The Chinchilla paper only addresses the contrived use case of a model that is trained once and never used for inference. Since most of the real world compute co

The Chinchilla paper only addresses the contrived use case of a model that is trained once and never used for inference. Since most of the real world compute cost will be in inference, Chinchilla seems to offer little practical guidance.