AI That Learns With Light: A 200-Metre Fibre Loop Does the Thinking
Today's AI runs on electricity, and lots of it. Vast numbers of transistors switch on and off, the chips get hot, and the power bill grows. But what if the calculating could be done by light instead?
On 5 November 2026, researchers at the University of Tokyo and the Japanese company NTT published a paper in the journal Nature Computational Science describing exactly that: a new way of building and training AI that uses light to do the work, demonstrated on a real optical system built around a 200-metre loop of fibre.
Why light?
Light is fast, and it can carry many signals at once. That makes it an attractive way to process information — in principle, a beam of light passing through a physical system can do in one step what a conventional chip does with thousands of tiny calculations. Researchers call this "physics-based AI": instead of simulating every step on a chip, you let the natural behaviour of a physical system, such as light travelling through fibre, do the processing for you.
The catch has always been training. Ordinary AI learns by a method called backpropagation, which needs to remember every intermediate result along the way — a poor fit for a physical system, where you cannot easily peek inside and record what happened at each step.
Two simplifications, one system
The team, led by Associate Professor Kohei Nakajima of the University of Tokyo's Laboratory for Intelligent Systems and Informatics, combined two ideas. The first is a new training method called state-skipping direct feedback alignment (SS-DFA), which learns mainly from how wrong the final answer was, instead of tracking every step in between. The second is a simpler network design, called a single-map architecture, in which many units follow the same simple rules rather than relying on a web of complex internal connections.
They tested the combined approach in computer simulations and on a real optoelectronic platform: information was turned into an optical signal, sent through a 200-metre loop of optical fibre, and turned back into an electrical signal for analysis. The simplified design matched, and in some tests outperformed, more complicated neural networks — while needing far less computation. The authors report roughly 40% lower computational resource requirements when the approach is built into modern network designs.
What it is not — yet
The researchers are clear about the limits. They have not tested the approach on today's largest AI models, such as the big language models, and the energy savings are a hoped-for benefit rather than a measured result. Turning a laboratory demonstration into practical hardware will take more work.
Still, the step matters: most optical computing stories so far have stayed in simulation. This one ran on real hardware, with light doing the calculating.
For students, it is a neat reminder that computing is not only about chips and code. Physics itself — a beam of light, a loop of glass fibre — can be the computer. The same physics your child meets in a lesson on light and waves is now being used to rethink how machines learn.
If your child is curious about how computers work — or about light, waves and the physics behind them — a first lesson in science can start with the questions they already ask. Book your first lesson
Sources:
- Tech Xplore, "Let there be light — novel AI framework runs on optical hardware", 5 November 2026: https://techxplore.com/news/2026-11-ai-framework-optical-hardware.html
- Yongbo Zhang, Mitsumasa Nakajima, Katsuma Inoue, Toshikazu Hashimoto, Yasuo Kuniyoshi and Kohei Nakajima, "Scalable physical deep learning using optical dynamics with state-skipping training and single-map architecture", Nature Computational Science, 5 November 2026, DOI: 10.1038/s43588-026-01075-w
London Tutors — post your tutoring request and we will match you with a verified tutor.