Energy · Breakthrough
Ivana Drobnjak set out to prove AI doesn't have to waste this much energy, and did
Ivana Drobnjak, Maria Perez Ortiz, Hristijan Bosilkovski, Leona Verdadero · London, United Kingdom
A UCL professor and her research team measured, rather than assumed, how much power generative AI wastes, and found simple fixes that cut it by up to 90 percent without hurting accuracy.
The person and the place
Ivana Drobnjak is a professor at UCL's Computer Science department and a member of the UNESCO Chair in AI, in London. She worked with Maria Perez Ortiz, also at UCL, Hristijan Bosilkovski, a recent UCL graduate, and Leona Verdadero at UNESCO.
The problem they cared about
The usual answer to AI's energy problem is more data centers and more power plants. Nobody was asking a simpler question first: how much of that energy is being wasted before it ever reaches a user.
The agency moment
So Drobnjak's team stopped assuming and started measuring. They ran a set of original experiments on a real, working model, testing exactly how much energy ordinary engineering choices actually cost.
What changed, with evidence
Rounding the model's internal calculations to fewer decimal places cut energy use by up to 44%, while keeping accuracy at 97% or better. Swapping a general-purpose model for one built for a single task cut energy by 15 to 50 times, depending on the job. Cutting response length in half saved another 54%. Stacked together, the changes cut energy use by 90% for repetitive tasks and 75% for complex ones. UCL's own estimate: one day of those savings could power 30,000 to 34,000 UK households.
The move worth copying
Most people can't rerun a university's energy experiments. But the underlying habit travels: before reaching for the biggest, most general model for a job done the same way every time, ask whether something smaller, measured against that exact job, would do it for a fraction of the power. Know a researcher measuring what everyone else just assumes? Feed the scout at goodinprogress.org.
"Our research shows that there are relatively simple steps we can take to drastically reduce the energy and resource demands of generative AI, without sacrificing accuracy and without inventing entirely new solutions." — Ivana Drobnjak, UCL News
Sources
Here is where this story came from. Go read the originals.
Curated by Good in Progress from the public record.
Sources last verified .
