The free energy principle is a theory, developed primarily by the neuroscientist Karl Friston, that proposes a single imperative behind brain function: minimize variational free energy. Free energy here is a mathematical quantity that bounds surprise, meaning how poorly an agent’s internal model predicts its sensory inputs. A system that keeps free energy low is, in effect, maintaining an accurate model of its world and staying within the states it expects to occupy.
The principle ties together two ways of reducing surprise. Perception and learning update the internal model so that it predicts sensations better, which is closely related to predictive coding and Bayesian inference. Action changes the world itself so that sensations match the model’s predictions, a process Friston calls active inference. The same objective thus accounts for both why we perceive and why we act.
Friston presented the principle as a candidate unifying framework in a 2010 Nature Reviews Neuroscience paper, arguing that older theories including the Bayesian brain, predictive coding, and reinforcement learning can be seen as special cases. The breadth of the claim is also the source of its controversy, since a principle general enough to cover everything can be hard to test against any particular finding.
For a general reader, the free energy principle is worth knowing as the most ambitious attempt to state a single law for the mind, and because its picture of agents that model the world and act to confirm their predictions maps closely onto how world-model based AI systems are being designed.