Smarter AI Cuts Energy Use by 100x, Boosts Performance

Artificial intelligence systems and data centers in the United States consumed approximately 415 terawatt hours of electricity in 2024, a figure projected to double by 2030. This escalating demand raises sustainability concerns, prompting researchers to develop a highly efficient AI proof-of-concept. Their neuro-symbolic AI, presented at the International Conference of Robotics and Automation, merges neural networks with symbolic reasoning for improved performance and drastically reduced energy use.

The team's focus is on visual-language-action (VLA) models for robotics, extending language models with vision and physical action capabilities. Unlike conventional VLA systems that rely heavily on trial-and-error learning, the new approach uses rules and abstract concepts to plan more effectively. This significantly cuts down on training time and energy expenditure during operation.

In tests using the Tower of Hanoi puzzle, the neuro-symbolic VLA achieved a 95% success rate, compared to 34% for standard systems. Its training time dropped to 34 minutes, a fraction of the over 1.5 days required by conventional models. Energy consumption for training was reduced to just 1% and operational energy use to 5% of traditional methods, offering a more sustainable path for AI development.