4 ms·

I don't understand the "whatboutism" angle. Running LLMs requires massive amounts of electricity and water. Period. That beef consumption and other activities

I don't understand the "whatboutism" angle. Running LLMs requires massive amounts of electricity and water. Period. That beef consumption and other activities also require massive amounts of resources is independent. Both can be true. And both need to be addressed. We no longer have the luxury to pick and choose which climate interventions we find most convenient; we need to succeed on all fronts. If the patient is dehydrated and bleeding, do doctors sit around debating if they should give them water or gauze? No, they do both immediately.

metalcrow 1y ago
They would if the patient is alive and fighting them! In our case, the patient is society and it is quite definitely not all unified that we cannot pick and choose. If the doctors want to save their patient, they either need to tie them down and do what is needed, or they need to work within the confines of what the patient allows while trying to convince them to allow more.
- dgb23 1y ago
  The patient is kicking and screaming, because they are an addict. Regardless, I’m not convinced by the parent comment. It basically suggests we solve this without analysis of costs and effects. Sounds insane to me. Especially since the most promising solutions come from rapidly developing technologies and research. Turning off the lights just sends us backwards.
  - -__---____-ZXyw 1y ago
    > Especially since the most promising solutions come from rapidly developing technologies and research What solutions are you referring to? This sentiment baffles me when I see it. It's not like we learned we'd this problem yesterday, and now we've to guess how we might solve it based on guessing. For 35 years, "rapidly developing technologies and research" hasn't improved the numbers. I understand that there's a regular stream of news which touts one thing or another as a possible solution, but which one of them has done anything? How long do we wait for this techno-optimism to bear fruit?
    - dgb23 1y ago
      A few obvious examples are the large gains in grid stability and efficiency, electric cars, the adoption and progress of solar panels. The latter is outpacing the expectations. And it’s not just engineering and manufacturing. These things have to be installed, maintained etc. In different types of places, which means you need specialized knowledge for all sorts of things. On the research side there has been much progress on climate and weather models. Software also plays a role in all of these, and material science and… I’m not a techno optimist, but rather see these developments as incredibly necessary. Yes they should have happened yesterday. Yes I’m incredibly frustrated with political leadership all around the world and especially with outdated economic models. But the only way is forward if it’s not too late.
btilly 1y ago
According to https://www.theverge.com/24066646/ai-electricity-energy-watts-generative-consumption https://www.theverge.com/24066646/ai-electricity-energy-watt... a reasonable estimate is half a percent of global energy usage by 2027. On the one hand, this is a lot of energy. On the other, that's in the same general neighborhood as what we could save by maintaining proper tire pressure on our cars. (See https://www.factcheck.org/2008/08/the-truth-about-tire-pressure/ https://www.factcheck.org/2008/08/the-truth-about-tire-press... for more about that.)
- perrygeo 1y ago
  I get that. I just don't get why you frame it as an either/or question as though they were mutually exclusive. We can reduce data center emissions, and improve tire maintenance. Por que no los dos?
clbrmbr 1y ago
Naive question: why doesn’t the market regulate electricity consumption? I’ve heard the answers pre-AI, but I wonder how this new general purpose use of electricity changes the calculus?
teruakohatu 1y ago
> massive amounts of electricity and water ... we need to succeed on all fronts Datacenters' used for training or batch jobs can be placed in places where water and power are plentiful, ambient temperature are relatively low and government/society are stable. A datacenter is going to be built at the bottom of New Zealand where all these things are true. There must be plenty of places in the world where this holds true. At least for training, I think it is possible to have our cake and eat it. For real-time inference, probably not.
- distances 1y ago
  > A datacenter is going to be built at the bottom of New Zealand where all these things are true. There must be plenty of places in the world where this holds true. There are many datacenters being built in Finland. Water is not running out, and one of the cheapest electricity prices in Europe thanks to plentiful wind power.
- int_19h 1y ago
  Given typical inference speeds for SOTA models, I rather suspect that you could also do it for inference without latency being particularly notable. We might need to lay a few more cables to New Zealand though.
roywiggins 1y ago
Okay, but there is some reason why you personally are still spending electrons posting on the internet and not, presumably, entertaining yourself by rolling coal, and it's probably because the amount of energy you're using to browse and post on HN seems trivial to you. Does the energy use differential between Python and C++ seem important enough to always pick C++ for everything? What I'm saying is that casual LLM use is also essentially trivial, and scolding people for using querying ChatGPT when you wouldn't scold people for, say, making a dozen Google searches or watching a Netflix show or taking a long shower or driving to the mall, is quite silly and a waste of time. Similarly, there's a reason why doctors will tell you to worry more about your cigarette habit than almost any other bad habits you have, because it's so much more likely to kill you that it practically demands to be prioritized.
TheNewsIsHere 1y ago
I used to live in a place where there are a lot of data centers now. There were only a few just a few years ago. Now there are tens and more under construction. I returned to that place to visit family and friends last year. It was an eye opening experience. The people who live there have taken a keen interest on curtailing further development of any new data centers. One of the chief complaints is the constant power issues that didn’t exist before 2021-2022. The locals argue that this is the result of the dramatic infusion of AI into every technology product. They’re likely not wrong. The communities in the area became quite politically active over the issue and have retained all manner of analysts and scientists, journalists and investigators, and so on, to aide them in making a political case against future data center development. The thing that got me was the complaints about the noise. Over the past few years the locals and those in their employ have been monitoring noise levels near the data centers and they’ve tracked sustained increases in noise pollution with the timeline over which the use of AI has exploded. There it has become sort of a trope to measure the working day by how much noise pollution is produced by any data center nearby. Mostly belonging to a hyperscaler known by an acronym. The data centers have reportedly not been the boon that was promised, which seems to be seen as insult to injury. The area already has a vibrant tech scene independent of data center operations. So the locals don’t really see value in allowing more data centers to be built, and they’re starting to organize politically around the idea of preventing future data center construction and implementing heavy usage based taxation on utilities used by the existing ones.
- RetroTechie 1y ago
  Anyone else notice the parallels with crypto mining industry?
lolinder 1y ago
Whataboutism is bad, but there's a distinction between whataboutism and making comparisons to get a sense of the true scale of the problem. If you're not allowed to ever compare relative scales of things without being accused of whataboutism then we're paving the way to making every molehill a mountain. Two average American's beef consumption for training GPT-3 would place the cost in water of training GPT-3 at 1/150000000 of the US beef industry's average water consumption. That means it's a rounding error relative to all the other uses we place on water, and pointing that out is not whataboutism, it's putting the problem in its proper scale.
- roywiggins 1y ago
  if a few thousand westerners who want to atone for LLM water use skipped a few burgers, the planet would come out ahead on net I think
  - -__---____-ZXyw 1y ago
    No burgers and no LLMs is better, though, in absolute terms. Or at least, that's one of the points being made, I believe
joshjob42 1y ago
We have a constraint, say keeping warming below 1.5 or 2C, and we want to achieve it with as little pain as possible. You use a laptop, you probably play some video games, you see movies, you probably eat meat, you probably drive a car around town instead of a velomobile. So the question is, how much energy is used by AI and what does it get us? Google's latest TPUs consume ~250W per fp8-petaflop at the data-center level (ie including cooling etc, according to their recent presentations). Even assuming 50% utilization which is a bit poor, that's say 0.5kW per fp8-petaflop, or ~1kW per 10^15 dense-parameter-equivalent tokens per second. So using a huge model, say 10T dense-equivalent parameters, and doing inference at the rate of o3-mini or Gemini 2.5 Pro (around 150 tokens a second) consumes 1.5kW during generation time. But maybe a better way to think about it is just that you'd be using 10J of energy per token, or ~15J of energy per word. In that context then, generating a book the length of Fellowship of the Ring (~180k words) would consume ~0.75kWh of energy. That's about like playing a video game on a PS5 for ~3 hours. In the US, that's on average ~270g of CO2, like driving ~1.5 miles or so in a Prius. It'd be about like riding ~30 miles on an e-bike. Another way to think of it might be that if you are texting with a friend, wherein lets say you both aggressively type at ~120wpm, you'd be using ~15W of power to chat with an AI at a similar pace, about the power draw of your friend's Macbook Air he would be typing on. That's around 0.5-0.7 miles on an ebike per hour of chatting. So even compared to your typical leisure activities, chatting with an AI is comparable in resource use to whatever else you'd be doing in that time most likely. And of course, a human in the US on average generates ~1.5kg of CO2 per hour they're alive in the US. In Bolivia, it's ~150g/hr, and in Sudan it's ~50g/hr or so. So if you're a company and you want to be climate conscious merely hiring someone who is a ~ subsistence farmer to do a job such that their standard of living rises enough to roughly emit the CO2 of the average Bolivian would mean they emit more carbon in a typical day than running an AI to produce as many words as are in the entire Game of Thrones book series to date. And if you were going to help someone immigrate from Bolivia to the US, you could have the AI write ~10 Songs of Ice and Fire a day for the same net CO2 output. Not to say that we shouldn't do those things because of climate or whatever, but I'm just saying that the objective energy and climate impact of using these models for things is small compared to basically any alternative for completing a task and most entertainment activities people do (movie theaters for instance use ~7kW or so while running, so in the time you spend watching the Fellowship of the Ring in a theater, an AI could write a text as long as FOTR 28 times).

metalcrow 1y ago

They would if the patient is alive and fighting them! In our case, the patient is society and it is quite definitely not all unified that we cannot pick and choose. If the doctors want to save their patient, they either need to tie them down and do what is needed, or they need to work within the confines of what the patient allows while trying to convince them to allow more.

dgb23 1y ago
The patient is kicking and screaming, because they are an addict. Regardless, I’m not convinced by the parent comment. It basically suggests we solve this without analysis of costs and effects. Sounds insane to me. Especially since the most promising solutions come from rapidly developing technologies and research. Turning off the lights just sends us backwards.
- -__---____-ZXyw 1y ago
  > Especially since the most promising solutions come from rapidly developing technologies and research What solutions are you referring to? This sentiment baffles me when I see it. It's not like we learned we'd this problem yesterday, and now we've to guess how we might solve it based on guessing. For 35 years, "rapidly developing technologies and research" hasn't improved the numbers. I understand that there's a regular stream of news which touts one thing or another as a possible solution, but which one of them has done anything? How long do we wait for this techno-optimism to bear fruit?
  - dgb23 1y ago
    A few obvious examples are the large gains in grid stability and efficiency, electric cars, the adoption and progress of solar panels. The latter is outpacing the expectations. And it’s not just engineering and manufacturing. These things have to be installed, maintained etc. In different types of places, which means you need specialized knowledge for all sorts of things. On the research side there has been much progress on climate and weather models. Software also plays a role in all of these, and material science and… I’m not a techno optimist, but rather see these developments as incredibly necessary. Yes they should have happened yesterday. Yes I’m incredibly frustrated with political leadership all around the world and especially with outdated economic models. But the only way is forward if it’s not too late.

btilly 1y ago

According to https://www.theverge.com/24066646/ai-electricity-energy-watts-generative-consumption https://www.theverge.com/24066646/ai-electricity-energy-watt... a reasonable estimate is half a percent of global energy usage by 2027. On the one hand, this is a lot of energy. On the other, that's in the same general neighborhood as what we could save by maintaining proper tire pressure on our cars. (See https://www.factcheck.org/2008/08/the-truth-about-tire-pressure/ https://www.factcheck.org/2008/08/the-truth-about-tire-press... for more about that.)

perrygeo 1y ago
I get that. I just don't get why you frame it as an either/or question as though they were mutually exclusive. We can reduce data center emissions, and improve tire maintenance. Por que no los dos?

clbrmbr 1y ago

Naive question: why doesn’t the market regulate electricity consumption? I’ve heard the answers pre-AI, but I wonder how this new general purpose use of electricity changes the calculus?

teruakohatu 1y ago

> massive amounts of electricity and water ... we need to succeed on all fronts Datacenters' used for training or batch jobs can be placed in places where water and power are plentiful, ambient temperature are relatively low and government/society are stable. A datacenter is going to be built at the bottom of New Zealand where all these things are true. There must be plenty of places in the world where this holds true. At least for training, I think it is possible to have our cake and eat it. For real-time inference, probably not.

distances 1y ago
> A datacenter is going to be built at the bottom of New Zealand where all these things are true. There must be plenty of places in the world where this holds true. There are many datacenters being built in Finland. Water is not running out, and one of the cheapest electricity prices in Europe thanks to plentiful wind power.
int_19h 1y ago
Given typical inference speeds for SOTA models, I rather suspect that you could also do it for inference without latency being particularly notable. We might need to lay a few more cables to New Zealand though.

roywiggins 1y ago

Okay, but there is some reason why you personally are still spending electrons posting on the internet and not, presumably, entertaining yourself by rolling coal, and it's probably because the amount of energy you're using to browse and post on HN seems trivial to you. Does the energy use differential between Python and C++ seem important enough to always pick C++ for everything? What I'm saying is that casual LLM use is also essentially trivial, and scolding people for using querying ChatGPT when you wouldn't scold people for, say, making a dozen Google searches or watching a Netflix show or taking a long shower or driving to the mall, is quite silly and a waste of time. Similarly, there's a reason why doctors will tell you to worry more about your cigarette habit than almost any other bad habits you have, because it's so much more likely to kill you that it practically demands to be prioritized.

TheNewsIsHere 1y ago

I used to live in a place where there are a lot of data centers now. There were only a few just a few years ago. Now there are tens and more under construction. I returned to that place to visit family and friends last year. It was an eye opening experience. The people who live there have taken a keen interest on curtailing further development of any new data centers. One of the chief complaints is the constant power issues that didn’t exist before 2021-2022. The locals argue that this is the result of the dramatic infusion of AI into every technology product. They’re likely not wrong. The communities in the area became quite politically active over the issue and have retained all manner of analysts and scientists, journalists and investigators, and so on, to aide them in making a political case against future data center development. The thing that got me was the complaints about the noise. Over the past few years the locals and those in their employ have been monitoring noise levels near the data centers and they’ve tracked sustained increases in noise pollution with the timeline over which the use of AI has exploded. There it has become sort of a trope to measure the working day by how much noise pollution is produced by any data center nearby. Mostly belonging to a hyperscaler known by an acronym. The data centers have reportedly not been the boon that was promised, which seems to be seen as insult to injury. The area already has a vibrant tech scene independent of data center operations. So the locals don’t really see value in allowing more data centers to be built, and they’re starting to organize politically around the idea of preventing future data center construction and implementing heavy usage based taxation on utilities used by the existing ones.

RetroTechie 1y ago
Anyone else notice the parallels with crypto mining industry?

lolinder 1y ago

Whataboutism is bad, but there's a distinction between whataboutism and making comparisons to get a sense of the true scale of the problem. If you're not allowed to ever compare relative scales of things without being accused of whataboutism then we're paving the way to making every molehill a mountain. Two average American's beef consumption for training GPT-3 would place the cost in water of training GPT-3 at 1/150000000 of the US beef industry's average water consumption. That means it's a rounding error relative to all the other uses we place on water, and pointing that out is not whataboutism, it's putting the problem in its proper scale.

roywiggins 1y ago
if a few thousand westerners who want to atone for LLM water use skipped a few burgers, the planet would come out ahead on net I think
- -__---____-ZXyw 1y ago
  No burgers and no LLMs is better, though, in absolute terms. Or at least, that's one of the points being made, I believe

joshjob42 1y ago

We have a constraint, say keeping warming below 1.5 or 2C, and we want to achieve it with as little pain as possible. You use a laptop, you probably play some video games, you see movies, you probably eat meat, you probably drive a car around town instead of a velomobile. So the question is, how much energy is used by AI and what does it get us? Google's latest TPUs consume ~250W per fp8-petaflop at the data-center level (ie including cooling etc, according to their recent presentations). Even assuming 50% utilization which is a bit poor, that's say 0.5kW per fp8-petaflop, or ~1kW per 10^15 dense-parameter-equivalent tokens per second. So using a huge model, say 10T dense-equivalent parameters, and doing inference at the rate of o3-mini or Gemini 2.5 Pro (around 150 tokens a second) consumes 1.5kW during generation time. But maybe a better way to think about it is just that you'd be using 10J of energy per token, or ~15J of energy per word. In that context then, generating a book the length of Fellowship of the Ring (~180k words) would consume ~0.75kWh of energy. That's about like playing a video game on a PS5 for ~3 hours. In the US, that's on average ~270g of CO2, like driving ~1.5 miles or so in a Prius. It'd be about like riding ~30 miles on an e-bike. Another way to think of it might be that if you are texting with a friend, wherein lets say you both aggressively type at ~120wpm, you'd be using ~15W of power to chat with an AI at a similar pace, about the power draw of your friend's Macbook Air he would be typing on. That's around 0.5-0.7 miles on an ebike per hour of chatting. So even compared to your typical leisure activities, chatting with an AI is comparable in resource use to whatever else you'd be doing in that time most likely. And of course, a human in the US on average generates ~1.5kg of CO2 per hour they're alive in the US. In Bolivia, it's ~150g/hr, and in Sudan it's ~50g/hr or so. So if you're a company and you want to be climate conscious merely hiring someone who is a ~ subsistence farmer to do a job such that their standard of living rises enough to roughly emit the CO2 of the average Bolivian would mean they emit more carbon in a typical day than running an AI to produce as many words as are in the entire Game of Thrones book series to date. And if you were going to help someone immigrate from Bolivia to the US, you could have the AI write ~10 Songs of Ice and Fire a day for the same net CO2 output. Not to say that we shouldn't do those things because of climate or whatever, but I'm just saying that the objective energy and climate impact of using these models for things is small compared to basically any alternative for completing a task and most entertainment activities people do (movie theaters for instance use ~7kW or so while running, so in the time you spend watching the Fellowship of the Ring in a theater, an AI could write a text as long as FOTR 28 times).