You Are Your Own Existence Proof: Karl Friston on Free Energy, Life, and Self-Awareness

Friston begins by noting that the free energy principle inherits from data-analytic approaches to high-dimensional brain data and from machine learning. He frames it as a formal statement: any system that survives in a changing world faces existential imperatives that can be cast as an inference problem. The probability of existing can be treated as the evidence that one exists, so existence becomes a statistical problem. Once written that way, the mathematics of inference can characterize the dynamics that serve that inference.

Technically, anything that exists as a separate entity can be interpreted as minimizing variational free energy, which machine learning knows as the negative evidence lower bound. Equivalently, its dynamics look as if they maximize marginal likelihood, the evidence for its own existence. Friston stresses that this is the core of the principle before any biological details enter.

Fridman asks what it means for a blob to be alive or to exist, and Friston widens the question beyond living things. He uses 'existential' deliberately: the principle concerns existence first, not only life. Borrowing from non-equilibrium steady-state physics, he says existing things have properties that let their existence be interpreted as an optimization procedure.

Friston calls the principle both ambitious and deflationary. If something exists, the mathematics of non-equilibrium steady state makes it look as if it is optimizing a particular quantity. That quantity turns out to match the evidence lower bound in machine learning, Bayesian model evidence in statistics, and a bound on surprisal in information theory.

Friston imagines the simplest case as a single-celled organism identifiable from its environment, or even a drop of oil in solvent. The question is why the interface between the oil and the not-oil persists instead of dissolving. He describes the drop as an ensemble of molecules in a heat bath, with exchange across its surface. The free energy principle asks what properties such an interface must have to keep itself from disappearing.

Friston distinguishes the psychological reading of predictive processing from the physicist's reading he is using here. The principle does not answer why something exists; it says that if it exists, it must display certain properties. Those properties have an inferential, self-evidencing gloss because preserving the boundary requires optimizing a particular functional. He calls this an existential imperative.

He connects this to autopoiesis in biology and self-assembly in computational chemistry. A macromolecule can be understood as configuring itself to minimize thermodynamic free energy, which is mathematically the same construct as the negative ELBO. The emergence of a structure distinguishable from its environment therefore necessitates an objective function that it looks as if it is minimizing.

Fridman asks whether there is an interesting distinction between living and existing. Friston says rocks, trees, and grains of sand can also be separated from their environments, so at some level they too must be optimizing free energy under the model-evidence interpretation. That means they are self-evidencing. He turns the idea into a phrase: statistically speaking, you are your own existence proof.

To make this precise, Friston says one must define what probabilistic properties something needs so that it knows where it finishes. The answer is written in terms of statistical independence, what is not correlated and does not depend on what. This identifies the surface of the oil drop from the water, and it yields four kinds of states: internal, external, and the surface itself, which is a Markov blanket.

Once the partition is in place, standard results from non-equilibrium physics can be applied. The internal states and part of the blanket look as if they are hill-climbing or doing gradient descent on the same quantity. Carving the blanket into sensory and active states gives the system autonomous dynamics: active and internal states are not influenced by external states. Friston notes that this can even lead into panpsychist language, where everything that exists is implicitly making inferences.

Fridman asks what separates an oil drop from a tadpole or larva. Friston says the key is movement, but not just any movement: the movement must come from the inside. Internal states can influence active states, and active states can influence external states, even though external states do not influence active states. That causal structure gives a system something like autonomous action.

Friston contrasts this with the Sun. The Sun has internal activity, but its internal states are more like a hot gas and average themselves away, so no overall pattern expresses itself on the surface as swimming. In a brain, internal dynamics are deeply structured and can be expressed through muscles and effectors. He stresses that there is only one way to change the universe: move, whether by walking, speaking, or secreting.

That non-random movement is what makes a system alive in this picture. A tadpole's active states change external states, creating an action-perception cycle and a recurrent dynamic. Its internal dynamics not only model data impressed on its Markov blanket but actively resample those data by moving toward chemical gradients. Chemotaxis is a simple example of testing hypotheses about the world.

Friston says this emphasis on movement has not fully hit machine learning, though it has diffused into robotics and active vision. Deep learning often avoids active sampling by assuming access to big data. In his analogy, current machine learning is more like an oil drop that enjoys exposure to nearly all the data than like a tadpole swimming out to find the right data.

Fridman asks whether the framework can define consciousness or self-awareness. Friston says it can be used to think about consciousness, though he defers definitive philosophical answers. He focuses on agency and planning: a virus has a form of agency in finding hosts, but one would not say it plans. An ant seems to have purpose and a different kind of biotic autonomy.

The line may have to do with the complexity of planning. Friston warns that the philosopher's notion of vagueness complicates any sharp boundary, using the example of when a pile of sand becomes a pile. Asking where exactly consciousness begins may be like asking exactly when a collection of grains becomes a pile. Still, he agrees that systems able to plan are closer to the line.

Technically, self-evidencing must be evidence for a model of what causes sensory impressions on the Markov blanket. If that model can plan, it must include future consequences of active states, which is planning as inference. This moves away from big-data machine learning because the system must select data by looking or moving. It also requires a generative model that considers the future conditioned on different courses of action.

Once a system selects among policies, Friston says it begins to look as if it has free will. That selection makes it an inference machine that chooses among alternative ways forward. He thinks planning brings one much closer to a line for consciousness, but not necessarily all the way to self-awareness. Formalizing self-awareness would require committing to a formal definition.

Fridman asks what the free energy principle is actually good for. Friston answers that it can be seen as a theory of non-equilibrium steady states, but also as a deflationary tautology. Like natural selection, it is beautifully true but does not by itself explain why you have legs or eyes. It says existing things minimize free energy, and that existing things exist.

The practical payoff comes from generative models. If a system's dynamics look as if they minimize variational free energy, that free energy must be a functional of a generative model of causes and consequences. In principle, one can write down that model, work out the gradients, and let the system self-evidence autonomously. One could supply an objective function and self-organizing dynamics to create an artifact.

Friston says the mathematics is easy once the generative model is specified: write down the probabilistic model and make a robot perform gradient descent on that objective. The hard part is writing the generative model in the first place. That is where the heavy lifting lies, just as natural selection does not tell you why eyes exist without an account of the actual phenotype.

Friston asks what a generative model for self-awareness would need. If the world is largely populated by other things like me, then a model that treats them as similar creatures becomes useful. But if I am me and you are you, the model must also infer a sense of self. In a world of other minds, it becomes necessary to distinguish who is causing what.

He illustrates this with the feral child or a solitary existence on Mars. If there were nothing like me around, there would be no need for an inference that it is me and not you experiencing or causing these sounds. Ambiguity induced by others makes self-modelling necessary. Self-aware artifacts, in this view, are equipped with models that allow for many copies of things like them.

Fridman connects this to pets and humanoid robots: people readily project attributes onto them, perhaps because it is easier to read the world by assuming the other is like me. Friston agrees and points to conversation as the highest manifestation of planning. Turn-taking requires mutual inference and theory of mind, including a model of your model of my head.

The most sophisticated generative model contains a generative model of something else's generative model. Friston says the current conversation evinces the kinds of models that would support self-awareness. Without them, the participants would talk over each other or sing in a choir rather than have a discourse.

Fridman asks for the objective function of existence, the meaning of life. Friston is tempted to answer as a physicist: free energy expected consequent upon behavior. That leads to searching for information and resolving uncertainty about the kind of thing one is. He suspects, however, that the question calls for a more personal answer that remains consistent with the physics.

The personal answer is that people have beliefs about the kind of creature and person they are. Self-evidencing means fulfilling those beliefs in an embodied, enactive way. Those scripts and narratives are given early, often in bedtime stories and fairy tales, and are reinforced by parents, friends, and society. The objective function becomes to fulfill the narrative that family and culture have instilled, and that one also creates.

Because self-evidencing is active, people model their environment and change it, and others do the same back. This synchrony means individuals create their own culture across time scales. Friston describes his own childhood scripts as a mixture of Einstein and Sherlock Holmes: he smoked heavily, tried to avoid too much interpersonal contact, and also fantasized about being a popular scientist who would make a difference. His father was an engineer who loved science, including a book on space-time and gravitation by 'Sir Arthur' that offered an understandable version of general relativity.

He says the fairy stories he grew up with were about these characters, though he keeps the Hobbit out because it made him nervous. He describes growing up to be what he imagined a mild-mannered Sherlock Holmes would do in his shoes. The personal reflection ties back to the idea that self-evidencing is the fulfillment of an acquired narrative.