How a decentralized inference network catches cheating GPU hosts: spot checks, reputation, open weights
autoimago · reddit · 2026-09-17
A Gonka contributor explains how the open-source decentralized inference network engineers trust when GPUs belong to strangers: (1) random 1–10% of tasks are re-run by other hosts as spot checks, so cheating on any request is a gamble; (2) a reputation system lowers check frequency for consistently passing hosts and raises it (with reward cuts) for failures; (3) the served models (DeepSeek V4-Flash, GLM-5.3-Flash, MiniMax M2.7) are open-weight, so anyone can compare endpoint outputs against the public model — impossible with closed models. The unsolved part: different GPUs don't produce bit-identical outputs, so matching requires a tolerance a clever cheat could hide within, which remains an open research question. The takeaway: decentralized AI swaps "trust the company" for probability, reputation and open weights. (Disclosure: the author is a project contributor, so treat as promotional but substantive.)
More from Infra
- GlobalFoundries and Marvell expand Vermont SiGe capacity for AI optical networking — zephyr_z9 · 2026-09-17
- $26,100 desktop AI datacenter: dual RTX PRO 6000 Blackwell workstation goes open source — dee_hw · 2026-09-17
- Huawei Rumored Scale-Up Node with 4,096 Accelerators Could Pack 384TB of HBM — zephyr_z9 · 2026-09-17
- Brad Gerstner: 43GW AI compute forecast too aggressive, sees ~25GW with half for OpenAI/Anthropic — firstadopter · 2026-09-17
- Cohere's CUDA Megakernel Serving Hits 292 tok/s at Batch 1 on a 30B Model — dl_weekly · 2026-09-17
- Stanford's Ousterhout: AI workloads are outgrowing TCP — AI Engineer · 2026-09-17