The Muon optimizer and the nanoGPT speedrun are the real '1% cost' existence proof
Keller Jordan's Muon optimizer set a nanoGPT training-speed record on 2024-10-15 (+35% over AdamW) and has held through 12+ community records by 7 researchers; Moonshot adopted it for Kimi K2/K2-Thinking, and GLM uses it too. The compounding community speedups (plus data-quality stripping and MoE/attention tricks) are the verifiable mechanism behind the episode's '1% of the cost' claim — the efficiency overhang is real, not magic.
Why it matters (spine — moat). The scarce asset as intelligence goes free. Track the open efficiency stack (Muon-class optimizers, data curation, quantization) as the true driver of falling training cost; assume no proprietary architecture moat behind frontier labs.
What would prove it wrong (falsifier). A frontier lab demonstrates a non-transformer architecture delivering gains the open efficiency stack cannot reproduce.
Live-status check + triangulation. Keller Jordan's Muon optimizer set a nanoGPT training-speed record on 2024-10-15 (+35% over AdamW) and has held through 12+ community records by 7 researchers; Moonshot adopted it for Kimi K2/K2-Thinking, and GLM uses it too. The compounding community speedups (plus data-quality stripping and MoE/attention tricks) are the verifiable mechanism behind the episode's '1% of the cost' claim — the efficiency overhang is real, not magic. Source: Primary · Keller Jordan (2024-10-15); see source link. Every post-cutoff figure was tool-fetched against a live source during the Kimi K3 sweep (2026-07-20); see sweeps/2026-07-20-VERIFIED_DIGEST.md for the full provenance and any corrections of record.
Track the open efficiency stack (Muon-class optimizers, data curation, quantization) as the true driver of falling training cost; assume no proprietary architecture moat behind frontier labs.
domains: machine-layer · open-weights · moat