A Pattern That Makes Itself

June 09, 2026

This piece grew out of a philosophy course I took last semester. For the final paper, I compared and criticized two arguments against conscious machines: Roger Penrose’s Gödelian case that the mind exceeds any formal system, and Anil Seth’s biological naturalism, which holds that consciousness requires a living, metabolic body. While I originally tackled both, I have since reworked the Seth side into a self-contained piece, building out the computational scaffolding I initially assumed as background. By first exploring cellular automata and Dennett’s Real Patterns, my goal is to unify a set of ideas. With this framework in place, I aim to explain why Seth’s objection fails to convince me. Hope it will be a good read.

Imagine shrinking until you are small enough to walk across a computer screen, so small that the square beneath your feet is a single pixel. From down here there is no image at all. There is only your pixel, blinking on, then off, then on again, following some rule you cannot see, surrounded by millions of others doing the same. No landscape, no face, no meaning: just lights flickering in the dark. To find the picture, you would have to climb back out, up and away, until the pixels blur together into something whole.

This change of scale points toward one of the deepest puzzles in science (if you accept a physicalist stance, I suppose): how does the physical world, which inherently consists of particles and their movements, give rise to something as strange as a living mind? How do atoms (which know nothing and feel nothing) eventually arrange themselves into creatures that know they exist?

This question assumes some kind of emergence, and I want to approach it from a computational perspective. To do that, we first need to understand the concept of emergence, and what it means to approach it computationally. So let’s set the scene.

Why Emergence Is a Difficult Word

The word emergence is one of the most contested in all of science and philosophy. Skeptics have called it a placeholder for ignorance, a way of saying something happens here that we don’t yet understand while making it sound profound. Others argue it is not a single concept at all, but a family of loosely related phenomena that resist unified treatment. Emergence, in this sense, can operate like a god of the gaps, a name we reach for whenever the reductionist assumption, that the way to understand something complex is to take it apart, runs out of road.

These criticisms are not trivial. Those who study the concept tend to distinguish a stronger reading, in which higher-level properties are completely irreducible to lower-level ones, from a weaker one, in which higher-level behavior is in principle derivable from the lower-level rules but only through computation too expensive to be tractable. Neither reading has settled the debate. The strong version is widely contested, the weak version is sometimes accused of being so permissive it says nothing interesting. A reasonable strand of opinion holds that the word should be abandoned entirely until we have sharper tools for saying what we mean.

One way forward, I think, is to stop arguing about definitions and start looking at concrete cases. That is exactly what John von Neumann did, and it is where I want to start.

Von Neumann’s Question

By the late 1940s, John von Neumann had helped design the atomic bomb, described the architecture underlying every modern computer, and made foundational contributions to quantum mechanics, game theory, and economics. But beyond all of that, one problem bothered him, not a problem of physics or mathematics, but one where automata theory meets biology: what is the minimum logical requirement for a machine to reproduce itself?

The question sounds straightforward until you think about it carefully. A machine that builds another machine needs a blueprint. But the blueprint is itself a physical object, which means the copying machine must also copy the blueprint, which means the blueprint must contain instructions for copying the blueprint, and so on, forever. Von Neumann had run into an apparent infinite regress of descriptions. Nature, of course, had solved this effortlessly for three billion years. But how?

Working with the mathematician Stanisław Ulam, von Neumann arrived at a strange solution. He abandoned continuous space entirely and replaced it with something radically abstract: a discrete grid where time moves in steps, and each cell follows a simple, local rule.

He designed a 29-state automaton capable of self-reproduction, which was also the first discrete parallel computational model in history formally described as capable of universal computation. He had not just presented a solution to a puzzle, in the same stroke he had invented a new model of computation.

It is worth noting that von Neumann never ran his automaton. In the early 1950s the computation was beyond reach, and his design existed on paper. It is also worth noting that his description of a machine reading its own tape and copying it predates the discovery of DNA’s structure. When Watson and Crick revealed the double helix in 1953, biologists recognized that DNA does something close to what he had theorized: it functions as both the blueprint and the object being copied.

The grid was born from an interesting question. And the assumption built into his solution was that what matters about life is its logical structure, abstractable from any particular physical stuff. This is a large claim. If what makes something alive is a matter of organization rather than material, then living things might in principle be realized in almost any medium that can carry the right structure.

What Is a Cellular Automaton?

In the twenty years after von Neumann, the concept was simplified and democratized. Cellular automata are discrete, abstract computational systems: spatially and temporally discrete, abstract, and computational. At bottom, any cellular automaton has four ingredients:

The most important concept here is local action. A cell knows nothing about the board as a whole. It looks only at its immediate neighbours, consults the rulebook, and updates itself. And the gist is this: even perfect knowledge of individual decision rules does not always let us predict macroscopic structure. We get macro-surprises despite complete micro-knowledge.

Conway’s Game of Life

In 1970, the mathematician John Conway introduced what would become the most famous cellular automaton in history. It runs on an infinite two-dimensional grid, and its transition rule contains exactly four lines:

Dennett, in his 1991 paper Real Patterns, notes that this single deterministic rule set makes the Life world a toy that perfectly instantiates Laplace’s determinism: given the state at one instant, you can in principle predict every future instant by applying the rules. The Life world has no hidden variables. Nothing is backstage. And yet what emerges from this transparency cannot be read off from the rules at all.

From these four rules emerge still lifes that hold their shape, oscillators that pulse in place, and spaceships that travel. The most famous is the Glider: five cells that, through nothing but the blind, local application of the rules, appear to move diagonally across the grid:

 Gen 0      Gen 1      Gen 2      Gen 4
· ■ · ·    · · · ·    · · · ·    · · · ·
· · ■ ·    ■ · ■ ·    · · ■ ·    · · · ·
■ ■ ■ ·    · ■ ■ ·    ■ · ■ ·    · · ■ ·
· · · ·    · ■ · ·    · ■ ■ ·    · · · ■
· · · ·    · · · ·    · · · ·    · ■ ■ ■

After exactly four generations the Glider returns to its original shape, but has moved one cell diagonally. No code instructs it to move. No rule says “travel south-east.” It moves because the rules happen to produce motion.

If you’d like to see it for yourself, here is a live grid. Press Run and watch the glider walk diagonally, moved by nothing but the rules. Click a cell to toggle it and click and drag to draw.

Wolfram’s Classification Scheme

In the 1980s, the physicist Stephen Wolfram spent years cataloguing every possible one-dimensional cellular automaton. In a one-dimensional CA, the grid is a single line. Each cell’s next state depends on its own state and its two neighbours, giving eight possible configurations and therefore 2⁸ = 256 possible rules. Wolfram encoded each as a number (Rule 30, Rule 90, Rule 110…) and found that all 256 sort into four behavioral classes:

The edge of chaos

The fourth class is where everything interesting lives. Wolfram conjectured (and Christopher Langton later gave the idea a quantitative handle) that Class IV occupies a critical threshold, the edge of chaos, between the frozen regularity of Classes I–II and the entropic noise of Class III. Langton’s parameter λ measures the fraction of a rule’s outputs mapping to a non-zero state. As λ rises from 0 to 1, behavior shifts from uniformity through periodicity into chaos, and Class IV rules cluster at an intermediate value, where information can both propagate across the grid and persist over time.

The strong version of this idea (that computation lives at the edge of chaos) was later challenged: Mitchell, Crutchfield, and Hraber showed that several of the headline results were partly artifacts of how the experiments were run. So treat it as suggestive, not proven. What survives is modest but useful (complex behavior needs a regime where information neither freezes nor dissolves), and that much is more than metaphor: a real relationship between a rule’s structure and the behavior it can support.

An Example: Rule 110

It helps to see one of these rules in full. Rule 110 runs on a single line of cells, each either 1 (filled) or 0 (empty). To update a cell, you look at a neighbourhood of three: the cell itself together with its immediate left and right neighbours. Three binary cells give eight possible neighbourhoods, and the rule simply specifies an output for each:

Read the right-hand column from top to bottom (01101110), and you get 110 in decimal. That is where the name comes from. The other 255 elementary rules are numbered the same way. You apply the rule to every cell at once to produce the next line, then repeat.

What makes Rule 110 remarkable is what happens when you let it run from a random starting line. It does not dissolve into uniform static, and it does not settle into a clean repeating pattern. Instead it throws off intricate, asymmetric structures: localized patterns that travel and collide, the one-dimensional cousins of the Glider. In 2004 the mathematician Matthew Cook proved that this behavior is enough to make Rule 110 Turing complete: with the right starting line of 1s and 0s, this single trivial-looking rule can in principle carry out any computation any computer can: run a calculator, a game, an entire program. Universal computation, hiding inside an eight-row lookup table. Cook’s original construction was slow, simulating a Turing machine only with exponential overhead, which left room to suspect the universality was a technical curiosity. However, Neary and Woods later showed the same emulation can be done in polynomial time, so the computation hiding in the lookup table is not just possible but efficient.

The Glider Problem, and What Counts as Real

Ask a physicist to look at the Game of Life and they will say: there is no such thing as a Glider. There are only cells turning on and off according to rules. The Glider is an illusion, a story we tell ourselves about pixels.

And technically, the physicist is right. There is no line of code that says move_glider_diagonally(). If you describe the game entirely at the level of the rules (the fundamental level) you will never find it. But in Real Patterns, Daniel Dennett asked a simple question: if the Glider isn’t real, how do you predict where it will be in 10,000 generations, short of grinding the rules cell by cell through every step?

He grounded this in information theory. Drawing on Gregory Chaitin’s definition of randomness, he argued that a data series is random if and only if its description is incompressible: nothing shorter than the verbatim bit-map will do. But if some more efficient description exists, then the series has a pattern. The Glider is real by this criterion: ‘five cells moving diagonally every four ticks’ is more compact than listing every cell it passes through. The compression is not a cognitive convenience, it is a feature of the data.

Dennett calls this level of description the design stance: you treat a system as if it were designed to behave in a certain way, and you predict accordingly. The Glider is a design-stance object. His deeper aim in Real Patterns is to show that the same compressibility argument applies one level higher, at what he calls the intentional stance: treating a system as a rational agent with goals. That is where folk psychology lives, and that is the real target of the argument.

Dennett’s deeper motivation for this argument is to offer a new way of thinking about the status of folk psychology in cognitive science. Folk psychology is the everyday practice we all use to explain and predict behavior in terms of mental states: beliefs, desires, intentions, fears. You expect a friend to take an umbrella because she believes it will rain and wants to stay dry. In early cognitive science this ordinary framework became a battleground. The question was whether folk psychology is a kind of proto-scientific theory, one that a mature science of the mind would either vindicate, by finding real beliefs and desires inside the head, or else discard as a primitive misdescription, the way chemistry discarded phlogiston. The stakes were high: if there is nothing belief-shaped in the brain, does that mean beliefs are not real?

Both Fodor and Churchland share one assumption: that for a higher-level pattern to be real, it must correspond to some matching structure in the underlying mechanism. They simply draw opposite conclusions from it. Fodor, the realist, holds that beliefs and desires are real because the brain contains internal representations with the right causal structure, a language of thought whose logic mirrors that of folk psychology. Churchland, the eliminativist, accepts the same demand for a structural match but bets it will fail: if neuroscience finds nothing in the brain shaped like a belief, then folk psychology is simply a false theory with no scientific standing, destined to go the way of phlogiston.

Dennett occupies the middle: higher-level patterns can be real without being locally encoded or mirrored by the mechanism that produces them, defended on the hard-nosed grounds of predictive success through compression. The Glider is real because betting on it pays off. Folk psychology is, in some sense, real for the same reason. A pattern is real, here, when it is useful to an observer, when someone standing outside the system can use it to compress and predict.

With this move, Dennett claimed that beliefs and desires can be perfectly real without being things you could find by opening the skull, real in just the way the Glider is, as patterns that earn their place by letting an observer predict what would otherwise be intractable. The same goes all the way up the ladder of the living world. You are made of an almost unimaginable number of atoms, none of which has the faintest idea it belongs to a person. No one tracks them one by one, because cell, organ, organism, and belief are the real patterns that compress that incomprehensible churn into something usable, each derivable in principle from the level below but never in practice. Mind, on this view, is not an illusion to be explained away, nor a ghostly extra ingredient, but a real pattern in the physical stuff beneath it, an organism a Glider at a vast scale, or so the analogy goes.

The Ghost in the Grid

Your brain contains billions of neurons. Each can be caricatured as a cell in a grid following one local rule: if I receive enough signals from my neighbours, I fire, otherwise I stay quiet. (Of course, real neurons are far richer than binary cells: they vary in timing, are modulated chemically, and trade in graded rather than strictly all-or-nothing signals. But the structural point survives the simplification.) Map every spark across all neurons and you would see what a physicist sees in Conway’s game: a storm of local events with no obvious meaning at that level. You would not see a memory. You would not see a belief. You would not see a self.

And yet you are reading this sentence, understanding it, evaluating it, perhaps pushing back. Something is happening that is not described in the firing rates of individual neurons. Just as the Glider is a pattern in the pixels, the thought is a pattern in the neurons, or so the computational story goes.

Dennett’s argument is of a specific kind, and the kind matters. It is not an argument from ignorance. It is an argument from computational intractability: even with complete and perfect knowledge of the lower-level rules, predicting macro-behavior by computing through them is in practice impossible at the scales that matter. The intentional vocabulary (beliefs, desires, intentions) is not a placeholder awaiting reduction. It is a compression scheme that captures real predictive patterns.

Emergence, Computationalized

The CA framework lets us replace a vague philosophical concept with a precise one: an emergent pattern is a pattern whose compressed description offers a computational saving over the physical description. The Glider is derivable from the rules, but only through computation that grows astronomically with time. At 10,000 generations, refusing to use the Glider description is not integrity, it is stubbornness. Patterns below this threshold are just noise or simple periodicity. Patterns above it, the Glider, the organism, the belief, are real by a single criterion: they are good to bet on.

But notice the shape of that criterion, because the rest of the essay turns on it. Everything here is an account of patterns licensed by an observer’s ability to compress and predict. It is, by design, a view from outside, and it works beautifully for what those systems will do. Yet there is a question it cannot reach on its own terms: what it is to be one of these systems. A Glider does not feel itself gliding. Does an animal feel its hunger? Do you feel your beliefs? The compression criterion is the wrong instrument for that question, not because it gives the wrong answer, but because it is not built to ask it.

Autopoiesis: The View From Inside

For all its power, Dennett’s criterion has a blind spot built into its strength. A pattern is real if an observer can use it to compress and predict. That says nothing about what it is to be that unit, if being it is like anything at all.

In the early 1970s, two Chilean biologists set out from the opposite end. Humberto Maturana and Francisco Varela asked not “when should an observer count something as alive?” but “what does a living system do that makes it a system in the first place?” Their answer was autopoiesis, from the Greek for self-producing. A living system is a network of processes that continuously produces the very components that produce the network, and in doing so specifies its own boundary. A cell manufactures the membrane that contains the chemistry that manufactures the membrane. There is no blueprint read by an external machine, as in von Neumann’s design. The organization and the thing organized become the same loop, sustained from within.

Maturana and Varela were explicitly rejecting the standpoint the first half of this essay relied on. A Real Pattern is something we find useful. An autopoietic unity is something that constitutes itself, whether or not anyone is watching. Their 1972 book’s original Spanish title was, tellingly, De máquinas y seres vivos (On Machines and Living Beings), and its burden is that a living being is not merely a machine described from the outside, but a unity that continually brings itself forth, something an external description never fully captures.

And they pressed the point further than biology. Living systems, they claimed, are already cognitive systems: living as a process is a process of cognition. A bacterium swimming up a glucose gradient is, in their sense, already knowing its world, discriminating, responding, maintaining itself against perturbation. Cognition is not a trick brains bolt onto life. It is what being alive already is.

That difference, on this view, is what the Glider lacks. A Glider does not produce its own components. It does not maintain a boundary against a world. It does not resist its own dissolution. Nothing is at stake for it. Remove the observer who calls it a Glider and there is nothing but cells obeying the rule. So the autopoietic verdict is that the organism is not simply a very complex Glider: it is the one thing a Glider supposedly can never be, a pattern that holds itself together, and for which holding together matters.

It is worth mentioning that autopoiesis is contested as a definition of life. Some purely physical, clearly non-living systems (Bénard convection cells, a candle flame) display a kind of self-maintaining organization without being alive, which suggests the criterion is at least incomplete. Maturana and Varela’s later extension of the idea to societies and language strained it further. So read autopoiesis here not as a settled definition but as a reorientation: from the pattern as seen to the system as self-made. Even if it does not draw the boundary of life perfectly, it isolates what the compression criterion leaves out entirely.

The Beast Machine

If autopoiesis is the abstract claim, the neuroscientist Anil K. Seth has given it a modern, empirical body. In Being You, Seth argues that perception is not a transparent window onto the world but a controlled hallucination: the brain constantly generates predictions about the causes of its sensory signals, and what we experience is the brain’s best guess, continuously reined in by sensory data. Seeing is a hallucination held on a leash, and an ordinary hallucination is simply perception that has slipped the leash.

Seth’s deeper move is why brains do this. Not to build accurate pictures, but to keep the body alive. Beneath vision and hearing lies interoception: the brain’s prediction and regulation of its own internal physiological state, the many variables that must stay within narrow bounds for the organism to persist. The conscious self, on this account, is not a thing that has a body. It is the felt interior of a body regulating itself in the service of staying alive. He calls this the beast machine, to insist on our animal, metabolic nature against any picture of the mind as a disembodied computer.

When Seth turns this into an argument against conscious machines in his article Conscious Artificial Intelligence and Biological Naturalism, it comes in two steps. The first is an attack on substrate flexibility, the assumption, built into computational functionalism, that the functions underlying consciousness can be lifted off their biological hardware and re-run on any medium that preserves the right organization. Seth thinks this assumption fails. Neural function, he argues, is not cleanly decomposable into discrete, Turing-computable operations: it depends on continuous electromagnetic field interactions, fine-grained temporal dynamics, freely diffusing neurotransmitters, and metabolic regulation, real-valued, temporally extended processes that, on his view, are not merely hard to simulate but in an important sense lie outside computation, because their functional power is grounded in material dynamics rather than symbol manipulation. Evolution has generatively entrenched these functions, coupling them so tightly to specific material features that they resist decomposition into modular, swappable parts. As Seth puts it, brains may be the kind of thing for which it is hard, perhaps impossible, to separate what they do from what they are.

From this follows his main point: the difference between simulating a process and instantiating it. A digital model might reproduce the input–output behavior of a conscious brain without thereby being conscious, just as simulating a rainstorm leaves everything bone dry. Simulating neural dynamics, on this view, no more produces experience than simulating rain produces water. He reinforces this with an appeal to causal and informational closure: in evolved, deeply entrenched systems like the brain, the high-level functions stay coupled to the low-level material dynamics, so you cannot cleanly lift the software off the wetware and expect the relevant powers to survive the move.

The second step is the positive account. Seth grounds it in two frameworks: the Free Energy Principle (FEP), which characterizes living systems as those that act to minimize free energy (roughly, surprise about their own sensory states) and so resist the entropic slide toward disorder, and predictive processing (PP), which models perception, action, and selfhood as hierarchical Bayesian inference aimed at minimizing prediction error. Seth insists these are not disembodied abstractions: they are realized through the concrete physical dynamics of metabolism, homeostasis, and boundary maintenance in an autopoietic system. Consciousness, for him, arises from prediction-driven self-regulation in a metabolically active, self-organizing body, not from computation in the abstract.

This is where the two traditions, computational functionalism and biological naturalism, come to a head over a single question: does the substrate matter?

Seth does not claim to have proven this. He offers it as a principled reason to doubt that consciousness is substrate-neutral, and it grows directly out of autopoiesis: the system must be one for which its own continued existence is at stake.

Notice the lineage. Varela spent his later career developing enactivism, the thesis, set out with Evan Thompson and Eleanor Rosch in The Embodied Mind, that mind is not computation in the head but the activity of a living body bringing forth a world. Thompson later pushed this into a full life–mind continuity thesis: the very properties that make something alive are continuous with those that make it a mind. Seth’s interoceptive, life-first picture is the descendant of this line. Where Dennett asks what patterns an observer should bet on, this tradition asks what it takes to be a system for which anything is at stake at all, and answers: you have to be alive.

Turning the Lens

Seth’s life-first view is a serious objection to the computational emergence story, so it deserves the hardest scrutiny. When you give it that scrutiny, a pattern emerges: several of his load-bearing claims either turn on themselves or rest on a step that is never quite taken. Let me go through them.

The frameworks cut the other way

Start with the foundation of Seth’s positive account: the Free Energy Principle and predictive processing. In essence, both are information-theoretic, substrate-neutral formalisms. FEP describes a system’s behavior as variational inference on a generative model. PP describes cognition as hierarchical Bayesian prediction-error minimization. Neither says anything about carbon, neurons, or metabolism in particular. A defender might reply that the formalism is neutral while the territory it describes, a system that actually maintains a thermodynamic boundary, is not. I take up that move below. But notice first that the neutrality is not an accident or a simplification, it is what gives these frameworks their power. Even their action-oriented form, active inference, has already been implemented in artificial agents, and is closely related to planning as inference, a standard machine-learning way of treating decision-making as probabilistic inference over action trajectories.

So Seth faces a dilemma. Either FEP and PP capture what is essential to consciousness (in which case, being substrate-neutral, they can in principle be instantiated in a non-biological system, and substrate-dependence collapses) or they do not capture what is essential, in which case Seth owes us an account of the missing ingredient. He cannot have it both ways: he cannot rest his theory of consciousness on formalisms that abstract away from biology and simultaneously insist that biology is indispensable.

The way to put the pressure is as a burden-shift. Biology can name a location, but not a mechanism. To block machine consciousness, it is not enough to observe that the only conscious systems we know of happen to be biological. One must identify the specific causal property of the wet substrate that is, first, necessary for instantiating these functional principles and, second, functionally irreducible, impossible to realize in any other medium. Seth gestures at metabolism, continuity, and entrenchment, but has not yet isolated such a property, shown it to be both necessary for these functional principles and impossible to realize in any other medium. Until that is done, the claim that biology is constitutive rather than incidental remains an intuition, not an argument.

Autopoiesis Is Computable

Recall the hinge of the life-first case: that a pattern on a grid can never truly hold itself together, that self-production belongs to life alone. This is where the view overreaches. The claim is false, and the proof has been sitting in this essay’s opening all along.

Recall where we began: von Neumann’s self-reproducing automaton, a configuration on a grid that builds a copy of itself by reading and copying its own description. This is the strongest example, because it does not merely model self-production: running it is self-production, the genuine article on a grid, decades before anyone tied the idea to biology. And it was not a fluke of his particular design. Christopher Langton’s loops are minimal self-replicating structures in a cellular automaton, and Bert Chan’s Lenia (a continuous generalization of the Game of Life) produces soft-bodied, lifelike creatures that maintain their own boundaries, hold together under perturbation, and repair local damage. Together they show how far the same idea extends, from bare replicators to things that look alive. (Whether the lifelike cases are genuinely autopoietic or only compelling models of it is contested. But on the organizational definition Maturana and Varela actually gave, the burden now runs the other way.)

This is not a historical curiosity but an active research program. Mordvintsev and colleagues’ neural cellular automata learn a local update rule under which a pattern grows from a single seed and then regenerates after being cut: self-repair as a learned, computable dynamic rather than a designed-in trick. Recent work on computational life shows that when random, non-self-replicating programs are left to interact in a substrate with no imposed goal, self-replicators tend to arise on their own, and complexity keeps building once they do. Self-production emerges from the dynamics rather than being engineered into them. A parallel line of argument, developed at book length by Blaise Agüera y Arcas, takes the von Neumann thread to its conclusion: that self-reproduction, and so life itself, is at bottom computational. None of this settles what life is. But it makes the life-first assumption, that self-maintenance and self-production are the signatures of something computation can only ever mimic from the outside, look increasingly like a bet against the evidence.

This matters because Maturana and Varela themselves defined autopoiesis as an organizational principle (a particular topology of self-producing processes), not as a property of any specific material. Nothing in the definition mentions carbon. If autopoiesis is organizational, and organization is exactly the kind of thing computation can capture, then there is no principled barrier to a computational system being autopoietic. The Glider was simply the wrong example: it is a rigid, fixed pattern that does not regulate anything. But the Glider does not exhaust what grids can do. Self-maintenance, boundary formation, and recursive self-production are all within reach of computable dynamics. The view from inside turns out not to be the exclusive property of meat.

Some will object that this still misses the thermodynamics. A living cell does not merely instantiate the organization of self-production, it pays for its persistence, doing metabolic work to hold itself far from equilibrium against the second law, while a pattern on a grid maintains itself for free. However, it is important to notice that any running computation is itself a physical, dissipative process: the machine updating the grid consumes energy and exports entropy exactly as Landauer’s principle requires, so the costlessness is a feature of the mathematical description, not of any actual system doing the computing. And insofar as the thermodynamic story is made precise, it is made precise through the Free Energy Principle, which is itself a substrate-neutral formalism, already implemented in artificial agents. So the objection faces the same fork as before: either far-from-equilibrium self-maintenance is captured by a formalism that can be realized in other media, in which case it is no barrier, or there is some further ingredient that resists formalization, in which case we are owed its name. What is true is narrower: today’s computational examples have not tightly coupled a pattern’s self-maintenance to its own energy budget the way a cell does, so nothing on a grid yet gets hungry. But that is a fact about how far the engineering has come, not about what computation can do, and reading a current limitation as a permanent one is exactly the slide this essay has been resisting.

This points to something many philosophers take for granted too quickly: that a program’s formal description is cleanly separable from the physical medium running it. Conceptually it is, which is exactly why we can write algorithms without reasoning about them voltage by voltage. But a running program is not its description, it is a physical process that the description merely specifies, a pattern in how real voltages actually change over time. To run the algorithm just is to drive some medium through those changes. So if a program can generate behavior complex enough to model autopoiesis, then any physical medium capable of running it does not describe that process from outside, it instantiates it.

Complexity is not uncomputability

Seth’s strongest-sounding claim is that the biological processes underlying consciousness (metabolism, neurotransmitter diffusion, homeostatic regulation) are not just complicated but non-computable.

In its strict, formal sense, non-computability is a precise and demanding property. A function is non-computable only when no algorithm can compute it, not merely that no efficient algorithm is known, but that none can exist, on any machine, given any amount of time. The canonical case is Turing’s halting problem, and the family of undecidable problems that reduce to it: questions for which one can prove, by diagonal argument, that no general procedure can ever return the answer. To earn the claim that a physical process is non-computable, you have to show that its behavior encodes something like that, a problem demonstrably beyond the reach of any algorithm. Mere intricacy does not do this. Neither does continuity, nor sensitivity to initial conditions, nor the sheer number of interacting parts. A system can be staggeringly complex, even chaotic, and remain perfectly computable. Seth never supplies the reduction. What he supplies is the intuition that these processes are too rich, too continuous, too embodied to be captured by an algorithm, and an intuition, however well-motivated, is not a proof.

The history of artificial intelligence is in part a graveyard of confident predictions that some task lay beyond mechanism. Vision and fluent language once looked obviously non-algorithmic, too high-dimensional and context-soaked to yield to a machine, and decades of hand-written symbolic rules failed badly enough to make that conclusion look safe. Then convolutional networks did vision and transformers did language, once the architecture was right. The lesson is not that everything is easy. It is that apparent intractability usually reflects the limits of our current models, not a formal limit on computation. What looks impossible from inside one paradigm routinely falls to another.

There may be a further slip worth naming. It is not obvious that continuous and real-valued imply non-computable. Continuous dynamics can often be approximated to arbitrary precision (neural networks that learn differential equations are one example), and at least for self-maintenance and regulation, the organizational facts that matter seem to survive the discretization. If that is right, then approximating a real-valued process to whatever tolerance the functional story requires would still capture what matters about it. So the move from biological and continuous to beyond computation may be doing more work than it can bear: it has to slide first from complexity to uncomputability, and then again from the analog to the non-algorithmic, and neither step is obviously licensed.

Put plainly, this is a conflation of practical intractability (we cannot yet simulate it well) with formal uncomputability (no algorithm could). A great deal of the life-first argument leans on that single equivocation. Once you separate the two, the claim that the biology relevant to consciousness lies beyond computation loses almost all of its support.

The rainstorm begs the question

Which leaves the most quotable objection: simulating a rainstorm doesn’t make anything wet, so simulating a brain doesn’t make anything conscious. It feels decisive, but I believe it is question-begging.

The reason the rain analogy works is that wetness is a substrate property: it is about water. So of course simulating it does not instantiate it. But that is only a good analogy for consciousness if you have already assumed that consciousness is a substrate property rather than an organizational one, which is precisely the point in dispute. Swap in a different kind of thing and the intuition reverses: a simulated chess game is not a fake chess game, it is a real one. The simulation instantiates the very thing it models, because chess is defined by its organization, not by its stuff. Whether consciousness is rain-like or chess-like is the entire question, so the simulation argument cannot be used to answer it. It smuggles in the conclusion it is supposed to earn.

I will admit to a doubt about the chess example. Is it really as clean as I am making it sound? How I play chess is normative, a matter of following rules I could get right or wrong, while a machine’s playing looks merely causal, electrons doing what physics makes them do. There is a real Wittgensteinian worry in the neighbourhood. But notice it rescues nothing for Seth: if the normative character of rule-following meant a machine only ever simulates chess, it would mean the same for me, since the rules of chess are no more written into my neurons than into silicon. What fixes the norm is participation in a practice, and nothing in that demands the participant be alive. The worry bites brains and machines alike, so it cannot be the thing that makes biology special, and it leaves the computational picture untouched.

One last point, in passing. Seth’s own framework arguably already commits him to a kind of multiple realizability. If FEP and PP describe a continuum of conscious or proto-conscious systems from bacteria to humans, he has accepted that the same principles are realized across radically different physical implementations, and there is no non-arbitrary place to draw the line that lets in the bacterium but rules out the machine.

Can the Grid Be Conscious?

It is worth being candid about how we got here, because the route was not a straight line. The first half of this essay leaned on Dennett: a pattern is real if an observer can use it to compress and predict. Cellular automata were the ground for this: they showed that organizational properties like self-reproduction, boundary maintenance, and self-regulation are not features that require biological encoding, but can be realized in computation. The life-first challenge then showed that Dennett’s criterion was not quite enough: a Glider is real to a viewer, but an organism is real to itself, and that difference is not nothing. The right response was not to retreat to biology but to upgrade the criterion: from real to an observer to intrinsically self-maintaining, and then to notice that self-maintenance is itself an organizational property within computation’s reach, as the automata already showed. That is why Seth’s argument does not hold: he may be right about what matters, that the deepest patterns are the ones that hold themselves together, but the CA framework shows that such patterns do not require the biological substrate he insists on. The defeat comes from his own premises, turned against his conclusion. We have, in a sense, climbed back out from the single pixel to the whole picture, and found the picture was never made of anything but organization.

But there is one thing all of this leaves completely untouched. Suppose we build a system that instantiates the whole functional organization: autopoiesis, interoceptive prediction, a self that regulates its own viability. We will still face the question: is there something it is like to be it? Why does any of that organization come with an inner light rather than running in the dark? This is David Chalmers’s hard problem, and nothing in the argument above dissolves it.

What I wanted to say is that the hard problem is not a special liability of the computational account. It is a problem for every physicalist account of consciousness, biological ones included. Seth’s brain-based, metabolic story faces it exactly as squarely as a silicon one: explaining all the self-regulating dynamics in the world still does not explain why they are felt, whether those dynamics run on neurons or on transistors. The hard problem taxes the biological naturalist and the computationalist at the same rate, which means it cannot be recruited as a reason to think machines in particular cannot be conscious. It is everyone’s mystery, not computation’s alone.

That is the shape of the conclusion, and it says almost nothing about consciousness, like almost every essay on consciousness. But at least we have said something. The case that consciousness requires living matter, or that it exceeds computation, does not survive scrutiny. What survives is the hard problem, and it points at no substrate in particular. On this terrain, machine consciousness is not a confusion to be dismissed but a defensible possibility.

Final Words

So, who is right: computationalists, biological naturalists, or someone else? To be honest, I don’t really care. I accept that consciousness really is a mystery once you adopt a physicalist stance ontologically, but I don’t think the problem will be settled by arguments alone, and yet here we are, making them. I also don’t believe consciousness is the ultimate question. I believe there are many questions that matter more than the consciousness question, which too often turns into a kind of academic playground, where ideas are thrown around for the sake of throwing them. But I believe a science of consciousness is possible, if we ask the right questions.

I want to end this essay with a thought from Bas van Fraassen’s “Transcendence of the Ego (The Non-Existent Knight),” whose conclusion I share. Van Fraassen invites us to consider Agilulf, the non-existent knight of Calvino’s novel, an empty suit of armor that nonetheless rides, speaks, and lifts a sword. If such a being existed, he argues, science would face no special crisis. It would simply do what it always does: observe the phenomena, build a first rough model with almost no predictive power, replace it with a better one, and pair each model with explanatory stories that can be tested. There is no argument that science could fail at this, because wherever there are observable phenomena, a scientific account can in principle depict them, inventing unobservable posits where the depiction needs them.

The emergence of consciousness, he says, is no different. The transition from a world without organisms to one with them, and from no consciousness to consciousness, is a change like any other. The question “how does consciousness emerge?” already has an answer at any given moment: whatever model our current science provides for that stage, disappointingly thin today but improving all the time. He likens the perplexity to Hume on miracles: to demand how there could be observed exceptions to a fully adequate science is to ask something self-contradictory, since science simply expands to accommodate whatever actually happens. Philosophers who remain perplexed by how consciousness could possibly have emerged at all in a merely physical world must, he concludes, mean something other than the scientific question, and unless they can keep the two cleanly apart, their problem is a pseudo-problem.

His parting line is the one I want to leave you with. There is a mystery of consciousness, he grants, but it is not among the mysteries that the sciences confront, the ones they so habitually and fortunately address and solve.

Further reading

• Daniel C. Dennett, Real Patterns (https://ruccs.rutgers.edu/images/personal-zenon-pylyshyn/class-info/FP2012/FP2012_readings/Dennett_RealPatterns.pdf)
• Francesco Berto and Jacopo Tagliabue, Cellular Automata (https://plato.stanford.edu/archives/Spr2016/entries/cellular-automata/index.html)
• Eric W. Weisstein, Elementary Cellular Automaton (https://mathworld.wolfram.com/ElementaryCellularAutomaton.html)
• Matthew Cook, Universality in Elementary Cellular Automata (https://wpmedia.wolfram.com/sites/13/2018/02/15-1-1.pdf)
• John von Neumann, Theory of Self-Reproducing Automata
• Stephen Wolfram, A New Kind of Science (https://www.wolframscience.com/nks/)
• Melanie Mitchell, James P. Crutchfield, and Peter T. Hraber, Dynamics, Computation, and the “Edge of Chaos”: A Re-Examination (https://melaniemitchell.me/PapersContent/dyn-comp-edge.pdf)
• Andy Clark, Mindware: An Introduction to the Philosophy of Cognitive Science
• Francisco J. Varela, Evan Thompson, and Eleanor Rosch, The Embodied Mind
• Humberto R. Maturana and Francisco J. Varela, Autopoiesis and Cognition: The Realization of the Living
• Evan Thompson, Mind in Life: Biology, Phenomenology, and the Sciences of Mind
• Anil K. Seth, Being You: A New Science of Consciousness
• Anil K. Seth, Conscious Artificial Intelligence and Biological Naturalism (https://www.cambridge.org/core/journals/behavioral-and-brain-sciences/article/conscious-artificial-intelligence-and-biological-naturalism/C9912A5BE9D806012E3C8B3AF612E39A)
• Karl J. Friston, The Free-Energy Principle: A Unified Brain Theory? (https://www.nature.com/articles/nrn2787)
• Karl J. Friston, Jean Daunizeau, and Stefan J. Kiebel, Reinforcement Learning or Active Inference? (https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0006421)
• Noor Sajid, Philip J. Ball, Thomas Parr, and Karl J. Friston, Active inference: demystified and compared (https://arxiv.org/abs/1909.10863)
• Christopher G. Langton, Self-Reproduction in Cellular Automata (https://fab.cba.mit.edu/classes/865.18/replication/Langton.pdf)
• Bert Wang-Chak Chan, Lenia: Biology of Artificial Life (https://www.complex-systems.com/abstracts/v28_i03_a01/)
• Ricky T. Q. Chen et al., Neural Ordinary Differential Equations (https://proceedings.neurips.cc/paper_files/paper/2018/file/69386f6bb1dfed68692a24c8686939b9-Paper.pdf)
• Alexander Mordvintsev, Ettore Randazzo, Eyvind Niklasson, and Michael Levin, Growing Neural Cellular Automata (https://distill.pub/2020/growing-ca/)
• Blaise Agüera y Arcas et al., Computational Life: How Well-formed, Self-replicating Programs Emerge from Simple Interaction (https://arxiv.org/abs/2406.19108)
• Blaise Agüera y Arcas, What Is Life? Evolution as Computation
• David J. Chalmers, Facing Up to the Problem of Consciousness (https://consc.net/papers/facing.pdf)
• Bas C. van Fraassen, Transcendence of the Ego (The Non-Existent Knight) (https://www.princeton.edu/~fraassen/articles/pdfs/TranscendenceRatioOffprint.pdf)