A New AI Approach Uses One Percent of the Energy

A research team at Tufts University has built an AI system that learns to do tasks using one percent of the energy a standard neural network needs to train, and five percent of the energy during operation. The catch, if it is one, is that it performs better than the conventional baseline on a classic planning problem.

The work, led by Matthias Scheutz and his lab, will be presented at the International Conference on Robotics and Automation in Vienna this May. The approach is called neuro-symbolic AI. Rather than training a single large neural network to handle every part of a task, the system combines a neural front end with a symbolic reasoning engine that operates on rules and abstract concepts such as shape, balance and position. The neural part recognises what is in front of it. The symbolic part decides what to do next.

Tested on a logic puzzle

Scheutz's group tested the system on the Tower of Hanoi, a puzzle that requires moving stacked discs between three pegs without ever placing a larger disc on a smaller one. The neuro-symbolic system solved it correctly in 95 percent of trials. A standard vision-language-action (VLA) model, trained end-to-end, managed 34 percent. The standard model, in other words, failed two-thirds of the time at a puzzle a thoughtful seven-year-old can solve.

The energy figures matter because AI workloads are already estimated to draw more than 10 percent of US electricity, and forecasts for the next five years are not reassuring. If a substantial fraction of that demand could be served by hybrid systems that need orders of magnitude less power, the grid implications are enormous. Data-centre operators are currently signing contracts for nuclear capacity that will not come online until 2028 at the earliest. A method that simply uses less might arrive sooner.

The honest caveat

Tower of Hanoi is a narrow benchmark. The result does not yet translate to the kind of open-ended language and visual reasoning that LLMs handle, and symbolic reasoning systems historically have been brittle outside their target domain. As Tufts Now notes, Scheutz's team has shown the approach works for one class of planning problem with the right hand-crafted symbolic rules. Generalising to the messy real world remains an open question.

Still, the direction is worth watching. The dominant assumption of the last five years has been that scale solves everything, that bigger neural networks fed with more data will eventually reason their way to anything humans can do. Results like this suggest the cheaper route may not be the bigger one. We covered a related debate in our piece on AI hallucinations, where OpenAI's own researchers conceded that pure scaling has mathematical limits.