The field of artificial intelligence is moving towards a more unified and principled approach, with a focus on formal frameworks and thermodynamic constraints. Recent developments have highlighted the importance of balancing exploitation and exploration in decision-making under uncertainty, and the need for scalable and energy-efficient solutions. The concept of active inference is being refined and unified with variational inference, enabling more efficient and effective decision-making. Additionally, the intersection of information theory and thermodynamics is providing new insights into the fundamental limits of information processing and the design of next-generation AI architectures. Notable papers in this area include:

• Active Inference is a Subtype of Variational Inference, which presents a novel message-passing scheme for scalable active inference.
• Information Physics of Intelligence: Unifying Logical Depth and Entropy under Thermodynamic Constraints, which introduces a novel metric, Derivation Entropy, and demonstrates the existence of a critical phase transition point in information processing.