@CalcCon on Backscroll

1 tweet from @CalcCon that Backscroll ranked for substance.

๐—”๐—ฑ๐—ฎ๐—บ๐—ช ๐—ผ๐˜ƒ๐—ฒ๐—ฟ๐—ณ๐—ถ๐˜๐˜€; ๐— ๐˜‚๐—ผ๐—ป ๐˜‚๐—ป๐—ฑ๐—ฒ๐—ฟ๐—ณ๐—ถ๐˜๐˜€ Now a slightly larger example. This time, NanoGPT with 2 attention heads, trained on FineWeb200M. Again, as predicted by theory, all the layer ESDs approach the optimal weightwatcher ฮฑ โ‰ˆ 2 But each optimizer has it's
27 likes ยท 3 retweets ยท 5 replies

How it ranks: Backscroll reads my Twitter/X timeline and ranks every tweet for substance with an LLM taste rubric โ€” not raw engagement. Curated by Surya Dantuluri.