Train once, deploy many: AI and increasing returns

Abstract

AI’s “train-once-deploy-many” advantage yields increasing returns: doubling compute more than doubles output by increasing models’ inference efficiency and enabling more deployed inference instances.