Not only are ML techniques finding ways to do physics with less energy, the amount of waste by non-ML computing absolutely swamps ML computing. I don't care if training a model uses the amount of CO2 that a translatlanic flight uses, because those flights are flying constantly, some of them not filled to the max. The whole criticism about energy in the paper seemed to miss major details that I'd expect top researchers to get right.

BTW, I ran a large scale computing platform at Google for several years. It made absolutely enormous, stunningly huge amounts of CPU available to scientists and we used it to make several important breakthroughs. However, after a few years, I finally was shown what the power cost implications were, and I shut down the project because the costs were not justified by the scientific benefits. Compared to what I was doing, ML is relatively skimpy on CPU.