Climate impact studies of artificial intelligence (AI) often assess the weight of their design – an electricity-intensive “training” – rather than the weight of their use. This is precisely the ambition of a study published on November 28, 2023 to determine the carbon footprint of AI once it leaves the laboratories.
Researchers from Hugging Face and Carnegie Mellon University (Pennsylvania) measured the energy consumption of eighty-eight AI neural learning models, divided into different families, by subjecting each group to the same tests, on the same computer. An ambitious work that “doesn’t seem to have been done before”judge Anne-Laure Ligozat, professor at the National Higher School of Informatics for Industry and Business (Ensiie), interviewed by The world.
As for the results, we shouldn’t hope for a ranking of much-discussed generative AI in order of greediness – the editors of ChatGPT and other Midjourneys do not make their data available to researchers. The research therefore distills more subtle lessons, of course about generative AI, but also about AI that works in the shadows and uses, for example, anti-spam, anti-pornographic image filters or automatic subtitle generators.
The effectiveness of ‘old-fashioned’ AI
The researchers’ main observation is the following: for simple tasks, specialized AIs are much more effective than generative AIs, which are versatile and capable of performing various tasks such as translation, summarization, search, etc. “They are about thirty times less energy intensive”, specifies Sasha Luccioni, co-author of the study, who recommends the use of AI for general purposes only in the absence of any other choice. This is confirmed by Maud Picq, director of ‘data science’ at Capgemini Invent: “It is important to choose the best technical approach depending on the context. »
The question arises for many use cases, including Internet searching, M continuesme Luccioni. “Do AIs really need to formulate their answers in three paragraphs of text? Can’t they just underline a few sentences that answer the question in a text available on the Internet? In any case, this would use much less electricity. » The researcher draws a global question from this: without even mentioning the tendency of generative AI to come up with factually incorrect answers, is it really necessary to “put generative AI everywhere” ?