How do you verify an untrusted GPU host actually ran the model? Gonka's design notes on 3 cheating vectors
autoimago · reddit · 2026-09-17
A contributor to open-source project Gonka published design notes on verifying that untrusted GPU hosts in a decentralized inference network actually ran the paid-for open-weight models (DeepSeek V4-Flash, GLM-5.3-Flash, MiniMax M2.7), when re-running every request would double network costs.
Three cheating vectors and defenses:
- Swapping in smaller models/cheaper quants → random 1–10% re-run by other hosts; open question: can subtle quants hide in floating-point noise?
- Claiming more compute than you have → 'Proof of Work 2.0' sprint rounds measure real compute; open question: how well do sprints track serving capacity?
- Occasional cheating → reputation scoring raising check rates; open question: is expected-loss deterrence enough for high-value requests?
Problems posed to the community: what output tolerance can't be gamed across GPUs/kernels; whether checks should be cost-weighted given MoE routing and 400K-token prompts; and whether sampling suffices or some requests need TEEs/ZK proofs. Code is public (inference-chain/), with HackerOne bounties for security breaks.
More from Infra
- huggingface_hub 1.32 lets UV scripts declare runtime images for HF Jobs — vanstriendaniel · 2026-09-17
- Open-sourced Qwen-1B-RLCD runs type-safe JSON inference 5x faster on-device — JiliJeanlouis · 2026-09-17
- Fathom Speeds Million-Token KV Scans 1.67x with Per-Query Read Depth — Vivek Kalyanarangan · 2026-09-17
- Edge0 Streams a 35B MoE from SSD at 20 tok/s on a Single 24GB GPU, Open Source — Edge0 · 2026-09-17
- Daytona Benchmarks NVIDIA Vera CPU on Agentic Workloads, Cited in Official Blog — mattturck · 2026-09-17
- CXMT's LPDDR5X lands in flagship phone as Nubia ships $885 Doubao AI handset — pstAsiatech · 2026-09-17