On the other hand, the energy consumption isn’t static and my impression is that compute scaling is still a dominant mechanism for increasing model performance, both for RL post-training and test time compute scaling
If we assume OpenAI will have similar power consumption and that emissions are linear in compute, then it’s only two more rounds of 10x increase in compute capacity to hit parity with animal agriculture
I think that’s plausible enough to be worth more careful consideration and discussion. Hopefully investments in more efficient models continue to pay off quickly enough to offset this, but failing that I’m not convinced the frontier labs will stop pursuing compute scaling just because it’s environmentally unsustainable
The less power they use the more money they make whereas in most other industries they either dont care, or want you to use more energy. Seems like the incentives are more aligned here.
Doesn't the growth projection of it being half of all animal agriculture in ten years not concern you? Using the same power as the activity to feed 5 billion people seems a bit much.
Top comment: whataboutism. AI threads never disappoint here.
> While it is possible to care about two things at once
It is, actually. Full stop.
And there is no reason not to fight against a new threat particularly hard, even when old threats are already there.
One doesn't necessarily write about plant based diets each time one writes about AI's environmental impacts, because mixing things isn't particularly effective.
there are people who even hate EVs just because producing batteries pollutes a lot and proceed to get a normal car because it's not just "a battery on wheels"
or the people who don't eat animals to save the planet and have a better environment
This was flagged dead, but I vouched to revive it. It seems like a decent overview article of a couple recent papers on the potential climate impact of AI. It doesn't feel AI generated itself---which admittedly would be ironic! Even if you think the whole climate narrative is bunk, please explain why in a comment rather than flagging this away. Or am I missing some reason this particular writeup isn't worthy of discussion?
I think as a planet (forget about borders and superpowers already) we need to ask ourselves where we want this to go.
Wrecking the little natural resources we have left just to pretend we’ve invented “intelligence” seems rather illogical. We’re still very much just simple “post-industrial barbarians”.
Film industry, video game industry, car industry, air travel industry, iPhone industry, Toys R Us industry... were they worth it?
The "highly educated" YOLO'd it the last 60 years cause they'll be dead. Such an enlightened people all those "knowledge workers" who exploited labor and destroyed the environment.
3C in 50 years is locked in. Get ready for Mad Max as famines, equator becomes a heat bubble around the globe force mass migrations at the last minute since no one will pay attention until the last minute.
Not to say I don’t share your skepticism, but I think the theory is that it does if it accelerates progress on open scientific questions like Fusion power generation, battery chemistry, etc.
I think Spain had the perfect solution to this that you mandate on a nation level that all data centers have to be 80% powered by renewables they made sure to put in the right caveats, for instance you’ll have to directly buy the renewable electricity for that 80%, you can’t offset it by buying carbon offsets
Anthropic's ethics points are already exhausted by their pledge not to eradicate humanity through uncontrolled runaway superintelligence. (They only do controlled runaway superintelligence). You can't expect them to also deal with small problems like climate change at that level!
What used to take many, many hours of research is now nearly instant. Software projects are greatly compressed. One worker, with AI, can do the work of many without it.
> One worker, with AI, can do the work of many without it.
> All this means less resource usage.
Can you explain your math here, because this doesn't make any sense whatsoever.
Workers are still working the same amount of hours they always did, they're just doing way more work in that time via AI systems. How exactly does this result in _less_ resource usage?
The only change AI is bringing to the future of work is that it's squeezing way more productivity out of employees at no benefit to them.
OK, that's a start. Is this because the studies are wrong and increased use of AI isn't actually going to dramatically increase fossil fuel usage, or because increased fossil fuel consumption isn't actually something that we should worry about?
I'm sure there are arguments for both of these (clean energy will quickly replace the dirty initial sources/why worry about decadal scale global warming in the face of FOOM?) but it would help if you actually made some arguments rather than just gesturing.
It’s hard to see how increased fossil fuel consumption will not be a problem given that we haven’t modeled tipping points at all.
However, focusing on the AI industry is beyond silly.
For example, this article focuses on the 25.7 million metric ton Co2E the 400 largest data centers emit (and I suspect this is data centers, not AI data centers specifically), and let’s say the total data center emission in the U.S. is 3 times that.
So that’s about 75million MTCO2E
Well, the US dairy industry alone emits about 160 million MTCO2E in a year. And dairy is a trivially substitutable product (with half a dozen plant based direct alternatives, but frankly, the US dairy industry would likely collapse if it wasn’t for the subsidies showered on it by the government), nd has a whole host of other issues.
AI on the other hand is doing completely new rhinfs, and is likely more efficient than the alternative ways of doing the same stuff.
If you’re seriously concerned about emissions start with eliminating subsidies to the dairy industry. That alone would wipe out half the industry and buy you enough emissions saved to more than offset AI emissions for another decade.
What this really shows is that this is not being taken seriously but is close to essentially clickbait (the research is necessary, but the idea that AI is even close to the biggest emissions concern is the clickbait aspect).
It’s wiping out any gains we have made in the past few years and accelerating in the opposite direction. This article only discussed the pollution aspect, not the heating or water loss aspects. The effect is much much greater than a number of increased emissions. It’s backsliding work we’ve taken for decades to get to and saying “fuck it all” for a chatbot that has none to negative benefit to all of society.
The 75 million are in addition to what the diary industry is already emitting and evidently much letts entwined with and necessary for the wider economy as the diary industry.
It's a bit like sitting in a car, driving in the opposite direction of where you have to go, and then accelerating - but then saying "well, I only accelerated a little!"
For that analogy to work, it’s as if your accelerator is being driven by a stack of blocks pushing the pedal down.
Your lowest block is 25lbs, the next one is 15lbs, the next one is 10lbs, and now you’ve added a block that’s another 5lbs.
It would be complaining about the last 5lbs gold block and asking for that to be removed and thrown away, but ignoring the 15lbs brick that’s made of stone and not very valuable.
The fact that the AI emissions are newer is no reason to give that additional scrutiny. The scrutiny should be directed to the parts of the problems that are bigger and have less value and are more easily substitute able.
> Animal agriculture produces an estimated 6.2 billion metric tons (6,200 megatons) of CO₂-equivalent emissions per year
While it is possible to care about two things at once, if this was really about the climate and CO2 emissions I think people would focus elsewhere.
This argument falls flat with me, and I care about climate change.
If we assume OpenAI will have similar power consumption and that emissions are linear in compute, then it’s only two more rounds of 10x increase in compute capacity to hit parity with animal agriculture
I think that’s plausible enough to be worth more careful consideration and discussion. Hopefully investments in more efficient models continue to pay off quickly enough to offset this, but failing that I’m not convinced the frontier labs will stop pursuing compute scaling just because it’s environmentally unsustainable
Are we assuming that chips just stop getting more power efficient?
Are we assuming that AI never reaches saturation? About half of US adults already use AI.
Over half of US adults are already using AI regularly.
https://img-9gag-fun.9cache.com/photo/a7oxg6L_460s.jpg
What's it "really about"
> While it is possible to care about two things at once
It is, actually. Full stop.
And there is no reason not to fight against a new threat particularly hard, even when old threats are already there.
One doesn't necessarily write about plant based diets each time one writes about AI's environmental impacts, because mixing things isn't particularly effective.
You need to be an eco-extremist (or just disconnected from reality) to read GP as anything other than a very reasonable value judgement.
What's an eco-extremist, already?
Don't bother answering, actually. I prefer not knowing.
there are people who even hate EVs just because producing batteries pollutes a lot and proceed to get a normal car because it's not just "a battery on wheels"
or the people who don't eat animals to save the planet and have a better environment
A much more interesting post is also on my front-page, which is: https://betterdatacenterproject.com/2026/11/16/how-ai-is-pow...
Wrecking the little natural resources we have left just to pretend we’ve invented “intelligence” seems rather illogical. We’re still very much just simple “post-industrial barbarians”.
https://news.harvard.edu/gazette/story/2023/01/harvard-led-a...
Film industry, video game industry, car industry, air travel industry, iPhone industry, Toys R Us industry... were they worth it?
The "highly educated" YOLO'd it the last 60 years cause they'll be dead. Such an enlightened people all those "knowledge workers" who exploited labor and destroyed the environment.
3C in 50 years is locked in. Get ready for Mad Max as famines, equator becomes a heat bubble around the globe force mass migrations at the last minute since no one will pay attention until the last minute.
What used to take many, many hours of research is now nearly instant. Software projects are greatly compressed. One worker, with AI, can do the work of many without it.
All this means less resource usage.
A fair article would factor these in.
> All this means less resource usage.
Can you explain your math here, because this doesn't make any sense whatsoever.
Workers are still working the same amount of hours they always did, they're just doing way more work in that time via AI systems. How exactly does this result in _less_ resource usage?
The only change AI is bringing to the future of work is that it's squeezing way more productivity out of employees at no benefit to them.
There are many way, way more applicable reasons to be disapproving of ai than climate change, this ain't it whatsoever
I'm sure there are arguments for both of these (clean energy will quickly replace the dirty initial sources/why worry about decadal scale global warming in the face of FOOM?) but it would help if you actually made some arguments rather than just gesturing.
However, focusing on the AI industry is beyond silly.
For example, this article focuses on the 25.7 million metric ton Co2E the 400 largest data centers emit (and I suspect this is data centers, not AI data centers specifically), and let’s say the total data center emission in the U.S. is 3 times that.
So that’s about 75million MTCO2E
Well, the US dairy industry alone emits about 160 million MTCO2E in a year. And dairy is a trivially substitutable product (with half a dozen plant based direct alternatives, but frankly, the US dairy industry would likely collapse if it wasn’t for the subsidies showered on it by the government), nd has a whole host of other issues.
AI on the other hand is doing completely new rhinfs, and is likely more efficient than the alternative ways of doing the same stuff.
If you’re seriously concerned about emissions start with eliminating subsidies to the dairy industry. That alone would wipe out half the industry and buy you enough emissions saved to more than offset AI emissions for another decade.
What this really shows is that this is not being taken seriously but is close to essentially clickbait (the research is necessary, but the idea that AI is even close to the biggest emissions concern is the clickbait aspect).
It's a bit like sitting in a car, driving in the opposite direction of where you have to go, and then accelerating - but then saying "well, I only accelerated a little!"
Your lowest block is 25lbs, the next one is 15lbs, the next one is 10lbs, and now you’ve added a block that’s another 5lbs.
It would be complaining about the last 5lbs gold block and asking for that to be removed and thrown away, but ignoring the 15lbs brick that’s made of stone and not very valuable.
The fact that the AI emissions are newer is no reason to give that additional scrutiny. The scrutiny should be directed to the parts of the problems that are bigger and have less value and are more easily substitute able.