Decentralized training works? REAX doubles baselines in first model competitions
airesearch12 · x · 2026-10-17
REAX reports its decentralized training network's first-week results: 217 miner slots now serve decision models, 65 independent miner wallets are earning, and nearly all routed traffic is answered by miners' GPUs with zero failed verification checks. The first training competition produced two champions — demand forecasting (UID 71) scored 91.1 vs a 38.2 baseline, and risk assessment (UID 251) scored 64.2 vs 31.7 — both more than double the starting performance, verified via confirmation draws, head-to-head playoffs, and copy/twin checks. The author argues this proves decentralized training competitions for use-case-specific decision models actually work.
More from Research
- NVIDIA's open midtraining recipe lifts SWE-bench Multilingual from 55.8 to 62.8 with 500B tokens — dair_ai · 2026-10-17
- Audit of 160,860 citations: nearly 40% of links shown by Chinese AI search go unused — Sanbi_Ai · 2026-10-17
- Dogs show the Bouba–Kiki effect, study finds — abenitezburraco · 2026-10-17
- rasbt's Reasoning From Scratch Finale: Distilling Reasoning Models Hands-On — rasbt · 2026-10-17
- Terence Tao: OpenAI's quasi-Riemann result is the field's biggest advance in 50 years — basedjensen · 2026-10-17
- After math and physics, AI finds explicit counterexample to 1967 fusion conjecture — mawcopolow · 2026-10-17