Model runs an ML experiment unprompted, complete with optimizer scaling analysis

teortaxesTex · x · 2026-11-16

teortaxesTex marvels that you can now ask a question and a model will go off and run a machine learning experiment on its own.

The screenshot shows genuinely substantive analysis: comparing SGD, Adam, and Muon on small power-law feature problems, finding optimizer gains live only in the transient — Muon led Adam by 1.4-2.8x early on an ill-conditioned problem (condition number 4,096) before noise-limiting erased the lead — and estimating 2-4x headroom above AdamW, of which the Muon family has claimed 1.3-2x.

Original post →

More from Models

Models channel →