The actual emissions for training a language model of the scale examined by Strubell et al. is 3261X smaller than this flawed estimate even if you consider the "P100 in average US data center on average electrical grid" scenario they were evaluating
…

→ View original post on X — @jeffdean