A System Overview for Near-Term, Low-Trust AI Compute Verification

This is a working paper with two primary goals: A) present opinionated recommendations for what the treaty-oriented field of AI compute verification should focus on. B) probe the research community on points of consensus vs. disagreement on objectives, threat models, and the proposed technical approaches.

‍

Abstract

This piece is aimed at researchers and institutions working on technical solutions for AI compute verification and assurance between rival nation states—including slowdown agreements or other forms of mutual restraint. The paper proposes a verification system fitting the criteria of being quickly retrofittable to existing data centers. The first part of the paper discusses prioritization: threat models, verification objectives, technical requirements and general design choices. Most of the paper proposes specific design details for such a system, with the goal of establishing consensus (or surfacing contention) in the field.  The proposed system involves network taps and recomputation to confirm that workloads were run as declared. This recomputation happens via unilaterally trusted compute provided by both the prover and verifier, with the results cross-checked and shared in a controlled and confidential manner. 

‍