> it seems that lots of ML stuff is just plain inefficient This is just like any other technology. Use it wrong, you will get burned. Doesn't matter how shiny

> it seems that lots of ML stuff is just plain inefficient This is just like any other technology. Use it wrong, you will get burned. Doesn't matter how shiny the container is. "ML stuff" also (mostly?) includes purely statistical methods from the 90s that are deterministic and arguably the best way to solve a large variety of non-generative problems. In fact, unless generation of arbitrary output is a major objective, it's likely you can solve whatever ML task on a workstation from 2010 that uses intel integrated graphics.