Comment by mike_hearn

It's all got much more complex than that in recent years. Training now involves large amounts of inference for RL rollouts and similar. You can't disentangle them computationally like that. "Inference" is just the word used to mean serving customer traffic now, and "training" means creating the model you serve.