#meatball discourse

[jadeharleyinc]

the scale of AI's ecological footprint

standalone version of my response to the following:

"you need soulless art? [...] why should you get to use all that computing power and electricity to produce some shitty AI art? i don’t actually think you’re entitled to consume those resources." "i think we all deserve nice things. [...] AI art is not a nice thing. it doesn’t meaningfully contribute to us thriving and the cost in terms of energy use [...] is too fucking much. none of us can afford to foot the bill." "go watch some tv show or consume some art that already exists. […] you know what’s more environmentally and economically sustainable […]? museums. galleries. being in nature."

you can run free and open source AI art programs on your personal computer, with no internet connection. this doesn't require much more electricity than running a resource-intensive video game on that same computer. i think it's important to consume less. but if you make these arguments about AI, do you apply them to video games too? do you tell Fortnite players to play board games and go to museums instead?

speaking of museums: if you drive 3 miles total to a museum and back home, you have consumed more energy and created more pollution than generating AI images for 24 hours straight (this comes out to roughly 1400 AI images). "being in nature" also involves at least this much driving, usually. i don't think these are more environmentally-conscious alternatives.

obviously, an AI image model costs energy to train in the first place, but take Stable Diffusion v2 as an example: it took 40,000 to 60,000 kWh to train. let's go with the upper bound. if you assume ~125g of CO2 per kWh, that's ~7.5 tons of CO2. to put this into perspective, a single person driving a single car for 12 months emits 4.6 tons of CO2. meanwhile, for example, the creation of a high-budget movie emits 2840 tons of CO2.

is the carbon cost of a single car being driven for 20 months, or 1/378th of a Marvel movie, worth letting anyone with a mid-end computer, anywhere, run free offline software that consumes a gaming session's worth of electricity to produce hundreds of images? i would say yes. in a heartbeat.

even if you see creating AI images as "less soulful" than consuming Marvel/Fortnite content, it's undeniably "more useful" to humanity as a tool. not to mention this usefulness includes reducing the footprint of creating media. AI is more environment-friendly than human labor on digital creative tasks, since it can get a task done with much less computer usage, doesn't commute to work, and doesn't eat.

and speaking of eating, another comparison: if you made an AI image program generate images non-stop for every second of every day for an entire year, you could offset your carbon footprint by… eating 30% less beef and lamb. not pork. not even meat in general. just beef and lamb.

the tech industry is guilty of plenty of horrendous stuff. but when it comes to the individual impact of AI, saying "i don’t actually think you’re entitled to consume those resources. do you need this? is this making you thrive?" to an individual running an AI program for 45 minutes a day per month is equivalent to questioning whether that person is entitled to a single 3 mile car drive once per month or a single meatball's worth of beef once per month. because all of these have the same CO2 footprint.

so yeah. i agree, i think we should drive less, eat less beef, stream less video, consume less. but i don't think we should tell people "stop using AI programs, just watch a TV show, go to a museum, go hiking, etc", for the same reason i wouldn't tell someone "stop playing video games and play board games instead". i don't think this is a productive angle.

(sources and number-crunching under the cut.)

good general resource: GiovanH's article "Is AI eating all the energy?", which highlights the negligible costs of running an AI program, the moderate costs of creating an AI model, and the actual indefensible energy waste coming from specific companies deploying AI irresponsibly.

CO2 emissions from running AI art programs: a) one AI image takes 3 Wh of electricity. b) one AI image takes 1mn in, for example, Midjourney. c) so if you create 1 AI image per minute for 24 hours straight, or for 45 minutes per day for a month, you've consumed 4.3 kWh. d) using the UK electric grid through 2024 as an example, the production of 1 kWh releases 124g of CO2. therefore the production of 4.3 kWh releases 533g (~0.5 kg) of CO2.

CO2 emissions from driving your car: cars in the EU emit 106.4g of CO2 per km. that's 171.19g for 1 mile, or 513g (~0.5 kg) for 3 miles.

costs of training the Stable Diffusion v2 model: quoting GiovanH's article linked in 1. "Generative models go through the same process of training. The Stable Diffusion v2 model was trained on A100 PCIe 40 GB cards running for a combined 200,000 hours, which is a specialized AI GPU that can pull a maximum of 300 W. 300 W for 200,000 hours gives a total energy consumption of 60,000 kWh. This is a high bound that assumes full usage of every chip for the entire period; SD2’s own carbon emission report indicates it likely used significantly less power than this, and other research has shown it can be done for less." at 124g of CO2 per kWh, this comes out to 7440 kg.

CO2 emissions from red meat: a) carbon footprint of eating plenty of red meat, some red meat, only white meat, no meat, and no animal products the difference between a beef/lamb diet and a no-beef-or-lamb diet comes down to 600 kg of CO2 per year. b) Americans consume 42g of beef per day. this doesn't really account for lamb (egads! my math is ruined!) but that's about 1.2 kg per month or 15 kg per year. that single piece of 42g has a 1.65kg CO2 footprint. so our 3 mile drive/4.3 kWh of AI usage have the same carbon footprint as a 12g piece of beef. roughly the size of a meatball [citation needed].