Zero knowledge verification for frontier AI training is possible

In three linesResearchers propose a zero-knowledge VM architecture to verify actual GPU floating-point computation during frontier AI training without disclosing model architecture. Protocol combines pre-committed training specs, inter-node network observations, and on-the-fly Merkle commitments. Estimated single-digit-percent overhead; deployable proof-of-concept in ~36 months.

Summary generated by Claude — human-verified