Karl Friston

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI insists that what separates a practitioner who grows from one who stagnates is not what they produce but how they engage with the uncertainty of the work. Friston’s framework supplies the deepest theoretical grounding available for this distinction. A system that minimizes free energy by acting to resolve uncertainty—seeking information rather than exploiting what it already has—is, on Friston’s account, doing something that no system optimized purely to predict the next token is doing. Curiosity, in the free energy framework, is not a personality trait; it is a mathematical necessity for any agent with a sufficiently rich model of itself.

The dominant machines of the AI age are, in Friston’s framing, extraordinarily sophisticated minimizers of surprise about their training distribution. They are very good at predicting what text follows the text they have seen. They do not have Markov blankets that define a self with stakes in its own persistence; they do not update their generative models through action in the world; they do not minimize the free energy of a self that exists over time. What they do is extraordinary and genuinely useful. But it is not what Friston means by intelligence, and the confusion between the two is, he argues, the deepest strategic error in contemporary AI development.

The cycle’s question is not whether AI is powerful but what kind of power it is—and whether the human who wields it is developing the self-directed, uncertainty-resolving, world-modeling agency that makes the power usable in genuinely novel situations. Friston’s framework makes the answer precise: a practitioner who uses AI to eliminate uncertainty rather than to engage with it is, in the free energy vocabulary, not an agent performing active inference but a passive system adjusting its beliefs to match an externally provided model. The result is competent output and atrophied agency.

He thus stands in the cycle alongside Andy Clark—who shares his commitment to the predictive brain and extends it into the extended mind thesis—and Humberto Maturana, whose autopoiesis framework addresses the same question about what distinguishes a living, self-maintaining system from its environment. Friston’s contribution is to give both frameworks a single mathematical foundation and to deploy it as a critique of the scaling paradigm that currently dominates AI research.

Origin

Friston trained as a psychiatrist before moving into computational neuroscience, bringing a clinician’s patience with complex systems and a physicist’s appetite for fundamental principles. His early work developed methods for brain imaging analysis, including statistical parametric mapping, that became the dominant tool in neuroimaging research and would alone have secured him a prominent place in the field. But the work he considers central began from a question that imaging could not answer: what must a brain be doing, at the most general level, simply in order to be a brain at all?

The answer he arrived at drew on Helmholtz’s nineteenth-century insight that perception is unconscious inference, on the Bayesian framework for rational belief revision, and on the thermodynamic concept of free energy, which in statistical physics measures the gap between an actual state of a system and its expected state. Friston argued that a brain—or any self-organizing system—minimizes a quantity formally analogous to free energy: the divergence between its internal model of the world and the sensory data the world actually provides. To perceive is to make predictions and update them; to act is to generate sensory data consistent with the predictions. The unification of perception and action under a single variational principle is the free energy principle’s central achievement.

The mature framework’s most contested and most original element is the Markov blanket: the statistical boundary, defined in terms of conditional independence, that separates a system’s internal states from the environment it models. Friston argues that anything that persists as a distinct thing must possess such a boundary—that the Markov blanket is the mathematical definition of individuation, of what it means to be a separate entity at all. This move connects the free energy principle to questions of consciousness, identity, and the boundary of the self that reach far beyond neuroscience.

Key Ideas

The free energy principle. Any system that maintains a boundary with its environment must, by the logic of thermodynamics, minimize the divergence between its internal model and the sensory evidence it receives. Friston calls this quantity—formally a variational bound on surprise—free energy. The free energy principle is his claim that this minimization is not a strategy a system might adopt but a necessary condition for its persistence as a distinct system at all.

Active inference. The free energy principle is satisfied in two ways: by updating the generative model to better fit the data (perception), or by acting on the world to bring the data in line with the model (action). Active inference is the second mode, and it is what makes an entity an agent. An agent does not only update its beliefs about an unchanging world; it selects actions that will resolve its uncertainty by generating informative new sensory data. Active inference thus unifies perception, action, and learning under a single principle, and curiosity—the drive to seek information that resolves uncertainty—falls out of the mathematics as a necessary consequence.

The Markov blanket. The statistical boundary, defined by conditional independence relationships, that separates an entity’s internal states from its external environment. The Markov blanket is not a physical surface but a statistical structure: the set of states that, once known, render the internal and external states statistically independent. Friston argues that anything that can be said to be a self must possess such a blanket—that the Markov blanket is what individuation means at the level of information theory.

Predictive coding. The specific neural implementation of free energy minimization in the brain. The cortex, on this account, is organized as a hierarchy of prediction machines: each level sends predictions down to the level below and receives prediction errors back up. What propagates is not raw sensation but the gap between what was predicted and what arrived. Predictive processing explains a range of perceptual phenomena—illusions, attention, hallucination, the placebo effect—as consequences of a single architectural principle.

Against the bigger hammer. Friston’s critique of the scaling paradigm is not that large language models are not impressive but that they are optimizing the wrong quantity. A model that predicts text has no self, no Markov blanket, no active inference, no intrinsic drive to resolve the uncertainty of a world in which it has stakes. Scaling makes such a model better at prediction and nothing else. Genuine intelligence, on Friston’s account, requires not more parameters but a different architecture entirely: one organized around self-maintenance, action, and the minimization of free energy over time.