We use artificial intelligence every day. We may ask chatbots tons of questions, speak to digital assistants, and use translation apps. Because the answers appear so quickly, we usually don't think about where the work happens.
Much of the difficult processing behind today's AI takes place in large data centers, where powerful computers handle our requests and send answers back. They make modern AI possible, but they also use a great deal of electricity, some even use as much power as a small city.
Sam Altman, the CEO of OpenAI, used to say that users saying "please" and "thank you" to ChatGPT has cost the company tens of millions of dollars in electricity.
That raises an important question: can we bring more AI into our pockets and into the devices around us? This is the goal of AI at the edge.
Edge devices are the phones, cars, cameras, robots, and sensors that collect information and interact with the world around us. Today, much of that information is sent elsewhere for processing. If the devices could understand more of that information themselves, they could respond faster, use less energy, and keep more data private.
But bringing AI to small devices is difficult. Phones and sensors have limited space and battery power; they cannot carry a data center inside them.
Our study, published this summer in Nature Communications, explores one way to address this problem. Researchers from UMass Amherst and our collaborators took inspiration from the human brain. In our brains, memory and thinking are not in separate places, they happen in the same tissue, at the same time. Traditional computers keep those two jobs apart, and shuttling information between them burns most of the energy. So we built chips based on a nanoscale device called a ‘memristor’, which lets them remember and calculate in the same spot, much the way brain cells do.
But this solution presents a challenge. Because memristors are physical devices, no two are exactly the same - just like our brains. Those small individual differences create a certain amount of randomness.
Engineers usually try to remove that randomness. But in our work, we went in the other direction. We designed the hardware and the software together, so that the natural randomness of the memristors could help the system.
By treating randomness as a feature rather than a problem, we achieved better performance with less hardware. We tested this approach on a tiny chip, smaller than our fingernails. The chip performed an important AI task: identifying the language of a written passage. Tested with a European language dataset, our system shows over 95 percent recognition accuracy while using about 90 percent less hardware than earlier systems of this kind. These results suggest that useful AI tasks can be performed more efficiently on small, battery-powered devices.
Future versions could help devices recognize speech, understand sounds, monitor health, or respond to their surroundings without constantly sending information to the cloud.
This work brought together researchers at the University of Massachusetts Amherst, the University of Southern California, and industry partner TetraMem.
Moving a lab demonstration toward a reliable and affordable product will require continued work from both universities and industry. But the direction is promising. Intelligence can move closer to us, into the devices we carry, drive, and wear every day. The goal is not simply to make AI bigger. It is to make AI small enough, efficient enough, and useful enough to be everywhere.
Yi Huang is an assistant professor in the Department of Electrical Engineering and Computer Science at the University of Tennessee, Knoxville. His research focuses on brain-inspired computing and energy-efficient AI systems for edge devices. Before joining UT Knoxville, he was a postdoctoral research associate at the University of Massachusetts Amherst where he earned his Ph.D. in 2025. The research featured in this commentary was conducted during Dr. Huang’s doctoral and postdoctoral work.
The views expressed by commentators are solely those of the authors. They do not necessarily reflect the views of this station or its management.