One miss in this post is that the author tries to make their point by comparing energy consumption of LLMs to arbitrary points of reference. We should be comparing them to their relevant parallels.
Comparing a ChatGPT query to an hour long Zoom call isn't useful. The call might take up ~1700 mL of water, but that is still wildly more efficient than what we used to do prior - travel/commute to meet in person. The "10x a Google search" point is relevant because for many of the use cases mentioned in this post and others like it (e.g. "try asking factual questions!"), you could just as easily get that with 1 Google search and skimming the results.
I have found use for LLMs in software development, but I'd be lying if I said I couldn't live without it. Almost every use case of an LLM has a simple alternative - often just employing critical thinking or learning a new skill.
It feels like this post is a long way of saying "yes, there are negative impacts, but I value my time more".
I basically do think that at some threshold it's important to weigh your time against negative impacts. I personally avoid taking flights whenever I can because of the climate and think that's worth my time relative to the emissions saved, but I also never worry about optimizing the energy use of my digital clock because that would take too much time relative to the emissions I could save. ChatGPT exists somewhere between those two things, and my argument in the post is that it's much closer to the digital clock.
10 gallons of fuel's worth of energy could be used to ask ChatGTP 100,000 questions (assuming 3 Wh per question) or power a digital clock (1-2 W) for 35 years. If you assume you ask ChatGPT 8 questions per day, it's using exactly as much energy as a digital clock. Personal use of ChatGPT is much closer to the clock.
The marginal difference of someone sitting or not sitting on a specific flight isn’t 10 gallons of fuel. It’s the capacity of the entire aircraft making that trip / number of passengers that gives 10 gallons on average because if more people make a trip eventually you need an extra aircraft.
So on an apples to apples comparison ChatGPT including training, hardware, etc uses a fuck of a lot more than just 3 Wh per question.
The training cost adds a minuscule amount to the marginal cost of each search. The best sources I can find on training GPT-4 say it used 50 GWh of energy to train. GPT-4 is being used for a lot more than ChatGPT, but even if we assume that ChatGPT is all it's used for if we divide the total cost of energy by the total number of ChatGPT searches it increases the cost of each search by 10% (I go into this in the post) raising the marginal cost including training from 3 Wh to 3.1 Wh.
Not so fast, billions of investment wasn’t spent on just 50 GWh of electricity.
The final step of training one specific model might have used 50 GWh which only costs what 0.011 billion? What the fuck do you think they are spending the rest of that money on?
And what do you think environmental impact of that spending was?
Most of that money went into infrastructure + human talent, with energy costs at the tail end. I'd be happy to include the energy costs of producing the materials, feeding the humans and moving them around etc. but those aren't usually included in environmentalist critiques of the energy used in training. I couldn't use the original article to unpack every possible interpretation of how much energy was used without it turning into a book. If we did the same for plane rides we'd also need to include the energy cost of assembling the planes, training and feeding the human pilots etc.
Fair enough, I’m not saying you have to do the analysis.
I was working from a rough order of magnitude estimate from their spending and calling their numbers BS.
Airlines are a better studied industry so you can find academic literature or use direct fuel costs which is a large enough cost you can get fairly close just using it alone.
Sorry but this doesn't add up and it's not fixable because the argument is broken and I doubt the numbers are even in the right order of magnitude when considered in context. (That's before considering that it also had the effect of you writing and sharing this post - what's the second-order-effecf of people seeing it increasing their use as result of that and resharing the same idea with their peers etc? What's the network effect of your digital clock?)
You are rationalizing your laziness in an attempt to cope with the negative effects that you truly deep down know are there.
Sounds to me like you're rationalizing your laziness in researching and sharing more accurate numbers because you want to justify "the negative effects that you truly deep down know are there".
If I'm going to be accused of rationalizing my laziness I need at least one source saying my numbers are off. I'm going off the numbers provided by environmentalist critics of ChatGPT. I hope that by sharing my post people feel exactly as guilty about using ChatGPT as they do about buying one additional digital clock.
Comparing a ChatGPT query to an hour long Zoom call isn't useful. The call might take up ~1700 mL of water, but that is still wildly more efficient than what we used to do prior - travel/commute to meet in person. The "10x a Google search" point is relevant because for many of the use cases mentioned in this post and others like it (e.g. "try asking factual questions!"), you could just as easily get that with 1 Google search and skimming the results.
I have found use for LLMs in software development, but I'd be lying if I said I couldn't live without it. Almost every use case of an LLM has a simple alternative - often just employing critical thinking or learning a new skill.
It feels like this post is a long way of saying "yes, there are negative impacts, but I value my time more".