The Honest Climate Case for AI
15 points by darccio
15 points by darccio
"Honest"
The article doesn't state anything negative and it's up to everybody but the companies and their users to do change the behaviour/rules/...
Out of the three positive cases it mentions, two are old or very old and are certainly far from the technology that is currently facing criticism. Even the third seems pretty different but it's at least quite recent.
I'm not saying the content of the article is wrong (computing power usage is dwarved by several other fields) but the article is certainly not "honest".
There is also another indicator that the author might be somewhat biased. The sentence structures follow a very typical repetitive phrasing and overal very formulaic approach of story telling. Of course, it might just be their writing style, or just heavily LLM assisted. But I do get the clear sense that the author is just somewhat biased.
Besides all of that. It is just an overall a very bland, middle of the road, not very inspiring article regardless of who of what the actual author is.
A mix of both: writing style and LLM assistance. And a bit of structure from a storytelling deck I like to use.
All the statements are based on actual reports, that's where the honesty lies.
Also, I think that it's clearly negative that agentic workloads are compounding the consumption. Or that the demand is outrunning the efficiency. Or, even, there is a significant increase on non-clean energy feeded to new datacenters.
These are the three sections before the positive cases, which are just a sample.
Stop feeling guilty about prompts. Your Wh per query is not the lever that matters. You'll do more climate good by eating one less steak, taking one fewer flight, or voting for better energy policy than by boycotting LLMs.
First off, these things aren't mutually exclusive, so there's no reason you couldn't do all of them.
But I also take issue with this view where (non-)usage of LLMs supposedly doesn't matter. By rejecting LLMs we can signal to the industry that maybe they don't need to build so many new data centers. I don't think you should treat the industry's trajectory as an inevitability.
In that light, I think focusing solely on the inference costs is misleading. The costs of training and data center construction should be factored in. Although, to be honest, I don't even agree that the inference costs on their own are negligible: according to the post, newer models may require 10 to 100 times more energy and agentic flows can trigger an unbounded number of requests.
Any industry that uses about 1.5% of the world's electricity, like air conditioning or industrial motors, shouldn't be a cause for concern.
Is that supposed to sound like a small amount of energy for an industry in its infancy?
By rejecting LLMs we can signal to the industry that maybe they don't need to build so many new data centers.
They were built and nobody thought they needed a coding agent before these existed.
Is that supposed to sound like a small amount of energy for an industry in its infancy?
No, the three sections below disarm that idea. The "shoulnd't be" is the key also.
I'd also argue that we're basically in the mainframe era of this tech. We've seen this story many times before where new technology starts out needing big data centres to operate, then over time people learn how to optimize it and it moves to edge devices. I don't see any reason why this tech should be any different. We're still in the very early days and it's silly to assume that energy requirements are going to stay roughly constant going forward. We've already seen a huge improvement in efficiency where models you can run on your laptop outperform frontier models that needed a whole data centre just a few years ago. That's the trend that's important to keep in mind when projecting what we can expect in the future.
The other argument of course is that as this tech becomes more efficient than demand will grow negating the efficiency gains. But that's just a question of whether people find this tech genuinely useful or not. If usage grows that implies that people are finding reasons to use it because it solves some problem for them. It's not different from any other technology in that regard.
Could be, there certainly is a pattern for a lot of technologies but it doesn't have to be a given. I mean, last time I checked I still couldn't get my hands on an actually working crystal ball.
I mean we're already running models on laptops today. It's not a hypothetical at this point.
Are the costs of LLM training documented somewhere?
I haven't seen citations for that cost, though I'd like to.
I think people overlook how AI impacts the demand for rare earth metals used in chip manufacturing. As demand for datacenter chips grows, there's a pretty environmental big impact caused by scaling up mining.
This is an interesting read and and a welcome addition to the discussion of LLM and datacenter usage.
And generally speaking I don't think it helps when people are being shamed for inquiring or using LLMs. But I'm also quite pessimistic about how these laws and foundations for efficiency will be laid out, because in my current climate almost everything seems to revolve around short term gains. And I don't know where to start fighting that.
The IEA predicting that "widespread adoption" of AI will decrease meet emissions is, um, surprising to say the least.
(Their examples look more ML/constraint solver shaped than LLM shaped, though)
People always want to be upset about new things, because they seem easier to change, thank about impactful things, which seem hard.
Since I bike everywhere and don't own an ICE car and don't fly for vacations etc my carbon footprint is way lower than average no matter what I do with computers.
If I can cut all beef (working on it) it'll be even lower.
But most people like cars and vacations and beef a lot.
Okay. I also don't eat meat, at all. Nor do I fly. Am I allowed to be upset about a technology we all got along fine without three years ago creating heat islands and driving increased emissions through higher base load requirements? Do I pass your moral purity test?
Not yet, you also gotta forsake HVAC and get all your power by running in a really big hamster wheel