Thijs van de Laar - Sustainable AI for Decision-Making - 04th February 2026 скачать в хорошем качестве

Thijs van de Laar - Sustainable AI for Decision-Making - 04th February 2026

In 2024, Statistics Netherlands (CBS) computed that 4.6% of all electricity usage in the Netherlands is due to data centers. With the growing demands of generative AI, this number is projected to grow further; on an already constipated electricity grid. We don't need more data centers, we need more energy-efficient AI. I will argue that the key to energy-efficient AI for decision-making lies in reasoning with beliefs. In classical control a policy is a function of states, while in the developing framework of active inference (AIF) control is a function of beliefs about states. This allows an AIF agent to select controls that inform its beliefs. An AIF agent then actively explores its environment to reduce uncertainty, which leads to improved data-efficiency and reduced computational demand. Probability theory is the principled mathematical framework for reasoning with beliefs. Through probabilistic programming (PP) techniques, computations for AIF control can be automated.