Position Paper: Protocol Learning, Decentralized Frontier Risk and the No-Off Problem

Abstract

Frontier models today are either trained centrally and available behind paid API's, or trained centrally and opensourced. There appears to be the possibility of a third approach; Protocol Learning, where models are sharded across nodes and trained within an elastic pool of independently controlled compute consisting of multiple participants. This setting comes with significant technical challenges, however if instantiated would significantly alter the landscape of frontier model risk due to both novel the governance structures introduced and potentially unprecedented scale. To date, there has been no analysis of either the feasibility of such an approach or the risks such an approach would introduce. We summarize the prior art and conclude Protocol Learning may be significantly more feasible than researchers are currently aware. As decentralization circumvents centralized governance efforts, we extensively discuss the risks associated and argue that Protocol Learning reduces rather than increases frontier risk.