Artificial intelligence keeps getting more capable, but also more power-hungry. Training large models consumes vast amounts of electricity, so researchers are looking for ways to make AI more efficient rather than just bigger. Researchers have now demonstrated one promising route: a physics-based AI system that computes with light. Physics-based AI uses the natural behavior of a physical system, rather than conventional chips, to process information.
Ordinary chips shuttle electrical signals through huge numbers of transistors, which uses energy and produces heat. Light can carry information extremely fast and can carry many signals at once, so optical computing is an attractive platform for AI. The team's demonstration system converts information into an optical signal, processes it as the light travels through a 200-meter loop of optical fiber, then converts it back into an electrical signal, using only a small number of electronic components.
Learning without looking inside
The advance is not only the hardware but also how the system is trained. Conventional training requires measuring and recording the system's internal states as it evolves, then backtracking to work out what to change. The team's method, called SS-DFA (state-skipping direct feedback alignment), skips most of those intermediate states and uses the final error to guide learning, cutting measurement, data storage and computation.
According to the study, published in Nature Computational Science, the simplified design matched and sometimes outperformed more complex neural networks while needing far less computation. When inserted as a building block into several modern neural network architectures, it improved performance while cutting computational resources. The approach also appears to challenge the scaling laws that describe how conventional networks improve with size.
This matters because the energy cost of AI is one of the field's biggest concerns. Optical systems generate little heat, so if the approach scales up it could lower the energy needed for computing and cooling in data centers, and even reduce water consumption.
The researchers caution that they have not yet tested the framework on today's largest language and multimodal models. More work on miniaturization, stability and manufacturing is needed before it could fit on a chip inside phones, cameras or self-driving cars.