A New Approach to AI Could Use 100x Less Energy

A team of researchers at Tufts University says it has built an AI system that matches the performance of conventional deep-learning models while using a small fraction of the energy and training data, according to ScienceDaily. The work, led by Karol Family Applied Technology Professor Matthias Scheutz, will be presented at the International Conference on Robotics and Automation in Vienna in May.

The approach is known as neuro-symbolic AI. Instead of leaning entirely on pattern recognition the way most modern systems do, it pairs neural networks with explicit symbolic reasoning, the kind of step-by-step rule-following that older AI systems were built around. The two halves cover for each other's weaknesses: the neural network handles messy, ambiguous inputs, while the symbolic layer keeps the reasoning structured and inspectable.

What the benchmarks show

In tests using a robot solving variants of the Tower of Hanoi puzzle, the neuro-symbolic system reached a 95 percent success rate. A standard vision-language-action model managed 34 percent. Training the hybrid took 34 minutes; the conventional model needed roughly a day and a half to reach a comparable level. According to the paper, the neuro-symbolic system used about 1 percent of the energy required to train the standard model, and roughly 5 percent of the energy during execution.

If those numbers hold up at larger scales, the implications are substantial. The International Energy Agency estimates that AI systems and data centres consumed around 415 terawatt-hours of electricity in 2024, and the demand curve has only steepened since. Frontier model training runs are now routinely measured in megawatts, with the largest deployments reshaping regional grids. Approaches that genuinely cut compute requirements, rather than simply finding faster chips, are some of the more interesting research bets in the field right now.

Caveats worth keeping in mind

The Tower of Hanoi is a bounded puzzle with clear rules, and neuro-symbolic systems have historically been difficult to scale to the open-ended tasks that today's large language models handle. Whether the same gains carry over to long-horizon planning, natural language work, or robotics in unstructured environments is still an open question. The Tufts team's results are best read as an existence proof: the trade-off between accuracy and energy is not as fixed as recent trends might suggest.

Symbolic methods fell out of fashion during the deep-learning boom, and the pendulum has been firmly on the neural side for the better part of a decade. Several research groups have argued for some time that the next jump in capability will come from putting the two approaches back together. Results like these are part of why that argument is being heard again.