I don't think this is correct. I think that in essence what we've learned over the past 100 years is that we need to be more conscious of the totality of the co

I don't think this is correct. I think that in essence what we've learned over the past 100 years is that we need to be more conscious of the totality of the cost benefit analysis, because what happens is that a lot of industry can can make private profits by exploiting public goods. So it's actually encouraging that we're starting to more objectively look at different activities and "Ok, here are the private profits of this industry and here are the public costs". Energy is a really simple element of that, and in AI it's probably not the most important, but it is very easy to measure. It's very easy to say "If you want to do this AI stuff, it's going to mean X tonnes of carbon or Y tonnes of rare earth metals or Z square miles of wind turbines". Having said that, in Crypto it's a much more compelling argument because it's like "Here are some obvious, big costs that are fundamental to the system, and there are almost no benefits" whereas in AI the nature of these algorithms is that we're likely to improve them, so comparing training GPT-3 is kind of like measuring the power draw of the LHC. It's certainly big and it's certainly something we can talk about, but no one thinks we're going to be building 100,000 LHCs.