the counterpoint is since then we've killed all the easy ways to scale. the datasets can't get bigger because it's already the whole Internet. model sizes can't grow that much because you start running into RAM constraints. efficiency definitely can be improved, but probably not more than 100x on our current architectures.