The research also shows how large the carbon footprint of large language models is, says Lynn Kak, an assistant professor of computer science and public policy at Hertie College Berlin, who was not involved in the Hugging Face research. She says she was surprised to learn the scale of the life-cycle emissions figures, but more work needs to be done to understand the environmental impact of large language patterns in the real world.
“We need to do a better job of understanding the more complex downstream effects of the uses and abuses of AI, which is hard to appreciate much, which is why this aspect is often overlooked.”
For example, advertising tips and algorithms are often used in advertising, which in turn pushes people to buy more things, causing more carbon emissions. It’s also important to understand how to use AI models, says Kak. Many companies, like Google and Meta, use AI models to do things like rank user comments or content recommendations. These actions consume very little energy but can happen a billion times in a day, and this increases the amount of emissions.
Attempts to reduce the carbon footprint of artificial intelligence
The tech sector is estimated to be responsible 1.8% to 3.9% of greenhouse gas emissions. While only a fraction of these emissions come from artificial intelligence and machine learning, the carbon footprint of AI is still very large for a single area within the technology sector.
By better understanding how much energy AI systems use, companies and developers can make choices about the balances they want to strike between pollution and costs, says Lucione.
“The authors of the paper hope that researchers and companies can think about ways to develop large language models to reduce their carbon footprint,” says Sylvain Viger, director of applications at semiconductor company Graphcore, co-author of the Hugging Face article. .
This research may also encourage people to turn to more efficient ways of doing AI research, such as tuning and improving existing models, rather than pushing for larger models, Lucione says.
Read also: 6 ways to prepare carbon neutral plans
The research findings are a “caution,” says David Rolnick, an assistant professor at the McGill University School of Computer Science and the Mela Quebec Institute for Artificial Intelligence, and one of the research co-authors with Kak, who was not involved in Hugging’s research. Face, for the people who use this kind of model, which are often the big tech companies.
“The effects of artificial intelligence are not deterministic, they are the result of the choices we make about how to use these algorithms and which algorithms to use,” he adds.