You Are Your Own Existence Proof

Lex Fridman opens by asking Karl Friston how much of the human brain we actually understand, from neurons up through function and psychiatric disorder. Friston, cited more than two hundred and forty-five thousand times for work in brain imaging and theoretical neurobiology, declines a score. We are in a better position than last century, he says, but how far remains almost unanswerable until you decide what counts as understanding.

A perfect cartography of every dendrite would send you into a life of microscopy. A neuropsychologist after strokes and lesions would stay with the whole person. Friston's own wager sits in between. If the brain is a machine doing inference, you do not understand its thermodynamics by chasing single molecules, any more than you understand a gas by tracking every particle. Go too coarse, though, and you miss the canonical wiring that makes message passing possible at all.

That is why he is excited about high-field magnetic resonance imaging. For the first time it can resolve activity at a few millimeters, enough to distinguish cortical layers whose microcircuitry is repeated across the brain. Somewhere between a millimeter and a meter, he says, is where the interesting questions live. Asked for the brain's most beautiful trait, he does not reach for consciousness or language. He names its hierarchical, recursive anatomy.

A liver, Friston notes, does exquisite biochemistry without which you would not last long. It does not have the brain's crafted sparsity, recurrence, and specialization. Fridman admits he has always pictured the brain as a giant interconnected mass, a tangle too dense to parse. Friston says that is what he thought too, before medical school. "It's not a magic soup."

Hierarchy, he argues, is not something you build by adding links. You start from a fully connected soup and carefully remove the connections that would destroy any layered structure. A hierarchy is defined by a sparse and particular connectivity. The farther apart two regions are, the less likely they are to share a direct fiber. Anatomy, not metaphor, is what underwrites the depth.

Trace studies draw that sparsity as something like an onion. All sensory traffic and all outgoing commands to muscles and glands exchange at the surface. A layer beneath watches that exchange, another layer watches that one, and so on into the core. If the brain is a message-passing machine, the circuitry that shapes the passing also dictates the function. Delete the sparse recursive structure and the remaining soup really cannot do much.

Most human neuroimaging that appears in journals measures fluctuations in activity over time, usually in response to a stimulus or a task, sometimes at rest. Two complementary principles organize the analysis. Functional specialization says that some tissue is dedicated to a kind of processing, such as seeing visual motion. Functional segregation says that specialization is spatially restricted, so a moving bird should light up a circumscribed patch that a static image does not.

Friston is willing to be unkind about the resulting maps. Some people call the enterprise a neo-cartography; he also hears it described as phrenology with better pictures, a hunt for blobs that do this or that. The hope is not empty. In twenty or thirty years the field has gone far beyond incremental confirmation of a century of lesion studies. Patients who lost V5, also called the middle temporal area MT, could not see moving things. Electrodes in animals found motion-sensitive receptive fields in the same neighborhood. Imaging then showed the region lighting up in living human brains.

Lesions alone had never been enough. Lashley's law of mass action, inherited from animals that showed the same deficit no matter which spoonful of brain you removed, had once made the magic-soup view respectable. Positron emission tomography and then functional MRI in the early 1990s let researchers watch restricted responses in place. The bigger question, Friston says, is functional integration: how those regionally specific responses are orchestrated. Fridman uses the fashionable word "deep." Friston pounces. Deep, he says, is "the millennial equivalent of hierarchical." Put it in front of anything and you sound trendy, but you are still implying a layered architecture.

Non-invasive scanners first measure structure, a nuanced X-ray of water, fat, or iron. Dynamic function splits into two camps. The dominant one records metabolic and hemodynamic signals, proxies for neuronal work whose time constants are much slower than the spikes themselves. Neurons cannot run a deficit the way muscle can. A few millimeters of busy cortex have to be watered now, within a couple of seconds. That neurovascular coupling is what fMRI rides, at a spatial grain of millimeters and a temporal grain of seconds.

The other camp records electromagnetic fields as they are generated, with EEG and MEG. Temporal resolution drops toward milliseconds, fast enough to watch a cascade of hierarchical message passing after a stimulus. The cost is space. By the time the signal has crossed white-matter spaghetti, meninges, skull, and scalp, it is spatially diffuse. Friston calls the dilemma a catch-22. You can know within millimeters which patch activated and not know when, or know within milliseconds when something responded and not know where. The two measures, indirect via blood and direct via fields, have to be used together.

Once the modality is chosen, analysis again splits between specialization and integration. The popular stream is what Friston nicknames blobology, a neologism that would appear in a newspaper and not in a specialist journal: the study of little blobs on brain maps. The joke has a serious underside. The mathematics of those blobs, including Euler characteristics, is real, and statistical parametric mapping is, in his dry phrase, mathematically furnished blobology. A statistical measure crosses a threshold in a restricted patch and a blob appears. That is evidence of segregation. It is not yet a picture of how the onion talks to itself.

Fridman mentions a day at Neuralink and asks about interfaces that listen and write inside the skull. Friston's word is ambivalent, and he owns the mix. Stimulation is not new. A century of work at the same neurological institutions stimulated dogs and read out muscle or salivation. Deep brain stimulation for Parkinson's is now standard treatment and a window onto movement-disorder dynamics. Transcranial magnetic stimulation for depression is crude, he says, a bit like banging a television when the valves are failing. It still sometimes works.

The happier frontier is sensory substitution. Give someone a magnetic sensor that encodes heading as tactile pressure on the abdomen, and after a few weeks they take it for granted. They are not reading a gadget; they feel which way they face. That plasticity is why Friston does not dismiss clinical hopes for locked-in syndrome, paraplegia, blindness, or deafness. The deflationary side is bandwidth. Present interfaces, he says, deliver bits per second, like communicating in very slow Morse code, not even within an order of magnitude of what those aspirations would require.

The deeper objection is the kind of system you are trying to join. The brain is high-dimensional, nonlinear, deeply structured, and held in a non-equilibrium steady state. He asks the listener to imagine aggressive satellites that can poke local patches of weather and then demand that the weather respond on cue. That is chaos control at an almost unimaginable scale. To be part of the system you would have had to evolve with it. It is not like rewiring a broken computer or plugging in a peripheral. It is like getting into the weather. Fridman, coming from machine learning, counters that the brutal part is the engineering, including a vascular web you cannot nick, and that once the satellites exist the same methods that win games can play in the soup. Friston accepts the hoped-for conclusion: really hard, hopefully not impossible.

The free energy principle, Friston says, inherits from high-dimensional time-series analysis of brains and from machine learning. Formally it is a statement that the existential imperative of any system that survives in a changing world can be cast as inference. The probability of existing is the evidence that you exist. Write existence as a statistical problem and two centuries of inferential mathematics become a description of the ensemble dynamics that must be in play.

Technically, anything that exists in virtue of being separate from its environment can be read as minimizing variational free energy. People in machine learning know that quantity as the negative evidence lower bound, the negative ELBO. Equivalently, the dynamics will look as if they are maximizing the marginal likelihood, the evidence for the system's own existence. Fridman hears an answer to what it means to be alive. Friston slows him down. The claim is not only about living things. It is about existence. If something exists, then by the mathematics of non-equilibrium steady-state physics it must exhibit properties that make it look as if it is optimizing a particular quantity, and that quantity happens to be Bayesian model evidence.

He answers as a physicist manqué, last taught physics in his twenties, and he keeps the deflationary clause in view. The principle does not say why a thing exists. It says that if it exists, it must display these properties, and the properties have an inferential gloss. The simplest picture in his head is a single-celled organism, or even a drop of oil in water. Why does the interface persist? Why does the oil not dissolve into the solvent? In computational chemistry that is a macromolecule configuring itself to minimize thermodynamic free energy, which is the same mathematical construct as the negative ELBO. Self-assembly, autopoiesis, the emergence of a form that can be told from its soup, all look like the same optimization. The twist of phrase he loves is that you are your own existence proof.

To say that something exists you first have to say where it finishes. Friston writes that down as statistical independencies: not what is connected, but what is not correlated and does not depend on something else. Sparsity again. Once you can identify the surface of the oil drop in the water, any universe that contains a thing falls into four kinds of states. There are internals, externals, and the surface itself. That surface is a Markov blanket.

Blanket states can then be carved into sensory states and active states. By the definition of the partition, external states cannot influence internal states except through the blanket. Active states, operationally, cannot be influenced by the external states either. Internal and active states together therefore have autonomous dynamics. Even a little oil drop now has a region of its surface that the outside cannot reach, and an interior that looks as if it is minimizing the same free energy, which is to say maximizing model evidence.

The force of the derivation, as Friston tells it, is that it grows out of probability and non-equilibrium physics without borrowing a single biological assumption. You do not need a theory of life. You need a boundary. Fridman says the picture almost talked him into pantheism: the drop already seems to sense, act, and maintain itself. The next question is what, if anything, still distinguishes that drop from a tadpole.

Fridman tries mortality, sensation, action. Friston says he has nailed it: movement. Not being moved, which the oil drop will do if you stir the bowl, but movement caused from inside. Internal states that can influence active states can cause the surface to go somewhere. There are, if you like, two kinds of oil drops. In one, the internal states are so random they average themselves away, and on balance the thing cannot move.

The sun is his extreme case. It has violent internal activity, but not the deep hierarchical structure a brain has, so no overall pattern expresses itself on the surface as swimming. The surface can be busy while the average position of that surface cannot relocate, because the interior is a hot gas. A living interior is structured enough to drive muscles, vocal folds, glands, and the autonomic nervous system. Then the whole organism actually moves.

There is, he says, no other way you can change the universe than by moving, whether that movement is a voice box, a walk, or juices squeezed from secretory organs. Doing so non-randomly is what makes you alive. The tadpole's active states change external states; those changes return as sensations; internal states update and drive the next movement. That loop is the prototype of chemotaxis, foraging, and behavior. An oil drop can be read as a perfect passive model of "I am sitting in a bowl of water." Equip it to push its surface around and taste chemical gradients, and it starts to look like a life form.

Since the turn of the millennium, Friston says, intelligence research has moved toward an enactive, embodied picture, and the central importance of movement has not really hit machine learning. It has diffused through robotics, and through vision problems where you must move a camera to sample. Data-mining deep learning has contended with the issue by declining it. We need not worry about moving, the field effectively says, because we have big data. We can see all the observations already.

In his frame that makes contemporary machine learning much more like the oil drop, except an oil drop granted exposure to nearly every datum. The tadpole has to swim out to find the right data. "It likes food" is a testable hypothesis: move, ingest, and see whether the sensory consequences are evidence that you are the kind of thing that likes this food. Active sampling is what turns statistical inference into behavior and a model into an agent. A system that never has to decide what to look at has skipped the loop that, on this view, is a large part of what intelligence is.

Asked whether the framework can talk about consciousness, Friston says you would want a qualified philosopher for a definitive answer, but information geometry and related ideas are being used to write down a calculus of minimal consciousness and even of selfhood. The first rung is agency. A virus has a form of it in the way it finds host cells, yet you would not endow it with planning that countenances the future. A foraging animal talking to its friends along the way has a different, biotic autonomy. Somewhere in the complexity of planning a line is often hoped for.

A few months before the conversation, a philosopher taught him the technical notion of vagueness. At what point does a collection of grains become a pile of sand? One grain is not a pile, two are not, and there is no honest grain at which the concept snaps into place. He had always assumed a line between conscious and unconscious would eventually be drawn. It might be the wrong kind of question. Having said that, he still thinks planning is the wrinkle that matters. If the generative model on your sensory surface includes the future consequences of your own action, you are in the game of planning as inference. That already requires movement, and a model of the sort a variational autoencoder might encode, thinking about the future conditioned on different courses of action.

Selecting among policies is, in this writing, what free will looks like: a self-evidencing artifact that appears to choose among ways forward. Friston does not think that gets you all the way to self-awareness. You would have to commit to a formal definition. Most people he knows would deny that a goldfish is self-aware, argue about a favorite cat, and grant the status to their mother. Fridman suggests that folding yourself into the plan is essential for complex planning. Friston agrees the line is unclear, and that the observation is more useful than it first sounds.

What is the principle good for? Friston compares it to natural selection: beautiful, undeniably true, and silent about why you have legs and eyes. It will not, by itself, let you build something. Survival of the fittest is circular in the same deflationary way. Things exist because they minimize free energy; they minimize free energy because they exist. You can go round and round.

The practical residue, analogous to genetic algorithms in machine learning, is that things which exist look as if their density dynamics optimize a variational free energy, and that quantity is a functional of a generative model, a probabilistic story of causes out there and consequences on the sensory blanket. Write down the model, work out the gradients, and you should be able to cause an artifact to self-evidence: an oil drop, a robot, in principle even the AGI you want, performing gradient descent on that objective. Mathematically he calls it easy. The heavy lifting is writing the generative model. That form is the phenotype. Natural selection never told you why you have eyes; the free energy principle will not either, until you specify the model that would have them.

Fridman's remark about putting yourself into the plan tells Friston what kind of generative model a self-aware artifact would need. The model has to be fit for the world in which it operates, and that world is largely populated by other things like it. If you can hypothesize that we are the same kind of creature, only I am me and you are you, then a sense of self becomes mandatory. In a social world you must infer that it is me talking and not you. You would not need that inference alone on Mars, or as a feral child with nothing like you around. Ambiguity induced by others is what makes "I" a hypothesis worth maintaining.

That is why we pour our attributes into pets and try to make robots humanoid. It is easier to read the world if you can assume that you are me and it is just your turn to talk. The highest realization of planning, Friston says, is the conversation they are having. Turn-taking and theory of mind mean I need a model of your model, including when you want to listen, interrupt, or drop off the line. Without that nested generative model they would talk over each other or sing in a choir. The dance of discourse is evidence of the models that support self-awareness.

Asked for the meaning of life, he is tempted to answer as a physicist: the free energy expected from his behavior, including searching for information about the kind of thing he is. The personal version is simpler and hops back to childhood. You are given beliefs about the person you are, in bedtime stories and fairy tales, from parents, friends, and culture. "I'm a princess and I'm going to meet a beast who's going to transform into a prince." All that self-evidencing then means is fulfilling those beliefs in an embodied way. Because active inference cuts both ways, you are also changing the culture that wrote the script. Friston's own mix was Einstein and Sherlock Holmes: smoke heavily, avoid too much contact, enjoy the fantasy of a slightly quirky popular scientist. His father was an engineer who loved Eddington's *Space, Time and Gravitation*. He grew up, he says, to do what a mild-mannered Holmes-slash-Einstein would do in his shoes.

Fridman closes with a line of Friston's that inverts the usual story of action. "Your arm moves because you predict it will, and your motor system seeks to minimize prediction error." Command-and-execute is the common-sense order. On this view the brain predicts a posture and the motor system abolishes the gap, so the arm arrives in the act of making the world match the forecast. The same move, stretched over decades instead of milliseconds, is a life spent turning a childhood narrative into evidence.