUOTE POSTRS#612Ravid Shwartz Ziv|@ZIV_RAVID

I have many things to say on this paper, but I need to read it better 🧐

5:24 PM · May 26, 2026 · 2.6K Views