So glad you’re vegan. I’ve been wanting to make this point for ages. People scream about personal responsibility when it comes to AI but not about their diets or lifestyle
I try not to be too aggro about veganism but when someone starts shaming me for using ChatGPT it’s hard to resist the temptation to say “bro you just paid to torture a pig”
I know, like I want to write a piece on veganism but honestly it’s so personal to me I can’t even be reasonable or nice about it so I try to avoid the topic all together 😭
Strong agree, I struggle to keep it calm on this one, and so somewhat counterintuitively don't talk about it much despite feeling so strongly about it.
This is brilliant and bookmarked, a very useful resource. The real issue here is that many people just generally don't like AI and therefore it gets a negative halo, including environmentally, which as you say makes no sense. One can legitimately disagree about the thing itself, but as you so comprehensively show, the environmental side is not a meaningful one.
Yup to be clear I think a lot of other concerns about AI are well-founded. Splitting things up into specific issues to avoid the negative halo's super important.
Social problems like deepfakes, hallucinations/getting important info wrong, algorithmic bias, and risks from advanced AI weapons systems/other issues. Listed them in the intro
As do I. However, using AI to assist with certain aspects of the creative process (like spell checking, for instance) does not constitute a lack of ability or courage in creating something.
Many respected authors pay an editor to check their work and make needed corrections. Yet, the author isn't generally thought of as someone lacking ability or courage.
Enjoyed the article in general, but I couldn't find support for this bit in the linked article - it seems to say the opposite (that thinking hard doesn't materially change calorie consumption). Have i missed something?
>Thinking hard about something uses 4 calories more per minute. All food we eat causes carbon emissions, around 10-30% of your carbon footprint is your food. The average American diet emits between 1-4.5 kg of CO2 per 1000 calories. So thinking hard for an hour emits between 0.24-1 kg of CO2. That’s the same as 200-800 ChatGPT prompts.
An elevator is 100-200 Wh to move one floor. By walking down the stairs from my sixth floor apartment instead of summoning the elevator, then riding it back to the ground floor, I conserve enough energy for 400 queries.
I've seen arguments that it's bad because it uses enough energy to charge your phone to a nontrivial percentage (for most Androids, that'd be around 15-20%, not sure about iPhone). But that just means phones are efficient!
1000 Wh worth of claps for this post! Much of what people complain about on social media is produced and provided without commentary or evidence. The world of memes and social media responses is rife with emotionally triggered people who read something that was designed to spark a reaction one way or another, and very little actual research or reading goes into it.
Bravo for pulling back the curtain in your OMG long post and cheat sheet. It was nice to see someone do the math and present the evidence (as clearly and accurately as possible). When you whip out the comparisons, then drive to Costco and see families buying so much bottled water, you think the apocalypse is coming, I feel like using ChatGPT is just fine.
IMHO, what people should be decrying is the wasted AI efforts in things like Google search, Facebook, email summarization, and all the other BS that you see out there in the world now, simply because AI is the cool thing and everyone has to have it. There are good places for AI, essential problems to solve, but because AI is used so frivolously, it's easy to wag your finger.
Brilliant post. I‘ll be sharing at future workshops. This does come up during my seminars. I usually respond, yes, our modern lifestyle costs a lot of energy. We should all be mindful. And yes, some ppl (in my case teachers) are just resisters because they haven’t found use cases for it yet and do not understand why or how it can become an integral part of workflows.
I was just told I'm a moron for asking ChatGPT a question about how a vintage dress should be worn.
Someone who has 114K followers posted a video to Instagram asking how her (absolutely beautiful) dress should be worn. I thought the V neck was front, and Y neck was back. ChatGPT hilariously disagreed with me.
Some random follower then started harassing me over using ChatGPT.
Dude. That IG post alone (and the fact that she was responding to it) cost about 2000 times as much energy as me trying to figure out the issue for her.
The numbers i’ve read is that one 100 word ChatGPT generation uses up more than 0,5 l of water. So that seems like a lot more than you write here? (No expert though, just asking about the numbers.)
That’s based on one study from the Washington that I can’t find any info on. Open to being wrong on this so if you can find any more info I’d appreciate it! I should look into it more too
First of all, this is why I love Substack. Every time I question something, I get a nice reply back. So, here you can actually have civilised discussions about things. Not the polarising madness you see on X.
Second, I think the article you're looking for is this: Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models. It's available as PDF on arxiv[dot]org.
I'm wondering if the difference is in how you count? For example, ChatGPT doesn't only use up energy in the household where it's being used, but also in the huge servers around the world needed to keep it running.
I know that my home country (Iceland) is now a very popular place for building data servers because there's lots of space and the climate is cold. So, anecdotal evidence only so far, I've heard that parts of that nature will be transformed into data centers to run AI.
But at the same time, I'm guessing AI video makers must consume by far the most energy, no? Creating, probably, 99% rubbish, lol. But that's another question, I guess...
This is a good follow-up to your previous article, I especially liked the illustration of individual vs global impacts and how they don't mix. While I broadly agree with your thesis, there are a few points where I would like to hear your thoughts.
First, you mention "at most all chatbots are only using 1-3% of the energy used on AI", could you share your source for that? Have there been parallel increases in non-LLM AI use in recent years that led strains on the energy grid? In other words, if "ChatGPT is not the reason coal plants are reopening to power AI data centers", then what is?
Second, what are the impacts of reasoning models on energy use? Based on inference cost, I'd expect it to go up by an OOM. Has anybody written about this?
Finally, it seems possible that even though individual ChatGPT queries have a negligible impact, if LLMs gets integrated into every facet of our digital lives (e.g Microsoft Word attempting to predict your next word whenever you type), then the aggregate individual impact could become substantial. Do you know of anybody writing on this?
I did a botec in my first post trying to figure out how much energy chatbots are using compared to everything else using sources on how much energy ChatGPT and others are using and comparing them to AI as a whole. Was really hard to get it above 3% as an absolute max. I suspect the single biggest culprit is recommender systems. Every time anyone swipes on TikTok or instagram there’s an AI model working out what they’d like to watch more.
Yeah reasoning models likely raise the energy cost a lot. If I had to bet myself, I’d guess 4o uses around 0.3 Wh and O3 uses 3 Wh, but very unsure about both. I talk about that more in my most recent post.
Yup that’s true! A lot of people are already covering the broader climate impacts of LLMs, I don’t know anyone who stands out though. Definitely worth being concerned about!
This is so great, Andy! As a vegan, I feel somewhat vindicated...
I'm curious about your first footnote. Do you have more sources for the climate footprint of upstream AI processes? Curious what broader climate impacts you're worried about.
Yeah really happy you liked it! It seems like people have wildly different expectations for how much energy AI will use more broadly. I'd like to eventually make a separate post about how a lot of AI scale out will still be useful even if it uses a lot of energy, because the value we get per unit energy will probably be really high, but that's not a fight I'm looking to have right now and want to acknowledge that AI will probably use a lot more energy in the future. I liked this post as an overview: https://www.sustainabilitybynumbers.com/p/ai-energy-demand
Cheers! I embarrassingly commented after only reading an excerpt (facepalm; self-reflecting). Appreciate your energy budget breakdown. :)
In a recent OpenAI Academy webinar, a staff member proposed an intriguing comparison: the total energy required to sustain a human worker (food, housing, etc.) versus that needed to run an LLM performing comparable tasks.
Fascinating and somewhat unsettling. Even if the math ultimately supports delegating certain work to LLMs from an energy efficiency standpoint, it seems more philosophical than mathematical: What are the deeper implications of optimizing cognitive work primarily through the lens of energy consumption?
This breakdown of ChatGPT’s energy use is really eye-opening. Hard to imagine that each prompt uses roughly the same energy as a google search. This perspective really helps in addressing environmental concerns about AI usage 🧐
So glad you’re vegan. I’ve been wanting to make this point for ages. People scream about personal responsibility when it comes to AI but not about their diets or lifestyle
I try not to be too aggro about veganism but when someone starts shaming me for using ChatGPT it’s hard to resist the temptation to say “bro you just paid to torture a pig”
I know, like I want to write a piece on veganism but honestly it’s so personal to me I can’t even be reasonable or nice about it so I try to avoid the topic all together 😭
IMO more good public writing about animal welfare from people with large followings seems really useful!
I will at some point 😭
Have a separate section on animal welfare here if it's interesting https://andymasley.substack.com/s/animal-welfareveganism
Strong agree, I struggle to keep it calm on this one, and so somewhat counterintuitively don't talk about it much despite feeling so strongly about it.
This is brilliant and bookmarked, a very useful resource. The real issue here is that many people just generally don't like AI and therefore it gets a negative halo, including environmentally, which as you say makes no sense. One can legitimately disagree about the thing itself, but as you so comprehensively show, the environmental side is not a meaningful one.
Yup to be clear I think a lot of other concerns about AI are well-founded. Splitting things up into specific issues to avoid the negative halo's super important.
Specifically what concerns?
Social problems like deepfakes, hallucinations/getting important info wrong, algorithmic bias, and risks from advanced AI weapons systems/other issues. Listed them in the intro
I’m liking this comment because it correctly used terminology related to a cognitive bias :)
Wow, a post about AI that isn't decrying it as evil or on the “look how much better I am than you because I don't use it” bandwagon.
I was starting to think such a thing didn't exist around here.
Meh. I still prefer people who have the ability and the courage to write and draw for themselves.
As do I. However, using AI to assist with certain aspects of the creative process (like spell checking, for instance) does not constitute a lack of ability or courage in creating something.
Many respected authors pay an editor to check their work and make needed corrections. Yet, the author isn't generally thought of as someone lacking ability or courage.
Equivocate harder Jamie
Be nice in my comments section please 🫵😇
Hey @andy,
Enjoyed the article in general, but I couldn't find support for this bit in the linked article - it seems to say the opposite (that thinking hard doesn't materially change calorie consumption). Have i missed something?
>Thinking hard about something uses 4 calories more per minute. All food we eat causes carbon emissions, around 10-30% of your carbon footprint is your food. The average American diet emits between 1-4.5 kg of CO2 per 1000 calories. So thinking hard for an hour emits between 0.24-1 kg of CO2. That’s the same as 200-800 ChatGPT prompts.
Thanks for catching! Will circle back
Yup this was wrong, will edit and remove this specific point. Got my wires crossed with another number I think. Thank you!
fyi there were like 4 duplicate copies of this comment so I just removed the others
An elevator is 100-200 Wh to move one floor. By walking down the stairs from my sixth floor apartment instead of summoning the elevator, then riding it back to the ground floor, I conserve enough energy for 400 queries.
More like 40-60 Wh. Not sure where you're getting your numbers.
I've seen arguments that it's bad because it uses enough energy to charge your phone to a nontrivial percentage (for most Androids, that'd be around 15-20%, not sure about iPhone). But that just means phones are efficient!
1000 Wh worth of claps for this post! Much of what people complain about on social media is produced and provided without commentary or evidence. The world of memes and social media responses is rife with emotionally triggered people who read something that was designed to spark a reaction one way or another, and very little actual research or reading goes into it.
Bravo for pulling back the curtain in your OMG long post and cheat sheet. It was nice to see someone do the math and present the evidence (as clearly and accurately as possible). When you whip out the comparisons, then drive to Costco and see families buying so much bottled water, you think the apocalypse is coming, I feel like using ChatGPT is just fine.
IMHO, what people should be decrying is the wasted AI efforts in things like Google search, Facebook, email summarization, and all the other BS that you see out there in the world now, simply because AI is the cool thing and everyone has to have it. There are good places for AI, essential problems to solve, but because AI is used so frivolously, it's easy to wag your finger.
One of the best posts about AI usage I have ever seen. I think each person complaining about it should read this.
Thanks so much!
Brilliant post. I‘ll be sharing at future workshops. This does come up during my seminars. I usually respond, yes, our modern lifestyle costs a lot of energy. We should all be mindful. And yes, some ppl (in my case teachers) are just resisters because they haven’t found use cases for it yet and do not understand why or how it can become an integral part of workflows.
I was just told I'm a moron for asking ChatGPT a question about how a vintage dress should be worn.
Someone who has 114K followers posted a video to Instagram asking how her (absolutely beautiful) dress should be worn. I thought the V neck was front, and Y neck was back. ChatGPT hilariously disagreed with me.
Some random follower then started harassing me over using ChatGPT.
Dude. That IG post alone (and the fact that she was responding to it) cost about 2000 times as much energy as me trying to figure out the issue for her.
But no, I'm a monster for using AI.
Good grief. 🤣
The hyperbole gets so exhausting. I don't know how people are constantly speaking like that.
Good insight 😌. Can i translate part of this article into Spanish with links to you and a description of your newsletter?
Please do!
Many thanks.
The numbers i’ve read is that one 100 word ChatGPT generation uses up more than 0,5 l of water. So that seems like a lot more than you write here? (No expert though, just asking about the numbers.)
That’s based on one study from the Washington that I can’t find any info on. Open to being wrong on this so if you can find any more info I’d appreciate it! I should look into it more too
First of all, this is why I love Substack. Every time I question something, I get a nice reply back. So, here you can actually have civilised discussions about things. Not the polarising madness you see on X.
Second, I think the article you're looking for is this: Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models. It's available as PDF on arxiv[dot]org.
I'm wondering if the difference is in how you count? For example, ChatGPT doesn't only use up energy in the household where it's being used, but also in the huge servers around the world needed to keep it running.
I know that my home country (Iceland) is now a very popular place for building data servers because there's lots of space and the climate is cold. So, anecdotal evidence only so far, I've heard that parts of that nature will be transformed into data centers to run AI.
But at the same time, I'm guessing AI video makers must consume by far the most energy, no? Creating, probably, 99% rubbish, lol. But that's another question, I guess...
This is a good follow-up to your previous article, I especially liked the illustration of individual vs global impacts and how they don't mix. While I broadly agree with your thesis, there are a few points where I would like to hear your thoughts.
First, you mention "at most all chatbots are only using 1-3% of the energy used on AI", could you share your source for that? Have there been parallel increases in non-LLM AI use in recent years that led strains on the energy grid? In other words, if "ChatGPT is not the reason coal plants are reopening to power AI data centers", then what is?
Second, what are the impacts of reasoning models on energy use? Based on inference cost, I'd expect it to go up by an OOM. Has anybody written about this?
Finally, it seems possible that even though individual ChatGPT queries have a negligible impact, if LLMs gets integrated into every facet of our digital lives (e.g Microsoft Word attempting to predict your next word whenever you type), then the aggregate individual impact could become substantial. Do you know of anybody writing on this?
I did a botec in my first post trying to figure out how much energy chatbots are using compared to everything else using sources on how much energy ChatGPT and others are using and comparing them to AI as a whole. Was really hard to get it above 3% as an absolute max. I suspect the single biggest culprit is recommender systems. Every time anyone swipes on TikTok or instagram there’s an AI model working out what they’d like to watch more.
Yeah reasoning models likely raise the energy cost a lot. If I had to bet myself, I’d guess 4o uses around 0.3 Wh and O3 uses 3 Wh, but very unsure about both. I talk about that more in my most recent post.
Yup that’s true! A lot of people are already covering the broader climate impacts of LLMs, I don’t know anyone who stands out though. Definitely worth being concerned about!
This is so great, Andy! As a vegan, I feel somewhat vindicated...
I'm curious about your first footnote. Do you have more sources for the climate footprint of upstream AI processes? Curious what broader climate impacts you're worried about.
Fellow vegans rise up.
Yeah really happy you liked it! It seems like people have wildly different expectations for how much energy AI will use more broadly. I'd like to eventually make a separate post about how a lot of AI scale out will still be useful even if it uses a lot of energy, because the value we get per unit energy will probably be really high, but that's not a fight I'm looking to have right now and want to acknowledge that AI will probably use a lot more energy in the future. I liked this post as an overview: https://www.sustainabilitybynumbers.com/p/ai-energy-demand
Cheers! I embarrassingly commented after only reading an excerpt (facepalm; self-reflecting). Appreciate your energy budget breakdown. :)
In a recent OpenAI Academy webinar, a staff member proposed an intriguing comparison: the total energy required to sustain a human worker (food, housing, etc.) versus that needed to run an LLM performing comparable tasks.
Fascinating and somewhat unsettling. Even if the math ultimately supports delegating certain work to LLMs from an energy efficiency standpoint, it seems more philosophical than mathematical: What are the deeper implications of optimizing cognitive work primarily through the lens of energy consumption?
First time reader here (subscribed!) and I get the sense you have and will continue to ponder this question. Excited to dig in!
I'll have to find a way to cite this article it's definitely food for thought. Although I can only read something like this in bits and bites myself.
Yup tried to split it up to be read in bits and bites. Would be really excited to have it shared!
This breakdown of ChatGPT’s energy use is really eye-opening. Hard to imagine that each prompt uses roughly the same energy as a google search. This perspective really helps in addressing environmental concerns about AI usage 🧐