Team Idea
Verifiable LLM inference on untrusted, heterogeneous hardware. Nodes execute jobs and sign Ed25519 receipts; any verifier re-runs a single forward pass to audit what the node claims it computed — we measured that audit at 2.4-9.9% of the job's own cost, cheap enough to pay for out of the job price. The open question is whether an honest AMD node and an honest NVIDIA node compute the same thing: if not, naive fingerprint comparison punishes honest participants. That ROCm measurement is what we came here to run. Rust, libp2p, MIT. Proven end-to-end on Linux/CUDA and a 2012 Intel Mac (x86_64), including cross-platform receipt verification.
Team Members
Karol Szubert
Szpilman
Rust AI infra developer
Aleksander Cichucki
rejonmosciska15938
Filip Zarzyński
0x414141
Security Researcher
.png&w=256&q=75)
.png&w=256&q=75)
