Nevermind the total absence of an attempt to control for quality, or, y'know, that the human goes on breathing after she finishes drawing, there are some real p

Nevermind the total absence of an attempt to control for quality, or, y'know, that the human goes on breathing after she finishes drawing, there are some real problems here 1. the human CO2e takes national, annual averages of CO2e and attributes the fraction of a year it takes to make a drawing to that drawing. they then _add_ to that the operating cost of whatever machine the person is using, which presumably is already a component of the national individual average? 2. The 'ChatGPT' baseline is GPT-3, not -3.5 or -4. 3. Training CO2e is from a cite I can't access, but is based on GPT-3 only. 4. They cite to some guy's post on Medium to estimate ChatGPT inference CO2e!? That admits it doesn't know the number of queries per day, or the compute time used, or the energy per compute (besides the TDP of A100s), or the carbon intensity of that energy??? This is a mess, and doesn't deserve a Nature publication.