Artificial intelligence has tremendous potential to improve efficiency and help decarbonize the planet — but it also requires huge amounts of energy that could exhaust power grids across the U.S.
That dichotomy creates a dilemma for companies tapping into the power of AI technologies for climate and clean tech innovation. The issue was a hot topic of discussion last week in Seattle at two major climate events: PNW Climate Week and the Bloomberg Green Festival.
Training AI models and using AI-powered machines require tons of energy. But the amount is determined by the breadth of what they are designed to do and how they accomplish their task.
Not every problem needs a complex AI solution, said Jai Jaisimha, founder of Transparency Coalition.ai and an affiliate professor in the University of Washington’s Department of Electrical and Computer Engineering.
“I do have some substantive concerns about trying to swat flies with laser beams,” Jaisimha said. “AI can be great, but it should be built for the task that you intend to apply it to.”
Rebecca Hu is CEO of Glacier, a San Francisco-based recycling startup. The company is building AI-trained robots that are positioned above recycling facility conveyor belts. The robots use cameras to watch what passes by on the belt, and have an arm to retrieve and sort recyclables.
Hu said “right-sizing” a model can prevent the drain on resources caused by over-engineering. Every AI company, she said, should consider how to deliver function in a responsible and efficient fashion.
Some companies are using AI for products that use less energy than their predecessors.
Hansi Singh, the Seattle-based CEO of Planette AI, said her company’s model combines physics with AI to forecast temperature, wind and rain stretching up to five years in the future. Singh said Planette’s approach uses a fraction of the energy consumed by the traditional method: large-scale, Earth-systems models that run on energy-hogging supercomputers.
And the less power-hungry strategy also delivers better results, according to Singh.
Planette launched its first product, an El Niño forecaster called Umi, last week. Singh said their models can predict temperature with 80-95% accuracy three to six months in advance, making them more accurate than NOAA forecasts.
One of the challenges in curbing AI-driven energy use is that it is difficult to quantify how much power different AI models use, said Jaisimha. The energy needs vary depending on the task being performed and the way they work is often not transparent, he added.
In the future, Jaisimha said, AI models that are verified as energy efficient could have a market advantage.
