Jump to content

Talk:Environmental impacts of artificial intelligence

Page contents not supported in other languages.
From Wikipedia, the free encyclopedia

Did you know nomination[edit]

(
)
  • ... that training the model for ChatGPT used the equivalent energy footprint of driving 123 gasoline-powered passenger vehicles for a year?
Created by Bluethricecreamman (talk). Number of QPQs required: 0. Nominator has less than 5 past nominations.

Bluethricecreamman (talk) 03:58, 3 July 2024 (UTC).[reply]

General: Article is new enough and long enough
Policy: Article is sourced, neutral, and free of copyright problems
Hook: Hook has been verified by provided inline citation
QPQ: None required.

Overall: awkwafaba (📥) 16:04, 10 July 2024 (UTC)[reply]
@Bluethricecreamman: Thank you for your hard work. Please fix copyvios and sourcing issues and then we can proceed.

Ok, i've updated the template with the corrected source for the first one (I guess I mixed up the two scientific american articles, I think they bothy say that GPT-3 releases 552 metric tons of Co2, but only one talks about the comparison to cars. Thanks for catching that! Also fixed ALT1, thanks for fixing that as well! Apologies for incorrect info, I'll try to read a bit slower with these articles. Bluethricecreamman (talk) 21:45, 10 July 2024 (UTC)[reply]
I didn't dig to hard on that spammy site earlier, sorry. I agree that they stole from enwiki and not the other way around. All three hooks are good to go. I think ALT0 is the most compelling. awkwafaba (📥) 01:05, 11 July 2024 (UTC)[reply]
Specifically, I am a little concerned that statistics-heavy topics like this, when they're directly related to buzzwordy trending topics in the news cycle, tend to be very ill-suited to being patched out from headlines. That is to say, writing the shark attack and dog attack based solely on news coverage would give the impression that these were both uncommon things that do happen every once in a while, when in reality the yearly rates are eighty versus several million. This article quotes a bunch of figures from news stories back-to-back, and there doesn't really seem to be much attempt at comparison or context. For example, there is a paper cited that projects 85-143 TWh of global power consumption from neural networks by 2027, based on a conjectural optimistic scenario where NVIDIA/TSMC transition their entire manufacturing output to A100s (all of which are installed in data centers and used for nothing but LLM inference). By comparison, this paper estimated that video gaming consumed 34 TWh/year in the United States alone in 2019; this document from the DoE says that aluminum production in the US in 2000 consumed a total of 279.2 TWh/yr. The US consumes about 16% of the world's electricity, so a very rough approximation here would give a ballpark estimate of 212.5 TWh for gaming (the US does about 2% of aluminum production worldwide, giving us around 13960 TWh yearly for global aluminum production). It also seems like there is a lot of ambiguity between training (a fixed cost, as models are only trained once) and inference (the process that happens when running the model, which typically requires billions or trillions of times less computation). I am somewhat worried about having this on the front page. jp×g🗯️ 23:45, 12 July 2024 (UTC)[reply]
you are literally using us energy gov stats to WP:SYNTH a whataboutism argument… by the same logic, the criticism section of any industries environmental impact can be called misleading because theres always another more polluting industry you can point to. the comparison you suggest is “missing” is literally that training ChatGPT training took as much energy as 123 gasoline powered cars yearly footprints. and suggesting that including examples from news articles headlining the carbon footprint is a bad thing for wikipedia is inane. most articles on here use secondary souces such as news articles. and most DyK hooks are eyecatching Bluethricecreamman (talk) 00:34, 13 July 2024 (UTC)[reply]
For the specific task of literally writing an encyclopedia article about a topic, I would say that answering questions like "why is this statistic relevant" or "what does this number mean" is probably its most important purpose. I gave a couple examples to illustrate what this looks like, obviously there is no reason why aluminum manufacturing in particular is more relevant than anything else. The most appropriate source here would likely be something like a textbook or monograph about electrical consumption by various industries, and how this related to environmental concerns in general. Since this is a subject of rather large significance, it is important that our writing on it be accurate, and not news. I am opposed to running stuff on the main page where vague insinuations are made by factoids of unexplained significance -- e.g. what's a terawatt-hour? what's a gram of CO2? why are some figures given in one and some as the other, and still others given in folksy derivative units like average midsize sedan gasoline consumption rates over average suburban commutes? jp×g🗯️ 01:06, 13 July 2024 (UTC)[reply]
Moreover, ALT0 is not supported by the source. Indeed, OpenAI's well-known sleazy refusal to be transparent about any of its operations or research means that the source cannot say that, because the source does not know. The figure claimed as fact in this hook is, in the source, carefully and explicitly presented as an approximate figure derived by estimation (it's from https://arxiv.org/abs/2104.10350 -- note the methods they are using). jp×g🗯️ 01:16, 13 July 2024 (UTC)[reply]
I'm not arguing these bad faith arguments after this. The source link is literally right there for anyone to read. There are three links I cite with the 552 metric tons figure in the article. Most of the "environmental impacts of" are similar collections of "factoids" discussing various industry leaders. And by your logic of WP:NOTNEWS, we should remove most news articles discussing long term trends in any industry. WP:NOTNEWS specifically states no original reporting on wikipedia (we don't report the news ourselves, we cite it), and calls for enduring notability of information (we don't do an article if its just one or two articles). if you find a reputable source arguing that this is a false number or a false comparison, feel free to edit the article. I respect your work as an admin, and your experience, but this is silly. Bluethricecreamman (talk) 01:35, 13 July 2024 (UTC)[reply]


Feedback from New Page Review process[edit]

I left the following feedback for the creator/future reviewers while reviewing this article: Nice article!!

Broc (talk) 11:16, 3 July 2024 (UTC)[reply]

ENERGY usage could be a small background section, but it should be clearly separated from the CO2 discussions in a separate section.[edit]

These are clearly different topics, but the large energy usage could be a short background for the CO2 part. What I removed maybe could go in that separate section.

For example, AI in Iceland might have practically no "environmental impact" other than the building of new geothermal plants. ---Avatar317(talk) 23:00, 3 July 2024 (UTC)[reply]

Agreed. we can have a section about how AI methods may have differing footprints based on whether they are run in data centers close to renewable energy sources. Bluethricecreamman (talk) 05:25, 4 July 2024 (UTC)[reply]

Another possibility is to rename the article to "Energy usage and resulting environmental impact from AI" since I haven't heard of ANY other negative environmental impacts discussed other than energy usage. Are there any? ---Avatar317(talk) 23:04, 3 July 2024 (UTC)[reply]

I've added in a water usage section. I suspect there is e-waste involved with the chips used for AI, but can't find anything yet... I'm assuming GPU chips are still new enough that e-waste isn't a huge issue.
Main issue with making this article is we don't have a real environmental impact of computing article, just a redirect to green computing. This deserves its own article with all the sourcing, but it also should have context of energy/water usage/ewaste of data centers/ Bluethricecreamman (talk) 05:27, 4 July 2024 (UTC)[reply]
Retrieved from "https://en.wikipedia.org/w/index.php?title=Talk:Environmental_impacts_of_artificial_intelligence&oldid=1233725510"