Verifiable AI Inference
Open-source zero-knowledge prover for LLM inference, making international AI agreements verifiable. At the Singapore AI Safety Hub (SASH).
Problem. International AI agreements, such as limits on what compute is used for, are only as strong as our ability to check them. Verifying what a datacenter actually computed today requires access to model weights and other secrets that no party is willing to share.
Approach. We build an open-source prover that produces zero-knowledge proofs of LLM inference, from the prompt through every layer to the generated tokens. This lets a verifier check that approved inference was run without seeing the model.
Results. Our first measurements prove inference of a quantised Llama-3.1 8B (4,096-token prompt, 512-token answer) in roughly 1–1.5 hours on a single H100, with verification taking about two minutes.
Read more: Towards Proving Frontier-Scale LLM Inference and The Math Behind Pacing the Frontier.