CONCEPT
Active Inference
Karl Friston’s account of agency as the complement to perception—where perception updates beliefs to match the world, active inference changes the world to match beliefs, and from this single move curiosity, goal-direction, and the distinction between genuine agents and sophisticated predictors all follow.
Active inference is the step that converts a prediction machine into an agent. In
Karl Friston’s framework, free energy—the divergence between an internal model and sensory evidence—can be minimized in two ways. The passive way is perception: update the model to fit the data. The active way is action: generate behavior that makes the data fit the model. Active inference is the second mode, and it is the mode that makes a system an agent in the full sense: a system that acts on the world to bring about states consistent with its expectations, rather than a system that only updates its beliefs about an unchanging world. From this single move, Friston derives goal-directed behavior (actions that minimize expected free energy), curiosity (actions that reduce model uncertainty—epistemic free energy), and the distinction between exploitation and exploration as two modes of the same underlying drive. The implication for contemporary AI is precise: a
large