An AI System That Uses a Hundredth of the Energy

A team of researchers at Tufts University's School of Engineering has built a proof-of-concept AI system that uses up to a hundred times less energy than a standard neural network while matching or beating it on accuracy. The work, led by Professor Matthias Scheutz, was published this spring and has drawn growing attention as the industry's power consumption climbs into the tens of percent of US electricity demand.

The approach is called neuro-symbolic AI. It combines the pattern recognition of neural networks, which learn from large amounts of data, with the kind of structured logical reasoning that humans use when they walk through a problem step by step. Neither half is new. The interesting move is in how the two are wired together so that the slow, expensive neural component only handles what the lightweight symbolic component cannot.

The benchmark

To stress test the system, the team used the Tower of Hanoi, a classic puzzle whose difficulty scales sharply with the number of disks. On the standard version, the hybrid model reached a 95% success rate. A conventional vision-language-action model, trained on similar data, managed 34%.

The more striking number came on a harder variant the system had never seen during training. The neuro-symbolic system still hit 78%. The conventional models failed every attempt. The full writeup is available on ScienceDaily, with additional coverage at SciTechDaily and Tech Times.

The energy figures were equally lopsided. Training the hybrid model used about 1% of the energy a standard model needs, and running it consumed roughly 5%. The improvement comes from offloading the deterministic, rule-bound part of the task to symbolic logic, where a few operations replace billions of matrix multiplications.

What the paper does not yet show

The work is a proof of concept. The tasks tested, while non-trivial, are far smaller than the training and inference loads of a frontier language model. There is no neuro-symbolic equivalent of GPT-5 sitting in a Tufts basement. What the paper does show is that the assumption baked into much of the current scaling effort, that more parameters and more compute are the only path forward, is not the only available answer.

That matters because the energy debate has stopped being theoretical. AI workloads already account for an estimated ten-plus percent of US electricity consumption, and the major labs have been signing data-centre deals that would have looked excessive even a year ago, including Anthropic's lease of the entire SpaceX Colossus facility. A research direction that promises an order-of-magnitude cut in energy use, even if it takes years to scale, is one that funders and policy makers will read closely.

Whether neuro-symbolic systems can be scaled to the size of frontier language models is the open question. Earlier waves of symbolic AI ran into a problem known as the knowledge bottleneck, where the rules themselves are too expensive to hand-write at scale. The Tufts team argues that modern neural models can do much of that knowledge-extraction automatically, which is what makes the hybrid possible now. If they are right, the next round of AI scaling may look less like a contest over data-centre capacity and more like a contest over algorithmic design.