The Component That Actually Makes a 'Faster' Optimizer Faster

The optimizer wars have a new reading instrument. When a paper claims a new training method beats the old one, the conventional move is to credit the most exotic-sounding step and move on. Wang et al.'s ablation of Muon on a small modular-arithmetic task inverts that habit. They isolate each component and find that the part everyone credited, spectral scaling, is a passenger. The driver is Newton–Schulz orthogonalization, a math step that keeps parameter updates from interfering with each other.

The mechanism repeats across any training run where one component does the work and another rides along. Strip the driver, and the speedup evaporates. Spectral-only matches AdamW. Orthogonalize-only matches full Muon. The "faster" label was never about speed of a single step; it was about which transformation actually shaped the solution.

Wang et al. report both first-crossing time and sustained-grokking time. Under the first metric, one Newton–Schulz iteration looks faster. Under the second, it is fragile, prone to transient collapse, with that fragility growing as the learning rate rises. The canonical five iterations are the learning-rate-robust choice. The headline inverts depending on which clock you keep.

The reader's tool is the dual-clock read. A single NS iteration can look faster on first-crossing time but collapse on sustained-grok; the canonical five iterations are the learning-rate-robust choice. Any single-iteration "faster optimizer" claim deserves a second clock.