Much has been said about many companies’ desire for more compute (as well as data) to train larger foundation models. I think it’s under-appreciated that we have nowhere near enough compute available for inference on foundation models as well. Years ago, when I was leading teams

→ View original post on X — @andrewyng,