What Is Mind?

The third ingredient

Walk up the tower one more step. Dissipation is a structure held in order by energy flowing through it, and it is everywhere. Memory is a copyable record that turns dissipation into something cumulative, subject to selection, and it is what makes a system alive. But the memory of bare life is slow. The genome is a model of the world, but a model written across generations and edited only by who survives. It cannot learn within a lifetime. It cannot represent this predator, this gradient, now.

Mind is what appears when the record becomes fast. A mind is a dissipative structure carrying a memory that it updates within a single life, an internal model of its environment that it runs forward to predict, and against which it compares what arrives, and by which it acts. The progression is exact:

The genome is a model on the timescale of evolution; the brain is a model on the timescale of a heartbeat. The rest of this essay is the case that this single addition, a within-lifetime predictive model that eventually includes the modeller, accounts for everything we can explain about mind, and isolates with unusual sharpness the one thing it cannot.

Why prediction is thermodynamically favoured

The first claim to establish is that the model is not a luxury bolted onto life but a thermodynamic economy, something selection drives toward for the same reason it drives toward any efficiency.

Start at the floor. Landauer’s principle: erasing one bit of information in a system at temperature $T$ dissipates at least

Information is physical, and clearing it has an energy cost. Now consider a system living in a changing environment, holding some internal state that correlates with what its sensors have seen. Still, Sivak, Bell, and Crooks proved a sharp result about such systems: the energetic inefficiency of the system, the work it dissipates beyond the reversible minimum as the environment drives it, is governed by the memory it keeps about the past that is useless for predicting the future. Information about where the predator was, beyond what that tells you about where it will be, is not free to hold. It is thermodynamic dead weight that must eventually be paid for at the Landauer rate.

The consequence is an arrow. A system that retains only the predictive component of its sensory history, discarding the rest, dissipates less and persists longer on the same energy budget. Selection on energy efficiency therefore pushes adaptive systems away from mere recording and toward modelling, toward keeping exactly the structure of the past that forecasts the future and no more. The predictive model is the thermodynamically efficient way for a dissipative structure to stay far from equilibrium in a world that changes. Life builds minds for the same reason it builds anything: because the alternative wastes the gradient. This is the bridge from the second essay to this one, and it is derivable, not asserted.

Persistence implies inference

The second claim is stronger and more surprising: a system that merely persists against dissipation can be described, with full formal rigour, as doing Bayesian inference on its world. This is the content of the free energy principle, and it is worth deriving because it is the precise statement of “to stay alive you must predict.”

A persisting system maintains a boundary, a statistical separation between its internal states and the external states of the world, mediated only through sensory and active states. To keep existing is to keep its sensory states within the narrow set compatible with being that kind of system, a fish stays wet, a cell stays intact, which is to say it must keep the long-run entropy, the average surprise, of its sensory states low. Let $o$ be the sensory states and $s$ the hidden external causes the system cannot see directly. Surprise is $-\ln p(o)$, and it cannot be evaluated directly because computing $p(o)=\int p(o,s)\,ds$ requires summing over all hidden causes. So the system carries an internal approximate distribution $q(s)$, its model, and minimises a quantity it can evaluate, the variational free energy:

Expand the joint, $p(s,o)=p(s\mid o)\,p(o)$:

Because the Kullback-Leibler divergence is non-negative, $F \ge -\ln p(o)$. Free energy is an upper bound on surprise. Minimising it over the internal model $q$ drives $q$ toward the true posterior $p(s\mid o)$, which is perception; minimising it over actions, by changing $o$ itself, is behaviour. The same quantity decomposes a second way,

so a mind is a system balancing the accuracy of its model against the cost of its complexity, exactly the trade-off any good scientist makes. The conclusion is that any system which maintains itself can be read as one that models its world and acts to make the world match its model. Perception, action, attention, and learning all fall out as the minimisation of one functional.

Now question it, because the project demands it and because the principle invites overreach. In its strongest form the free energy principle is close to a tautology: a system that exists is, by definition, one whose states stay in their expected set, so of course it “minimises surprise.” Read that way it describes rather than explains, and it is hard to falsify, since almost any persisting system admits such a description. Its real value is not as a law that could have been otherwise but as a translation manual, a proof that the language of thermodynamic persistence and the language of inference are the same language. That is enough for the present argument: it shows the step from life to mind is not a new kind of magic but a change of description forced by staying alive.

From model to self-model

A thermostat predicts and acts, and no one thinks it has a mind. The free energy principle by itself is too generous; it makes inference cheap. The decisive ingredient, the one that turns a model into something we are tempted to call a mind, is reflexivity: the model comes to include the modeller.

A self-model is a system’s internal representation of itself as one more object in the world it is modelling, an agent with a boundary, a location, a history, and goals. The moment the model contains a model of the system, including the modelling, the structure becomes self-referential, and self-reference is mathematically a fixed point. This is the same machinery that closed the first essay: Lawvere’s fixed-point theorem, the engine behind Gödel and Cantor and the strange loops of self-reference. A mind, on this reading, is a predictive model that has found a fixed point in which it appears to itself.

This is not loose analogy; it is where several of the most concrete theories of mind actually sit. Metzinger argues that the self is precisely such a model, a phenomenal self-model, and that strictly speaking no one has a self, one is a self-model, a transparent representation the system cannot see through and therefore takes for real, the “ego tunnel.” Graziano’s attention schema theory is sharper still and almost mechanical: the brain models its own attention the way it models a limb, builds a simplified schematic of the act of attending, and that schematic, queried, is what the system reports as “awareness.” Hofstadter called the general phenomenon a strange loop. What all of these share is that they explain a specific, otherwise puzzling fact: why a physical system would insist it is conscious, report an inner life, and be unshakeably convinced there is something it is like to be it. Those are outputs of a self-model, and a self-model would produce them. Hold onto that last point. It will turn out to be the hinge of the whole essay.

The functions of mind, and what they cover

Assemble the ingredients, dissipation, predictive memory, a world-model, a self-model, and a large class of things we mean by “mind” follow without further mystery. Perception is inference over the hidden causes of sensory data. Learning is the update of the model. Attention is the allocation of limited modelling resources to the most informative parts of the input. Action is inference run in reverse, changing the world to reduce prediction error. Self-monitoring, the sense of being a located agent with a past and a stake in the future, is the self-model doing its job. Report is the self-model made available to the systems that produce speech. The felt unity of a single experiencing subject is what a single, integrated self-model represents.

These are what Chalmers named the easy problems of consciousness, “easy” only by contrast, since each is a vast scientific labour, but easy in the sense that we know what a solution would even look like: a mechanism, a function, a model. The framework of this essay gives a clean and unified account of all of them. They are what a self-modelling predictive dissipative structure does to stay far from equilibrium in a changing world. If “mind” meant only this, the essay could end here, and the answer would be: mind is the tower’s third ingredient, a within-lifetime model that includes itself.

The neuroscience, and where it stands in 2026

It is worth marking that this is now an empirical science with real instruments and adjudicable theories, not only a philosophical one. Two functional theories lead the field. Global neuronal workspace theory, from Baars and Dehaene, holds that a mental content becomes conscious when it is broadcast widely, “ignited” into a global workspace that makes it available to many systems at once for flexible report and control; consciousness is access. Integrated information theory, from Tononi, takes the opposite starting point, defining consciousness as the quantity of irreducible, intrinsic cause-effect structure a system has, its integrated information $\Phi$, and locating it in a posterior cortical “hot zone” rather than in broadcasting.

In 2025 these two were put through a preregistered adversarial collaboration, the Cogitate study, in which proponents of both theories agreed in advance on the experiments and on what would count as failure, then measured 256 people with fMRI, MEG, and intracranial electrodes. The result was, with admirable inconvenience, inconclusive for both. The posterior synchronisation that integrated information theory predicted was not sustained as required; the prefrontal ignition and content that global workspace theory predicted were weaker than required, especially at stimulus offset. Neither theory won, and the proponents disagree about whose wounds were deeper. The lasting result is methodological: the field has matured to where it can stage a fair fight between theories of consciousness and let nature referee. There is also live controversy about whether integrated information theory is even falsifiable, a public dispute over whether it should be called science at all, which is itself a sign that the empirical pressure has become real.

The honest summary is that we can now find robust neural correlates of access consciousness, of whether and how a content is represented, reported, and made available, and we can test functional theories against one another. This is genuine progress, and all of it is progress on the easy problems of Section V. None of it, by construction, touches what comes next.

The gap it cannot cross

Here the essay must do what the previous two did at their floors, and state plainly the thing the framework does not reach.

Everything in Sections I through VI explains functions and reports. It explains why a system perceives, predicts, acts, models itself, says “I am aware,” and is convinced beyond argument that there is something it is like to be it. It does not explain why there is, in fact, anything it is like. The redness of red, the painfulness of pain, the sheer felt presence of experience, what Nagel pointed at by asking what it is like to be a bat and Chalmers named the hard problem, is not entailed by any account of structure and function. You can specify the entire causal, computational, self-modelling machine, and the question “but why is it accompanied by experience rather than running in the dark?” remains formally open. The machine, fully described, is logically compatible with there being nothing it is like to be it. That compatibility is the gap.

This is the third member of a family. “What Is Nothing?” fixed everything about existence except the underivable generators and was precise that they were the residue. “Are We Alone?” fixed the entire shape of the question except the value of one rate and named that rate. “What Is Mind?” fixes every function of mind except the felt interior, and isolates it exactly. In each case the framework’s honesty is the same: it does not pretend the residue away, it draws a clean boundary around it.

And as before, the responsible thing is to present the two serious stances at their strongest, because reasonable people who have thought hardest disagree.

The first is illusionism, or strong functionalism, the line of Dennett, Frankish, Graziano, and in his way Metzinger. Its case is built on the hinge from Section IV. A self-model would generate the conviction of an ineffable inner quality whether or not there were any further fact, so that conviction cannot be evidence for the further fact. Once you have explained why the system is certain there is something it is like, and why it finds that something indescribable, and why it reports a gap, you have explained everything there is to explain. The “hard problem” is then a cognitive illusion manufactured by the self-model’s own transparency, exactly the kind of artefact a fixed-point self-reference would produce. Parsimony is on this side: no new ingredient is needed.

The second is realism about phenomenal consciousness, the line of Nagel, Chalmers, and Goff. Its case is that experience is the single thing we know more directly than any theory, more certainly than the existence of the external world, since it is the medium in which all evidence arrives. No description of structure logically entails why it is like something, demonstrated by the conceivability of a functional duplicate with the lights off, so function cannot be the whole story. The gap is not an illusion to be dissolved but a real feature of the world, possibly requiring an extra fundamental ingredient: experience as basic (panpsychism), or properties orthogonal to the physical (property dualism), or physics not yet written.

I will not pretend to settle in an essay what the field has not settled in fifty years. But notice what the framework does contribute, which is not nothing. It makes the residue sharp. It shows that everything functional is on one side of a clean line and the felt interior is on the other, orthogonal to all of it, untouched by any mechanism. It even explains, through the self-model, why we would be convinced of the residue regardless of its status, which is the strongest single consideration the illusionist has and the strongest single thing the realist must answer. The framework cannot tell you which side is right. It can tell you, with unusual precision, exactly what the two sides are fighting over.

Two cautions, and a careful analogy

Before the last turn, two guardrails.

First, a clean refusal. It is a recurring temptation to wire mind into the foundations of physics by saying consciousness collapses the quantum wavefunction. The framework gives no warrant for this, and it should be rejected. Decoherence accounts for the appearance of measurement through ordinary entanglement with an environment, with no observer’s mind required; the older von Neumann-Wigner idea that consciousness causes collapse is unsupported and unnecessary. A mind is a macroscopic dissipative model, not a privileged physical agent that reaches into the microscopic. Keeping this straight matters, because the genuine mystery of mind is hard enough without importing a false one.

Second, an analogy offered only as an analogy, with the same caveat I attached to the gauge argument in the first essay. The first-person perspective is, by construction, absent from the third-person observables of a brain, in the way that the gauge of a potential is absent from the gauge-invariant field. Something real, that does work, that no external measurement resolves. The parallel is suggestive and worth holding, but it is not an identity: gauge-invisibility is a feature we build into the theory on purpose, while the explanatory gap is a principled barrier we did not choose and cannot obviously remove. The shared structure, real, effective, invisible to the external observables, is the reason the analogy is worth posing. Whether it is more than a rhyme is open, and I leave it open.

Underivability again: the self-model at its own floor

Now the turn this essay exists to make. Recall the load-bearing result of “What Is Nothing?”: no system can derive the existence of its own generators from within the system they generate. Gödel and Löb in logic, gauge surplus in physics, the fixed point that relocates rather than dissolves the regress.

A mind is a self-model, a fixed point in which a predictive system represents itself. The hard problem, stated structurally, is the self-model attempting to derive its own phenomenal character from within its own description. And that is precisely the move the underivability thread says cannot be completed. A self-referential system can represent itself, can report on itself, can be convinced of its own interior, but it cannot, from inside, certify the ground of that interior, cannot derive why there is an inside at all rather than a dark machine. The felt residue may be exactly what the underivable generator looks like from within: the inside view of a structure that cannot, by the same theorem that limits every sufficiently rich system, justify its own innermost fact to itself.

I offer this as the new possibility, in the spirit of the project, not as a solution. It does not close the gap; nothing in this essay closes the gap. What it does is relocate the gap into a family we already understand. The hard problem becomes the mind’s version of Gödel’s wall, the place where a self-model runs into the limit on self-reference that the first essay derived in the abstract. If that framing is right, then the irreducibility of experience is not an anomaly peculiar to brains but the local, lived instance of the most general fact in the trilogy: a system cannot ground itself from within. The mind is simply the first place in the tower where that ungroundedness is felt.

What this opens

Set the trilogy’s third floor down beside the other two.

1. Mind is the tower’s third ingredient. Dissipation rides a gradient; memory makes the ride cumulative and alive; a within-lifetime predictive model makes it a mind, and a model that includes the modeller makes it a self.
2. The model is forced, not optional. Prediction is thermodynamically favoured, because non-predictive memory is wasted energy, and persistence against the second law is formally equivalent to inference. Minds are what efficient survival in a changing world looks like from the inside of the equations.
3. The functions are explicable, and now empirically so. Perception, learning, attention, action, self-monitoring, and report all fall out of one functional, and the field can at last stage fair tests between theories of access consciousness.
4. The residue is sharp, isolated, and orthogonal to all of it. Why there is something it is like is not entailed by any function. The framework does not dissolve this; it draws a clean boundary around it and even explains why we are convinced of it regardless of its status.
5. The residue may be underivability, felt. The hard problem reframes as the self-model encountering, from within, the same limit on self-reference that bounds every sufficiently rich system. The inside cannot ground the inside.

That is where the three essays meet. Existence rides a gradient between two silences and cannot derive its own generators. Life is that ride carrying a memory, scattered as a possibility across a galaxy of dead-star debris, one rate away from telling us whether it is alone. Mind is that ride carrying a model that has turned to face itself, and at the floor of it lies the same ungroundedness as at the floor of everything else, except that here, uniquely, there is something it is like to stand on that floor and look down. We are the part of the universe that models itself well enough to find the one question it cannot answer from inside, and to feel, rather than merely prove, that it cannot. The feeling is the wall. And the wall, this whole trilogy has argued, is not a flaw in the view. It is the view.