Karl Friston published “The free-energy principle: a unified brain theory?” in Nature Reviews Neuroscience in January 2010 (Vol. 11, pages 127-138). It is the most cited single statement of one of the most ambitious ideas in theoretical neuroscience.
Friston’s claim is that action, perception, and learning can all be understood as a system minimizing a single quantity, variational free energy, which is an upper bound on the surprise an agent encounters about its sensory inputs. A brain that minimizes free energy is, in effect, building a probabilistic model of its world and continually adjusting either the model or its own actions to make sensations less surprising. Perception updates the model; action changes the world so that it better matches the model.
The paper reviews how older theories, including the Bayesian brain, predictive coding, efficient coding, and reinforcement learning, can each be read as special cases or consequences of free-energy minimization. This unifying ambition is what makes the principle both influential and controversial. Supporters see a candidate first principle for cognition; critics argue it is so general that it is hard to falsify.
For a general reader, the free-energy principle is worth knowing because it frames the brain as a prediction machine that acts to confirm its own forecasts, an idea that resonates strongly with how modern self-supervised and world-model based AI systems are built.