Decentralized RL on Bittensor: 100 steps across 12+ GPU models with corrected sampling noise
const_reborn · x · 2026-11-01
RomAI ran 100 steps of fully decentralized RL on the Reliquary model on a Bittensor subnet, showing an up-and-to-the-right chart. Key techniques: rollouts come from untrusted miners on 12+ GPU models — which usually breaks RL — fixed by publicly correcting the sampling noise so the trainer knows exactly what the model would have sampled. Every eval sample is verified on-subnet.
More from Infra
- Reddit debate: cloud LLMs change rules on a whim — only local models are reliable — Substantial_Swan_144 · 2026-11-01
- DeepSeek v4.1 on 4x H200: 2,400 tps, $120k for 16 concurrent users — teortaxesTex · 2026-11-01
- Building a Kubernetes for agents: agentic loop as the new job controller loop — rakyll · 2026-11-01
- Local LLM newbies puzzled: why build a custom harness instead of using Unsloth? — Marino4K · 2026-11-01
- Grafana's Cloud Knowledge Graph now accepts Cypher queries — JeremyCMorgan · 2026-11-01
- NASA JPL stress-tests CXMT memory down to -100°C, unit stays online — jwt0625 · 2026-11-01