Implications of large language model diffusion for AI governance

Abstract

This article studies the diffusion of large language models, focusing on the effects of this diffusion on AI governance. The author argues that the diffusion of large language models is driven by compute, data and algorithmic insights, and that limiting diffusion is most important for compute, but also for data and algorithmic insights. The article assesses the importance of each input in relation to how difficult they are to create independently and how easy they are to share with others. The article also argues that the increasing cost of developing and training large language models will make future models more difficult to replicate. The author further suggests a portfolio approach to governing the diffusion of large language models, which involves prioritizing compute but also working to limit access to datasets and algorithmic insights. The author concludes by recommending that top developers reduce the publication of important algorithmic insights and strengthen information security and operations security to prevent theft and leakage. – AI-generated abstract