why do people think that efficiently trainable models will reduce GPU demand?

Comments