On a median day, I estimate I consume 1,300 Wh through Claude Code—4,400 “typical queries” worth. (For those wondering, this is $15-20 of token spend on a typical day.)
[...]
So, if I wanted to analogize the energy usage of my use of coding agents, it’s something like running the dishwasher an extra time each day, keeping an extra refrigerator, or skipping one drive to the grocery store in favor of biking there. To me, this is very different than, in Benjamin Todd’s words, “a terrible reason to avoid” this level of AI use. These are the sorts of things that would make me think twice.
Just highlighting a commenter on that blog post, who raised a similar point as some other folks: the cost of inference is low, but it ignores the cost of training, which is dramatically higher. If...
Just highlighting a commenter on that blog post, who raised a similar point as some other folks: the cost of inference is low, but it ignores the cost of training, which is dramatically higher.
If Anthropic at any point in time uses 1 GW and inference makes up 8% of that (see Infrastructure-side gut-check), should you not scale your own consumption by a factor 12.5 to arrive at a true run rate which accounts for the research and training time which enables the inference in the first place? 10 extra fridges?
If we stopped training new models tomorrow, this argument holds water (2ml or so :3). Seems shaky though given that we're still amortizing the costs of our current models, while developing new ones?
As a consumer I don't have direct control over that, though. One possibility is to use an older model or a cheaper model instead of the latest and greatest? I'm mostly using Sonnet instead of Opus...
As a consumer I don't have direct control over that, though. One possibility is to use an older model or a cheaper model instead of the latest and greatest?
I'm mostly using Sonnet instead of Opus because Sonnet seems good enough.
Oh certainly! I'm not calling you or anyone else (or myself) out for their individual actions; apologies if I came across as doing so. It is what it is, ultimately. My critique was thrust at the...
Oh certainly! I'm not calling you or anyone else (or myself) out for their individual actions; apologies if I came across as doing so. It is what it is, ultimately. My critique was thrust at the author's argument around the impact of a query sent to e.g. Anthropic, since it isn't taking the full cost into account (i.e. including the training cost), only a lower number.
I don't think the author goes into it, but I'm also not a fan of the "carbon footprint" metric being used to criticize an individual's actions. Ultimately, the bulk of the impact happens outside of an individual consumer's direct control, so trying to pin all the blame on them has always felt like a farce (doesn't stop me from eating vegan and riding a bike, though).
Interesting read! For me it's interesting when I see people criticizing the electricity use of AI citing environmental concerns, while conveniently forgetting to criticize the fact that it's only...
Interesting read!
For me it's interesting when I see people criticizing the electricity use of AI citing environmental concerns, while conveniently forgetting to criticize the fact that it's only an environmental concern because we're still reliant on fossil fuels and an outdated power grid that doesn't prioritize energy storage and energy sharing, at least here in America.
In fact, come to think of it, most popular criticisms of AI would be completely invalidated if you removed late stage capitalism from the equation.
Energy is never free. It doesn't matter if you're getting it from solar, hydro, wind, or coal. The demand you're placing on the grid requires more solar panels to be manufactured, more wind...
Energy is never free. It doesn't matter if you're getting it from solar, hydro, wind, or coal. The demand you're placing on the grid requires more solar panels to be manufactured, more wind turbines to be built, and more dams to be erected; all of which have an environmental impact.
Renewables have far less of an impact than fossil fuels do, but manufacturing a solar panel is still a very resource intensive process that has a negative environmental impact. If we're spending a ton of energy without much benefit, we're still destroying the environment for no reason.
What? Ai is a product of late stage capitalism... you're talking about a product built to hoover up incredible amounts of personal data, while relying on stolen work of others, funnelling all of...
What? Ai is a product of late stage capitalism... you're talking about a product built to hoover up incredible amounts of personal data, while relying on stolen work of others, funnelling all of it to the benefit of a corporate entity with direct ties to an imperialist military regime. This thing was funded with the ill gotten gains of investors and continues to drag down the economy in a variety of ways. If you removed late stage capitalism from ai, ai (in its current form) would not exist.
The companies wouldn’t exist, but the tech would - or something as close as you can reasonably assume in this kind of counterfactual. These are models that come from adapting scientific work, by...
The companies wouldn’t exist, but the tech would - or something as close as you can reasonably assume in this kind of counterfactual. These are models that come from adapting scientific work, by researchers who said “making computers handle natural language is incredibly useful for data parsing and machine translation” or “the statistical properties there are fascinating, I wonder how far we can push that?” or the old reliable “holy shit, I can make software talk, that’s goddamn cool”.
In being justifiably pissed off at the major companies behind most of the negative impact, you’re actually giving them far too much credit. They don’t own “AI”, they aren’t by any means an essential part of the technical ecosystem, and allowing your dislike for them to spill over onto the concept of an LLM or an image model or whatever is kinda buying into that very capitalist idea of the corporation being synonymous with the invention - rather than the corporation packaging and monetising the last few percent after university researchers and OSS developers and hyperfocused nerds spent decades laying the foundations. The major tech companies do have a lot of incredibly skilled people working for them too, and I absolutely don’t want to undermine that; if anything I’m saying the opposite, that a lot of those same people would still be doing equally good or better work without the profit motive.
The timeline might have ended up a few years differently, the burn would likely have been slower, and the hardware might not have been such a focus without eleventy zillion dollars pouring into NVIDIA. But we’d still have the tech - with all the possibilities and potential consequences that brings - without anything like as many misaligned incentives, and likely with a tiny fraction of the likely harm and resulting backlash because of that.
I mean, you're basically making my point for me. The worst parts of AI right now, the data harvesting, labor replacement, and stolen artwork, are all tied to the system it's being built under. If...
I mean, you're basically making my point for me. The worst parts of AI right now, the data harvesting, labor replacement, and stolen artwork, are all tied to the system it's being built under.
If people weren't forced to tie their survival to jobs, and artists weren't forced to monetize every piece of artwork as income just to live, this would be a very different conversation. Even calling it "theft" depends on a capitalist view of art as property and livelihood. And when people criticize AI in those terms, they often end up reinforcing the same capitalist assumptions they say they oppose.
To use a fictional comparison, Star Trek’s holodeck works in a very similar way. The holodeck computer in Star Trek also generates things from records of existing people, harvested data, artwork, published stories, historical data, and reference material, but nobody treats that as some inherent moral crime or art theft, because it exists in a post-scarcity world.
And I'm not trying to defend AI as it exists right now, I'm just saying that in the moments a lot of people criticize AI they tend to accidentally validate capitalistic power structures. And that seems interesting/odd to me.
(I agree with your point, but imo the Star Trek comparison isn't terribly compelling ... no one has concerns about the holodeck in fiction because none of the writers wanted to deal with that...
To use a fictional comparison, Star Trek’s holodeck works in a very similar way. The holodeck computer in Star Trek also generates things from records of existing people, harvested data, artwork, published stories, historical data, and reference material, but nobody treats that as some inherent moral crime or art theft, because it exists in a post-scarcity world.
(I agree with your point, but imo the Star Trek comparison isn't terribly compelling ... no one has concerns about the holodeck in fiction because none of the writers wanted to deal with that quagmire. See also the ethics -- and lack of objection to -- simulating people from your day-to-day life 🤢)
Fwiw, as a person who criticizes both AI data centre power consumption and power grid issues (but isn't American), the concern is that it's using such an incredibly large amount of power to...
For me it's interesting when I see people criticizing the electricity use of AI citing environmental concerns, while conveniently forgetting to criticize the fact that it's only an environmental concern because we're still reliant on fossil fuels and an outdated power grid that doesn't prioritize energy storage and energy sharing, at least here in America.
Fwiw, as a person who criticizes both AI data centre power consumption and power grid issues (but isn't American), the concern is that it's using such an incredibly large amount of power to perform tasks that most do not see as valuable (or at least, insufficiently valuable vs. the costs). Skimming this article, it looks like some projections claim that data centres will consume ~6.7-12% of the total yearly US energy production. That growth has been causing household electricity rates to increase, which is straining an already strained population.
So ... IMO at least, my criticisms of AI would not be invalidated if you removed late stage capitalism from the equation. Power generation has always taken a long time to come online, since it's expensive to scale up, and typically grows slowly enough to be predictable -- the current growth of AI data centres bucks that trend. ROI on that power generation + data centre buildout is poor, too: assuming one interprets the METR and those MIT researchers results as meaning that traditional work and business practices not being amenable to AI/LLM usage in their current forms (which goes both ways), we'd need reforms across all sectors, which is only loosely related to late stage capitalism.
From the article:
[...]
Just highlighting a commenter on that blog post, who raised a similar point as some other folks: the cost of inference is low, but it ignores the cost of training, which is dramatically higher.
If we stopped training new models tomorrow, this argument holds water (2ml or so :3). Seems shaky though given that we're still amortizing the costs of our current models, while developing new ones?
As a consumer I don't have direct control over that, though. One possibility is to use an older model or a cheaper model instead of the latest and greatest?
I'm mostly using Sonnet instead of Opus because Sonnet seems good enough.
Oh certainly! I'm not calling you or anyone else (or myself) out for their individual actions; apologies if I came across as doing so. It is what it is, ultimately. My critique was thrust at the author's argument around the impact of a query sent to e.g. Anthropic, since it isn't taking the full cost into account (i.e. including the training cost), only a lower number.
I don't think the author goes into it, but I'm also not a fan of the "carbon footprint" metric being used to criticize an individual's actions. Ultimately, the bulk of the impact happens outside of an individual consumer's direct control, so trying to pin all the blame on them has always felt like a farce (doesn't stop me from eating vegan and riding a bike, though).
Interesting read!
For me it's interesting when I see people criticizing the electricity use of AI citing environmental concerns, while conveniently forgetting to criticize the fact that it's only an environmental concern because we're still reliant on fossil fuels and an outdated power grid that doesn't prioritize energy storage and energy sharing, at least here in America.
In fact, come to think of it, most popular criticisms of AI would be completely invalidated if you removed late stage capitalism from the equation.
Energy is never free. It doesn't matter if you're getting it from solar, hydro, wind, or coal. The demand you're placing on the grid requires more solar panels to be manufactured, more wind turbines to be built, and more dams to be erected; all of which have an environmental impact.
Renewables have far less of an impact than fossil fuels do, but manufacturing a solar panel is still a very resource intensive process that has a negative environmental impact. If we're spending a ton of energy without much benefit, we're still destroying the environment for no reason.
What? Ai is a product of late stage capitalism... you're talking about a product built to hoover up incredible amounts of personal data, while relying on stolen work of others, funnelling all of it to the benefit of a corporate entity with direct ties to an imperialist military regime. This thing was funded with the ill gotten gains of investors and continues to drag down the economy in a variety of ways. If you removed late stage capitalism from ai, ai (in its current form) would not exist.
The companies wouldn’t exist, but the tech would - or something as close as you can reasonably assume in this kind of counterfactual. These are models that come from adapting scientific work, by researchers who said “making computers handle natural language is incredibly useful for data parsing and machine translation” or “the statistical properties there are fascinating, I wonder how far we can push that?” or the old reliable “holy shit, I can make software talk, that’s goddamn cool”.
In being justifiably pissed off at the major companies behind most of the negative impact, you’re actually giving them far too much credit. They don’t own “AI”, they aren’t by any means an essential part of the technical ecosystem, and allowing your dislike for them to spill over onto the concept of an LLM or an image model or whatever is kinda buying into that very capitalist idea of the corporation being synonymous with the invention - rather than the corporation packaging and monetising the last few percent after university researchers and OSS developers and hyperfocused nerds spent decades laying the foundations. The major tech companies do have a lot of incredibly skilled people working for them too, and I absolutely don’t want to undermine that; if anything I’m saying the opposite, that a lot of those same people would still be doing equally good or better work without the profit motive.
The timeline might have ended up a few years differently, the burn would likely have been slower, and the hardware might not have been such a focus without eleventy zillion dollars pouring into NVIDIA. But we’d still have the tech - with all the possibilities and potential consequences that brings - without anything like as many misaligned incentives, and likely with a tiny fraction of the likely harm and resulting backlash because of that.
I mean, you're basically making my point for me. The worst parts of AI right now, the data harvesting, labor replacement, and stolen artwork, are all tied to the system it's being built under.
If people weren't forced to tie their survival to jobs, and artists weren't forced to monetize every piece of artwork as income just to live, this would be a very different conversation. Even calling it "theft" depends on a capitalist view of art as property and livelihood. And when people criticize AI in those terms, they often end up reinforcing the same capitalist assumptions they say they oppose.
To use a fictional comparison, Star Trek’s holodeck works in a very similar way. The holodeck computer in Star Trek also generates things from records of existing people, harvested data, artwork, published stories, historical data, and reference material, but nobody treats that as some inherent moral crime or art theft, because it exists in a post-scarcity world.
And I'm not trying to defend AI as it exists right now, I'm just saying that in the moments a lot of people criticize AI they tend to accidentally validate capitalistic power structures. And that seems interesting/odd to me.
(I agree with your point, but imo the Star Trek comparison isn't terribly compelling ... no one has concerns about the holodeck in fiction because none of the writers wanted to deal with that quagmire. See also the ethics -- and lack of objection to -- simulating people from your day-to-day life 🤢)
Fwiw, as a person who criticizes both AI data centre power consumption and power grid issues (but isn't American), the concern is that it's using such an incredibly large amount of power to perform tasks that most do not see as valuable (or at least, insufficiently valuable vs. the costs). Skimming this article, it looks like some projections claim that data centres will consume ~6.7-12% of the total yearly US energy production. That growth has been causing household electricity rates to increase, which is straining an already strained population.
So ... IMO at least, my criticisms of AI would not be invalidated if you removed late stage capitalism from the equation. Power generation has always taken a long time to come online, since it's expensive to scale up, and typically grows slowly enough to be predictable -- the current growth of AI data centres bucks that trend. ROI on that power generation + data centre buildout is poor, too: assuming one interprets the METR and those MIT researchers results as meaning that traditional work and business practices not being amenable to AI/LLM usage in their current forms (which goes both ways), we'd need reforms across all sectors, which is only loosely related to late stage capitalism.