As Luccioni shows me a digital map that tabulates real-time data on electricity consumption and carbon intensity, she explains it’s her “favorite website.”
Her enthusiasm carries over into her role as AI & Climate Lead at Hugging Face, where Luccioni collects actionable data on AI’s environmental impact.
She has her work cut out for her, especially as companies like Nvidia and Google, which design AI chips, become more secretive about their proprietary offerings.
“The difficulty is that there’s this race to secrecy,” Luccioni said. “Since ChatGPT came on in November 2022, companies have really cracked down on how much information they share about their models.”
Luccioni codeveloped and regularly contributes to CodeCarbon, a program that helps developers estimate emissions and energy use from running AI models. Luccioni says tens of thousands of people have cited the program. She’s most proud of how it allows technologists to act on their concern for the environment and benchmark against actual data.
Meta recently used the tool to estimate emissions from running one of its latest Llama models.
“Where the energy is coming from is really the biggest impact on emissions,” Luccioni said. “The issue is that most supercomputers, be it Google Cloud, Azure, or AWS, are located in places that are not powered by renewable energy, mostly natural gas and coal, and that makes a huge difference.”
One of her latest projects should make it less opaque to understand how much energy and computing power popular AI models and tools use. This work, done with the Organisation for Economic Co-operation and Development, is expected to eventually help establish an energy-efficiency rating standard for AI models. She likens it to an Energy Star rating.
“The EPA adopted this trust-based approach, but in the case of AI it would be a bit harder to adopt that approach because there are so many variables,” she said. “You have to define specific data sets and specific hardware to compare these models.”
As AI’s value soars, possibly into the trillions, Luccioni’s work will help document the tech’s toll on electric grids and, ultimately, the environment.
“You train a model once, but you deploy for a while. Even quantifying the amount of energy per query or per day is really powerful,” Luccioni said. “In the long term, it really adds ups.”
See Business Insider’s full AI Power List
Read the full article here
