Scaling-fit discrepancy: output-head FLOPs are ~65% of compute at the smallest rung, ~3% at 535B
dlwh · x · 2026-10-27
Researcher dlwh highlights a methodological gap between scaling-law datasets: Uncheatable fit fixed E at 1.4 and counted output-head FLOPs, while Paloma excluded them. That term is 65% of compute for the smallest rung versus 3% at 535B — a big deal when extrapolating between scales. He also calls the 1.4 asymptote "made-up nonsense," noting he did the analysis with Codex.
More from Research
- arXiv moderator's tip: accept the classifier's primary category or risk a hold — tdietterich · 2026-10-27
- arXiv has 4,000+ papers on hold amid submission surge; moderator shares how to avoid it — tdietterich · 2026-10-27
- CoreWeave, You.com and NVIDIA lift BrowseComp 22.9% in 3-week RL post-training sprint — RichardSocher · 2026-10-27
- Open-source decision models sold their own base in Red Alert 2 — then learned until no human could beat them — heathercmiller · 2026-10-27
- DeepMind paper: LLMs can't sample randomly on their own — pass random numbers from code instead — rohanpaul_ai · 2026-10-27
- World Model Assigns Weight and Friction to Objects, Then Simulates with a Physics Engine — jonstephens85 · 2026-10-27