Decentralized RL Works: Untrusted Miners Lift LiveCodeBench 39.5→58.7 on Bittensor Subnet
bittingthembits · x · 2026-11-05
Reliquary AI's Bittensor subnet (SN81) demonstrated that a real reinforcement learning loop can be fed by mutually untrusted miners: miners ran 100 training steps across 12+ GPU models, improving all 8 benchmarks.
- LiveCodeBench rose from 39.5 to 58.7
- AIME25 rose from 75.8 to 87.9
- Sampling noise was publicly fixed so untrusted miners can feed a genuine RL loop, with every eval sample verified on-subnet
The argument: RL-based improvement is the expensive part of AI and until now only a handful of companies could afford it — if strangers can contribute compute to a real RL loop and results are verifiable, decentralized training becomes practical.
More from Infra
- LexiSuite: an all-Rust local LLM stack with training, LoRA and trimming that claims to beat llama.cpp in spots — W61k3r · 2026-11-05
- Gary Marcus predicts the data center binge will collapse, with many bankruptcies ahead — GaryMarcus · 2026-11-05
- Merge's eval claims its MCP connectors beat Microsoft's native servers on tool count and reliability — shensi · 2026-11-05
- TorchTPU compat study: 4,821 of 4,917 community models run crash-free on TPUs — jiqizhixin · 2026-11-05
- Foxconn targets 50% of global AI server market, up from ~40% today — Beth_Kindig · 2026-11-05
- antirez: prefetching MoE experts via previous-layer router lets K3.1 stream on a 128GB box — antirez · 2026-11-05