Richard Sutton, widely regarded as the father of reinforcement learning, has launched a new lab with an ambitious goal: building a trillion-parameter agent that learns in real time while consuming only 20 watts of power. This represents a dramatic challenge in efficiency and scale, aiming to create powerful AI systems that operate within extreme energy constraints.

The post references Sutton's influential 2019 essay, The Bitter Lesson, which argued that general methods leveraging computation consistently outperform approaches that try to encode human domain knowledge. This new lab appears to be a direct extension of that philosophy — pushing toward massively scaled, computation-driven agents while simultaneously targeting radical energy efficiency, a combination that could reshape how AI systems are designed and deployed.

Researchers at Tufts University, led by professor Matthias Scheutz, have developed a neuro-symbolic AI system that combines traditional neural networks with human-like symbolic reasoning, achieving up to 100x reduction in energy consumption while significantly improving accuracy. The approach targets visual-language-action (VLA) models used in robotics, where conventional systems rely on brute-force trial and error and consume enormous amounts of electricity — AI and data centers already account for over 10% of U.S. electricity production.

In tests using the Tower of Hanoi puzzle, the neuro-symbolic system achieved a 95% success rate compared to just 34% for standard models, and on unfamiliar complex variations it still succeeded 78% of the time while traditional models failed entirely. Training time dropped from over a day and a half to just 34 minutes, and training energy was reduced to just 1% of conventional requirements. During operation, the system used only 5% of the energy needed by standard approaches.

The research, set to be presented at the International Conference of Robotics and Automation in Vienna, suggests that current LLM and VLA-based approaches may not be sustainable long-term. By incorporating structured reasoning and rule-based logic rather than relying solely on statistical pattern matching, neuro-symbolic AI offers a potentially more efficient, accurate, and sustainable foundation for future AI systems — addressing growing concerns about AI's strain on power infrastructure as data center energy demands are projected to double by 2030.