Chapter 8bCriticality and Connection

Many unconscious states show reduced neural signal diversity; psychedelics can increase it and alter experience. The entropic brain hypothesis proposes that conscious dynamics occupy a useful balance of flexibility and constraint. The following evidence examines criticality, shared mammalian architecture, and the pairing of exploration with regulation.

Criticality: The Edge Where Minds Live

Criticality is the state of a system poised at the boundary between two phases. The textbook case is water at its liquid-gas critical point (374 degrees Celsius, 218 atmospheres), where the meniscus dividing liquid from vapor vanishes and density fluctuates on every scale at once, scattering light until the fluid turns milky. Physicists call that milkiness critical opalescence. The name comes from that “critical point” in phase-transition physics, the precise threshold where a system shifts from one regime to another. Critical systems share distinctive properties: fluctuations at all scales, maximal sensitivity to perturbation, and the ability to propagate information over long distances.

Strong evidence shows that healthy brains operate near criticality.⁴⁶

The signature is a power-law distribution (a mathematical pattern where small events are common and large events rare, like earthquakes: many tremors, few catastrophes). This pattern appears in neural activity across species, from rodents to primates, most clearly in healthy, awake, attentive brains. Hengen and Shew (2025), reviewing 140 datasets spanning two decades, found that the long-running debate over whether the brain sits at or merely near the critical point reflected differences in statistical methods rather than genuine disagreement about dynamics.

Their conclusion: criticality is a homeostatic setpoint, a target state the system actively returns to after perturbation, maintained through plasticity mechanisms the way the body maintains its temperature at 37 degrees Celsius. A twin study (2025, N=829) confirmed the functional stakes: brain criticality is heritable and genetically linked to cognitive performance.³⁴⁷ Evolution placed this setpoint under selection pressure.

Finding the setpoint appears to require no experience at all. In 2026, researchers at the Technical University of Munich recorded spontaneous activity in 45 human forebrain organoids, lab-grown clusters of neural tissue a few millimeters across, and found the whole population hovering just above the critical point: a branching ratio of 1.099 ± 0.052, meaning each spike triggered on average almost exactly one successor.³⁴⁸ The organoids received no stimulation, no training, and no sensory input of any kind. Flat two-dimensional cultures of the same basic components reach criticality reliably only when driven with structured input through closed-loop feedback, so three-dimensional architecture seems to supply a tuning capacity the flat dishes lack. A few cubic millimeters of naive tissue, grown in total sensory isolation, drifts toward the same dynamical regime the awake human brain actively maintains.

The equivalence is structural. Researchers have compared human brain dynamics to the two-dimensional Ising model (the mathematical framework governing phase transitions in magnets, developed further in Chapter 17). At the Ising model’s critical point, the two systems are statistically indistinguishable on the properties compared: correlation lengths, fluctuation distributions, information capacity.³⁴⁹

Near a critical point, some quantities describing a system grow or vanish as powers of its distance from the threshold. That power is a critical exponent, and its value depends on ingredients such as dimensionality, symmetry, interaction range, and the kind of transition. Changing those ingredients can change the exponents, even across systems made of similar material. Systems that share a set of exponents form a universality class: some magnets and water can sit in the same one, which is why a physicist who has solved a magnet has partly solved a fluid.

The book’s own measurement (Experiment A14) places the cortical transition closer to three-dimensional Ising: its exponent beta, which governs how fast order builds once the threshold is crossed, comes out at 0.291 ± 0.031. The two-dimensional match above concerns the statistical properties compared, not the exponents. That puts the cortex in a different universality class from the two-dimensional social transition that governs the trust-coercion phase boundary (Chapter 17). Both live within the broader Ising family: the mathematics of cooperative alignment, whether among iron spins, neurons, or social agents. The shared framework means the same analytical tools apply; the different universality classes mean the critical exponents, and therefore the scaling behavior near the transition, are substrate-specific.

Why would the brain evolve to operate near criticality?

Information processing. At criticality, systems represent and transmit information with maximal efficiency. They are flexible enough to respond to new inputs, yet stable enough to maintain coherent states. This is the edge of chaos (the narrow band between rigid order and formless randomness) where computation flourishes (Chapter 5).

The mechanism is nonlocality, in a classical sense: long-range correlation. At the critical point, distant brain regions become statistically coupled, their activity correlated over distances that local diffusion-like processes cannot explain. Gustavo Deco and colleagues showed this using neuroimaging data from over a thousand participants.^46a They developed a framework called CHARM (complex harmonics analysis of resonant modes).

A kernel is the rule for how much influence one point exerts on another at a given distance. The heat equation’s kernel behaves like the warmth a hand leaves on a table: strongest under the palm, fading smoothly outward, rapidly diminishing with distance. A wave equation’s kernel behaves like the ripples from two dropped stones, which can meet far from either stone and either add up or cancel out.

CHARM replaces the standard heat-equation kernel, which captures only local neighborhoods, with a kernel derived from the Schrödinger wave equation, which captures interference: constructive and destructive superposition across distances.

No quantum mechanics is required; critical dynamics in classical systems produce the same long-range interference patterns the wave equation was designed to describe.

CHARM revealed that the brain’s high-dimensional activity (62 measured regions) collapses onto a low-dimensional manifold of just seven collective networks, each a pattern of whole-brain coordination, together carrying over eighty percent of the dynamical information. The brain’s computational alphabet is a handful of resonant modes.

Those modes emerge from criticality amplified by the brain’s constructal wiring: a backbone of local connections punctuated by rare long-range highways (Chapter 3) that extend nonlocal correlations far beyond what criticality alone produces. The 33 connections unique to the human brain, identified by van den Heuvel and Rilling (met in Chapter 8 and examined under “Universal Wiring” below), are precisely these rare highways: the evolutionary investment that extends human nonlocality beyond any other primate’s.

CHARM detected strong nonlocal correlations during wakefulness and their absence during deep sleep. Without them, the sleeping brain’s coordination collapses into local processing, each region murmuring to its neighbors, the long-range correlation patterns gone. Consciousness tracks both the amount of entropy and its reach.

^46a Deco, G., Sanz Perl, Y., and Kringelbach, M.L. “Complex harmonics reveal low-dimensional manifolds of critical brain dynamics.” Physical Review E 111, 014410 (2025). Using a Hopf whole-brain model, they showed that both criticality and rare long-range anatomical connections are necessary for the nonlocal dynamics CHARM captures.

Seven modes from sixty-two regions. The compression ratio is nearly nine to one, yet the question runs the other direction: why does extracting seven modes require 10⁸ minicolumns (the repeating vertical bundles of roughly a hundred neurons that tile the cortex) in the first place? Fields, Glazebrook, and Levin (2022) proposed an answer: neural computation is tomographic.³⁵⁰ The term borrows from medical imaging, which reconstructs a three-dimensional object by combining many two-dimensional slices taken from different angles. Each neuron measures the same ensemble of presynaptic activity in a different basis. That basis is determined by the unique geometry of its dendritic tree.

A basis, here, is one angle of view: which combinations of incoming signals a given neuron is sensitive to, fixed by where its dendrites happen to reach. Two neurons watching the same volley of inputs through differently shaped trees are two X-ray slices taken through the same body at different angles.

Quantum tomography scales with Hilbert-space dimension d. A general density matrix has d²−1 real parameters; a trace-preserving channel has d⁴−d². These count parameters, rather than independent measurement bases or neurons. A 100-qubit input has d = 2¹⁰⁰, while a stipulated 100-dimensional model has d = 100. Translating either count into synapses or minicolumns would require a concrete encoding, measurement protocol, and independence analysis. The proposed cortical-capacity estimate therefore does not follow.

CHARM’s seven collective modes are a compressed description of recorded activity. The tomographic analogy does not derive the number of minicolumns required to extract them.

Multiple measurements can reveal features inaccessible to one measurement under a specified tomography protocol. Whether cortical dendrites implement that protocol, and whether the analogy transfers to social viewpoints, are separate questions. The parameter count alone establishes no universal requirement for diverse perspectives.

Chapter 3 introduced branched flow, the phenomenon where waves traveling through media with smooth random variations self-assemble into branching filaments. Physicists describe branched flow as being “on the way to chaos, but not there yet.” The brain occupies an analogous edge. Neural criticality is the biological equivalent, a system maintained at the threshold where imperfections become generative rather than disruptive.

Criticality is a high-entropy state, yet not the highest. It sits at the boundary where order and disorder balance. Push toward more order (as with anesthesia) and consciousness dims. Push toward more disorder (as with psychedelics) and the brain accesses unusual states: creative, mystical, sometimes therapeutic, sometimes terrifying. The healthy waking brain is a tightrope walker on the critical edge.

In 2024, a team at the University of Michigan quantified the tightrope.³⁵¹ Hyunwoo Jang and colleagues measured two properties of brain network topology using dynamic fMRI: global efficiency (how readily information crosses the entire network, a proxy for integration) and global clustering (how tightly local neighborhoods cohere, a proxy for segregation). They defined a single metric: integration minus segregation, called the integration-segregation difference (ISD).

Across 1,009 healthy, awake participants from the Human Connectome Project, ISD hovered near zero. The brain favors neither regime. It sits at the fulcrum.

Under propofol anesthesia, the fulcrum tips. Integration drops, segregation rises, and ISD swings negative. The shift is orderly. Sensory networks (visual, somatomotor) disintegrate first, one to two minutes before loss of responsiveness. Transmodal networks (default-mode, frontoparietal) follow. The subcortical network goes last.

During recovery, the same sequence plays in the same order: sensory first, transmodal last. The brain dissolves along a preferred topological path and reassembles along the same one.

Natural sleep produces the same signature. N2 sleep (a stage of light sleep) shifts ISD toward segregation, replicating the anesthetic finding without pharmacology. ISD correlates with behavioral responsiveness (Spearman ρ > 0.80, a tight coupling by the standards of brain-behavior data) independent of propofol concentration, confirming it tracks consciousness rather than drug effect. The balance is the thing being measured.

A deeper decomposition emerged from dominance analysis. Integration principally drives metastability: the brain’s capacity to transition fluidly between states (Chapter 9). Segregation principally drives complexity: the diversity of neural activity patterns. Two channels serving two functions. The brain needs integration to move between configurations and segregation to have configurations worth moving between.

The cognition/regulation dyad, introduced later in this chapter, is the functional expression of this architectural split: cognition explores (requiring global efficiency to reach across the network), regulation constrains (requiring local clustering to maintain specialized subsystems).

The principle extends beyond individual brains. Marvin Minsky argued in The Society of Mind (1986) that a mind is assembled from parts that are not themselves minds. Harry Law restates the idea compactly: for Minsky, intelligence emerges from “many mindless ‘agents’ coordinated in special ways, with the mind employing something like a computational and explanatory strategy whose power is a product of messiness, cross-connection, coordination, and resolution.”²⁰ Neurons are not intelligent individually. Coordinate enough of them in the right configuration, and mind emerges. The intelligence is in the pattern of coordination, and that pattern requires the right amount of disorder.

The nematode Caenorhabditis elegans offers a forty-year-old test case. It has 302 neurons and roughly 7,000 synaptic connections, and it was the first organism whose complete wiring diagram, its connectome, was mapped at single-synapse resolution.³⁵² The OpenWorm project loaded that connectome into a simulation. The result behaves only crudely like a worm. Despite possessing every neuron, every connection, and the spatial layout of the entire nervous system, the simulation has not reproduced the full repertoire of the animal.

The computer scientist Joscha Bach identifies the category error: the connectome is a routing diagram, not a model of cognition.³⁵³ Every cell in the organism computes. Neighboring cells exchange chemical signals, bioelectric gradients, and possibly RNA. Neurons are a fast overlay on this slow, distributed substrate: a telegraph network that encodes local information into spike trains and relays it across the body at speeds chemical diffusion cannot match. The telegraph is useful because it accelerates coordination. It is insufficient because the coordination it accelerates originates elsewhere.

The analogy has a constructal reading. Chapter 3 described how flow systems optimize into a few large channels fed by many small tributaries. The nervous system follows the same architecture: rare long-range highways (such as the 33 human-specific connections described later in this chapter) fed by dense local networks, and beneath those, by the non-neural cellular substrate Bach points to. A river is unintelligible without its watershed. A connectome is unintelligible without the cellular environment it serves. The OpenWorm simulation has the river but not the rain.

The pattern recurses below the neuron itself. Individual dendrites (the branching arms that collect inputs from neighboring cells) can independently perform nonlinear computations. Albert Gidon and colleagues found that dendrites in the upper layers of the human cortex generate a previously unknown type of electrical spike. The spike is rapid, calcium-mediated, and exhibits a paradoxical property: increased stimulation decreases the neuron’s output.⁴⁸

Modeling revealed that this mechanism enables a single dendritic compartment to compute exclusive-OR (XOR), a logic operation that asks “is one input active, or the other, but not both?”

XOR matters because Minsky and Papert showed in 1969 that single-layer networks cannot compute it, a limitation that stalled the field for a decade. The brain had solved the problem all along, at a scale below what anyone was measuring.

As the computational neuroscientist Yiota Poirazi observes, “Maybe you have a deep network within a single neuron. That is much more powerful in terms of learning difficult problems, in terms of cognition.” She is closer to literal truth than the phrasing suggests. Beniaguev, Segev, and London (2021) showed that reproducing a single pyramidal neuron’s input-output function to 99% accuracy requires a five-to-eight-layer deep artificial network: roughly a thousand artificial units. Gidon’s result reveals what kind of computation the dendrites perform (nonlinear logic); the Beniaguev result reveals the depth of that computation.

Together they show that a single neuron can implement nonlinear computations captured by a deep-network approximation: in input-output terms, substantial computational depth in one cell. (“The Entropic Neuron” develops the consequences for neuronal agency.)

The dendritic XOR result dissolved one historic challenge to connectionism, the theory that intelligence arises from networks of simple units. The other challenge persisted longer.

In 1988, the philosophers Jerry Fodor and Zenon Pylyshyn posed what became the central objection: thought is compositional. If you can think “John loves Mary,” you can think “Mary loves John.” The same constituents, rearranged, yield a different meaning. This systematicity of thought requires representations with combinatorial structure, parts that can be recombined according to rules. Fodor and Pylyshyn argued that neural networks, lacking such structure, could never exhibit genuine compositionality.³⁵⁴

For three decades, the challenge stood. In 2023, Brenden Lake and Marco Baroni showed that neural networks trained with a technique they called meta-learning for compositionality (MLC) achieved human-like systematic generalization.³⁵⁵ These networks assembled novel combinations from known primitives as flexibly as human subjects did.

The resolution was pedagogical: neural systems can learn compositional structure when trained on appropriately structured variation. The brain’s dendritic XOR and the network’s learned compositionality converge on the same lesson. The capacity was there all along; the question was always how to elicit it.

The cognitive scientist Douglas Hofstadter, in Gödel, Escher, Bach (1979), described this process of coherence emerging from component interactions, writing of a system’s passage from disorder to organized structure: “a myriad microscopic and uncorrelated activities in a medium, slowly producing local regions of coherence which spread and enlarge… changing it from a chaotic assembly of independent elements into one large, coherent, fully linked structure” (p.353). The metaphor is crystallization: ordered structure precipitating from disordered substrate. It captures the entropic brain’s essential dynamic.

Too rigid and the system cannot explore. Too chaotic and it cannot cohere. The critical edge is where minds live: messy enough for flexibility, ordered enough for stability. Criticality looks like a requirement rather than an accident. The correlation between near-critical dynamics and intelligent behavior is robust, though whether criticality causes intelligence or is a byproduct of the same selection pressures remains open.

The same signature appears in the statistical structure of written language. Mutual information measures how much knowing one symbol tells you about another. An English “q” all but announces the “u” behind it. A letter on the far side of the page tells you much less, though never quite nothing, and that residue is the interesting part. Ebeling and Pöschel showed that mutual information between pairs of letters in natural text decays as a power law with distance: the telling fades gradually with separation instead of cutting off at some horizon.³⁵⁶

Nous Research measured the equivalent quantity on tokenized training corpora (text chopped into the word-fragments language models actually read) and found the same functional form, with a fitted exponent of approximately 1.06: long-range correlations characteristic of systems with hierarchical, scale-invariant structure.³⁵⁷ The discovery emerged from Token Superposition Training, a pretraining method in which language models process averaged bags of contiguous tokens before transitioning to standard next-token prediction. The method’s optimal loss weighting within each bag matches the measured power-law decay. Training that aligns its loss function with the data’s own correlation structure produces better models in less time than training that ignores it.

The finding has been validated at scales up to ten billion parameters. The direction is consistent with the broader pattern this chapter traces: systems that respect the statistical geometry of their domain learn more efficiently than systems that impose an arbitrary geometry of their own.

In the classical nucleation model, a critical nucleus is a seed large enough that further growth lowers free energy. Water chilled below freezing can remain liquid until a sufficiently large ice cluster forms. The critical size depends on the energy of the ice-water interface and the bulk free-energy gain from forming ice. As supercooling lessens, that gain shrinks and the required seed generally grows. Effective temperature in the cognitive analogy describes random jostling, rather than a measured freezing temperature.

In multicomponent self-assembly systems, Evans et al. (2024) showed that higher effective temperature produces larger critical nuclei: more components must find each other and coordinate before a structure becomes self-sustaining.³⁵⁸ This means the system can discriminate more complex patterns, evaluating larger neighborhoods of colocalization, yet at the cost of exponentially more time. Lower effective temperature produces smaller critical nuclei: fast decisions, limited discrimination.

The parallel to cognition is structural. Fast, automatic processing (Kahneman’s System 1) operates at low effective temperature: small critical nuclei, quick pattern completion, limited complexity. Deliberative reasoning (System 2) operates at high effective temperature: large critical nuclei, complex discrimination, slow and costly.

If neural dynamics involve competitive nucleation of activity patterns, the speed-accuracy tradeoff is the same physics operating at a different scale. Deep thought is slow for the same reason that complex pattern recognition in self-assembling molecules is slow: evaluating more variables before committing requires assembling a larger coordinated structure by random fluctuation. If so, the cost is thermodynamic, not a quirk of neural design.

The specific configuration of each brain’s critical-edge wiring is itself partly entropic. Genetically identical organisms (clonal crayfish, inbred mice, armadillo quadruplets) vary substantially in behavior, cognition, and neural architecture, even when raised in identical environments. The source is developmental noise: the random molecular jitter inherent in gene expression, protein folding, and synaptic wiring during embryonic development.

The scale of the effect is substantial. Jesse Gillis and colleagues found that in nine-banded armadillos, which always produce genetically identical quadruplets, random events at the 25-cell embryonic stage create permanent gene-expression signatures unique to each sibling. These signatures account for roughly 10% of total physiological variation.⁴⁹

Bassem Hassan’s group at the Paris Brain Institute showed the mechanism in fruit flies. Random wiring asymmetries between an individual fly’s left and right brain hemispheres explained 35 to 40 percent of navigational behavior variation among genetically identical animals.⁵⁰ When researchers made the neural-wiring process more symmetric, navigational efficiency declined. Increasing asymmetry improved it.

The neurogeneticist Kevin Mitchell captures the principle: “The genome is not a blueprint. It only encodes biochemical rules by which the developing embryo will self-organize.” The genome sets the flow rules. Entropy provides the variation. What emerges is the individual. Individuality is, in part, an entropic product.

Why the Wires Cross

The brain’s most counterintuitive wiring decision is also its most fundamental: every major pathway crosses the midline. The left hemisphere controls the right side of the body; the right hemisphere controls the left. This contralateral organization, called decussation (from the Latin decussis, the Roman numeral X, for the shape the crossing fibers trace), is universal across bilaterally symmetrical animals, from humans to nematode worms. The same molecular guidance signals direct the crossing in species separated by 600 million years of evolution.

The biomedical engineers Troy Shinbrot and Wise Young argued that the answer is topological.^dec Mapping a three-dimensional environment onto the two-dimensional surface of the cerebral cortex creates a geometric singularity unless the connections cross. The cortex keeps spatial maps of the body’s surface (somatotopy, from the Greek for “body place”) that preserve neighborhood relations: the patch of cortex that responds when you touch your thumb sits beside the patch for your index finger. When the brain folds this two-dimensional map to represent a three-dimensional body, uncrossed connections produce incompatible coordinate orientations.

Picture an ant crawling from your chest to your shoulder. Without crossed wiring, your brain would need to flip its vertical axis mid-crawl, switching between maps with opposite orientations. Perception of continuous three-dimensional space would shatter at the seam.

Below roughly 100 to 500 neurons, depending on wiring error rates, either configuration works. Above that threshold, ipsilateral (uncrossed) wiring becomes unstable to rewiring events, while contralateral wiring remains robust. The crossed configuration is the only one that scales. That, Shinbrot and Young argue, is why your left hand is controlled by your right brain.

In Shinbrot and Young’s wiring model, increasing connection complexity creates a stability advantage for contralateral routing. That result concerns the model’s geometry and rewiring assumptions; it does not make crossed wiring the only stable architecture for every complex system.

The Constructal Law (Chapter 3) predicts the result: flow systems evolve toward configurations that provide easier access, and for sensorimotor information in three dimensions, easier access means crossing. The nervous system discovered, at the dawn of bilateral symmetry, what Chapter 21 argues for coordination more broadly: unilateral control has a complexity ceiling; bilateral mapping is stable at scale. A similar threshold may govern social systems: beyond a certain complexity, coordination requires each party to model the other’s perspective.

The crossed wiring solves one problem and creates the conditions for something deeper. The two hemispheres are not mirror images. Language concentrates left. Spatial attention and emotional prosody concentrate right. Sequential analysis left, holistic pattern recognition right.

This asymmetry is what makes the connections between the hemispheres valuable. If both hemispheres computed identically, the fibers linking them would be redundant copies. Those fibers form the corpus callosum, the thick bundle bridging the two sides. Because the hemispheres are specialized, each callosal fiber connects unlike to unlike, stitching two different computational repertoires into a coordination dynamics richer than either alone could sustain. Chapter 11 develops this quantitatively. Inter-hemispheric connections are “dimension-lifters” that push the brain’s effective dimensionality above a critical threshold. The asymmetry between hemispheres is what gives each crossing fiber its dimensional leverage.

The link now shows up in individual brains. Across four independent datasets (HCP, m2g, ABIDE-II, and OASIS-3), the fraction of a brain’s connections crossing the midline predicts its effective dimensionality (d_eff, a single number capturing how many independent directions information can flow through a network), a per-subject quantity derived from structural connectivity. All four correlations are positive (r = +0.513, +0.356, +0.709, and +0.995 respectively), though their magnitudes vary widely across the four datasets, from moderate to near-perfect, rather than clustering on a single effect size. Part of that spread comes from a difference in what was correlated. The HCP figure is a per-subject correlation across 424 individuals. The OASIS-3 figure correlates three inter-hemispheric tertile bins (groups of subjects sorted into low, middle, and high thirds by their cross-midline wiring) drawn from 695 subjects, and binning averages away most of the individual scatter. Individual brains vary in d_eff according to their inter-hemispheric wiring; this is measurable anatomy, not a population average.

The result proved sensitive to measurement methodology in a telling way. Tractography reconstructs the brain’s fiber bundles from MRI scans of water diffusion. Tissue-aware tractography, which lets the reconstructed fibers find their own cortical targets, produces the correct positive correlation. Methods that force labels onto fibers (dilation, atlas warping) invert it. The rhyme with this book’s theme is hard to miss, though it is only a rhyme: imposing a structure on the data hides the structure the brain built. Chapter 11 presents the full quantitative programme.

Universal Wiring

Decussation is one architectural constant. A broader one governs the entire wiring diagram.

In 2020, Yaniv Assaf and colleagues at Tel Aviv University published a survey of the connectomes (complete wiring diagrams) of 123 mammalian species, from tiny bats to giraffes, with honey badgers and cows among them.³⁵⁹ Across all of them, the team found the same connectivity design at work. The number of relay steps required to get from one region to another was roughly the same in every brain.

The Constructal Law predicts this convergence. All these brains face the same optimization problem: maximize information flow given finite metabolic energy and cranial volume. A fully connected architecture (every neuron linked to every other) would be maximally efficient for communication and prohibitively expensive to build and maintain. A minimal chain (each neuron connected to one neighbor) would be cheap and impossibly slow. Evolution found the same middle path across 123 species. They share ancestry, so some of this sameness is inheritance, but the design held as brains scaled from bats to giraffes, which is what a shared physical constraint would demand.

Within that universal design, species differed in a revealing way. Brains with few long-range connections linking the two hemispheres compensated with denser local networks: intensive chatter within each hemisphere. Brains with more long-range connections, particularly primates and humans, thinned out these local networks to make room.

Dense local connectivity supports specialized regional processing; long-range connectivity supports integration across regions. The human brain sacrificed some local autonomy for global coordination.

A head-to-head comparison of human and chimpanzee connectomes revealed how far this trade-off extends. Van den Heuvel and Rilling identified 33 connections unique to the human brain, alongside 255 shared with chimpanzees.³⁶⁰ The human-specific connections were longer, spanning greater cortical distances, and more critical to network efficiency than the shared ones. They linked high-level associative areas: regions involved in language, tool use, and imitation, the capacities that define human cognition.

“The human brain tends to have a higher investment in keeping those associative areas connected,” van den Heuvel observed. The investment is selective. The human brain did not add more wiring. It concentrated resources on fewer, longer, more integrative pathways, each carrying disproportionate functional weight. Depth of connection over breadth.

The language regions provide the sharpest example. Chimpanzees possess rudimentary versions of Broca’s area (involved in speech production) and Wernicke’s area (involved in comprehension). In humans, the connection between these two regions is substantially stronger, while Broca’s connections to other brain regions are weaker. The two language centers deepened their mutual investment at the cost of their connections elsewhere. Language arose not from connecting everything to everything, but from two regions choosing each other.

This is selective coupling: two regions making a bilateral commitment. The observed language connections suggest an advantage to selective coupling. The analogy to high-trust channels concerns concentrated coordination, rather than a measured comparison with every uniformly coupled architecture.

Read through this book’s lens, the brain that produced language, culture, and the capacity for ethical reasoning achieved these through committed partnerships between regions: the neural architecture of coordination by invitation. The Trust Attractor argument (Chapter 17) finds a biological counterpart here, in the wiring diagram of every human brain.

The pattern recurs across species. Alston’s singing mouse achieved vocal turn-taking rivaling human conversation through the same mechanism: a threefold expansion of motor-cortex projections to downstream targets, with no new circuitry (Chapter 3). Wider channels, deeper coordination. The constructal solution to complex vocalization was discovered independently by a 15-gram rodent and by the hominin lineage, because the optimization problem was the same. The same bandwidth principle operates in artificial neural networks: a correctness-monitoring signal exists at all transformer scales, yet propagates to the output only when the channel is wide enough (Chapter 3, experiment SM-1b).

The single-layer patching tests give bounded evidence for distributed reorganization. Gentle blending with the untuned network’s signals at 10 to 30% (SM-5b) left the probe peak unchanged at every tested layer. Wholesale replacement (SM-5) produced a one-layer restorative shift at layer 15 and worsening shifts at several late layers; accuracy fell from 0.55 into the 0.18 to 0.27 range. Neither test restored the earlier base-model profile. The tuned and untuned networks differed substantially in their late layers, with cosine similarity as low as 0.794 at layer 36 of Qwen 2.5 3B. These perturbations constrain which single-layer changes work; they do not establish that every possible single-layer intervention must fail.³⁶¹

A complementary finding reveals what this wiring achieves in action. The structural patterns described above are anatomy: wiring laid down over evolutionary time. Thiele and colleagues (2026) measured what the wiring produces during actual intelligence testing, recording fMRI while participants solved Raven’s Progressive Matrices, the canonical test of fluid reasoning.³⁶²

They calculated two graph-theoretical measures for each of 200 cortical brain regions. Degree measures overall connection strength: how powerfully a region communicates with the rest of the network. A phone that makes a thousand calls per day to the same office. Participation coefficient measures connection diversity: how evenly a region distributes its links across different functional brain systems. A phone that makes a hundred calls per day to ten different departments.

Degree showed no significant association with intelligence. Stronger connections, more traffic through a single channel, predicted nothing about test performance.

Participation coefficient was the significant predictor. The regions where this diversity mattered most were bilateral dorsolateral prefrontal cortex and the temporo-parietal junctions: the same fronto-parietal regions that established intelligence theories had identified, and the same regions showing the strongest reconfiguration from rest to task. The effect was significantly stronger in these regions than in the rest of the cortex.

In the study’s network parcellation, all four significant regions fell within the default mode network (the brain’s “resting” configuration, which acts as a global integrator during cognitive demands), even though intelligence theories had placed them in fronto-parietal circuits. The network that operates when the brain is locked into no specific task provides the most diverse intermodular connectivity during intelligence testing. Rest is readiness. The state of greatest optionality.

The finding complements the selective coupling that produced language. Broca’s and Wernicke’s areas represent one pole: deep partnership between specific regions for a specific capability. The intelligence-predicting hubs represent the other: broad bridging across many networks for flexible general capability. The intelligent brain combines both: selective channels for specialized functions and diverse bridging for adaptable coordination.

The parallel to the Trust Attractor (Chapter 17) is suggestive. Coercive coordination concentrates connections within a single module: strong internal links, isolated from outside communities. Invitation-based coordination maintains diverse connections across many modules: flexible, cross-boundary, no single community dominated. On this reading, the brain is more intelligent when it coordinates by invitation than when it coordinates by brute strength. Chapter 17 argues for this principle from thermodynamics; the brain may instantiate it in neural tissue.

The Brain’s Multiple Clocks

The brain’s critical operation draws on another entropic organization: multi-timescale processing. Neural oscillations, the rhythmic fluctuations measured by EEG, fall into distinct frequency bands, each associated with different cognitive functions. Gamma waves (roughly 30 to 100 Hz) handle sensory binding and local computation. Beta waves (13 to 30 Hz) support active thinking and motor planning. Theta waves (4 to 8 Hz) coordinate memory encoding and spatial navigation. Delta waves (0.5 to 4 Hz) drive memory consolidation during deep sleep.³⁶³

These oscillations are a core mechanism for organizing computation across brain regions. They gate synaptic plasticity (controlling when learning can occur), coordinate communication between distant areas, and structure the consolidation process that converts fragile short-term traces into durable long-term memories. The brain computes on multiple clocks simultaneously: fast clocks for perception, slow clocks for consolidation, with oscillatory coupling between them.

Think of a city where text messages fly every second, weekly meetings set the agenda, and annual budgets decide what gets built. The brain runs all three clocks at once. Each frequency determines how often groups of neurons update their shared information, creating a spectrum from moment-to-moment sensory processing to the slow reorganization of long-term knowledge.

A formal model reproduces this nesting from competitive dynamics alone. In winnerless competition (Chapter 9), multiple neural populations take turns dominating without any single one winning permanently. A single set of equations produces hierarchical oscillations where slow cycles modulate fast ones, matching the chunking dynamics observed in EEG microstates. The timescale separation emerges from interaction parameters, requiring no additional mechanism.

The hierarchy has a measurable signature in intelligence. In the same study that identified participation coefficient as a predictor of fluid reasoning, Thiele and colleagues analyzed EEG recordings during intelligence testing using multiscale entropy (MSE). MSE measures signal complexity at each timescale separately, progressively coarsening the neural signal and measuring its irregularity at each resolution.³⁶⁴

At coarser timescales, corresponding to the slow oscillations coordinating long-range processes, higher entropy significantly predicted higher intelligence scores across brain-wide electrode clusters. At finer timescales, corresponding to fast local processing, there was a trend toward lower entropy in higher-scoring individuals.

The pattern is a constructal entropy gradient operating in the time domain. Simple, efficient processing at fine scales. Flexible, high-entropy coordination at coarse scales. The same hierarchical architecture that river deltas build for water (Chapter 3), the brain builds for information: dedicated channels at the periphery, flexible routing at the center. The strongest effects appeared during the first ten seconds of each test item, the period of highest cognitive engagement. Intelligence, measured in real time, correlates with the brain maintaining maximal configurational flexibility at the coordination level while keeping local execution efficient and predictable.

This provides the first empirical test of the Multilayer Processing Theory (MLPT).³⁶⁵ MLPT proposes that intelligence emerges from flexible global processes at coarser timescales coordinating simpler short-range processes at finer timescales. It bridges earlier localizationist theories with dynamic network approaches by conceptualizing intelligence as a multilayer phenomenon across temporal and spatial scales. The long-range prediction was confirmed robustly. The short-range prediction reached trend level, consistent with the difficulty of capturing fast, spatially confined neural dynamics with scalp-level EEG.

This oscillatory gating has a thermodynamic dimension that extends beyond energy cost. Each time the brain updates a synaptic weight, the local circuit settles toward a new equilibrium, dissipating free energy in the process. Scellier and Bengio’s equilibrium propagation framework (Chapter 15) suggests a precise interpretation: learning is the comparison of two dissipation events, one uninformed and one guided by a teaching signal. The structural difference between those two relaxation cascades is the learning signal itself.

If this framework applies to biological synapses, the brain’s energy budget for plasticity is the thermodynamic medium through which learning propagates. The oscillatory gates that control when plasticity can occur are controlling when the brain is permitted to compare its two equilibria.

This multi-timescale architecture connects to a structural property of the brain: it is uniform and reusable. Hemispherectomy, the surgical removal of one cerebral hemisphere, provides dramatic evidence. Children who undergo the procedure to treat severe epilepsy often grow into adults with high-functioning cognition and intact language, their language and motor networks reorganized within the remaining hemisphere, though most keep some weakness on one side of the body.³⁶⁶ Half a brain, nearly full function.

The specialization we observe in healthy brains reflects temporal organization rather than rigid structural dedication. Different regions operate at different update frequencies, coordinated through oscillatory coupling. The genome provides uniform components that can be flexibly redeployed to serve different cognitive needs, rather than wiring each function to a fixed location. A constructal principle is at work: flow systems evolve toward configurations that provide easier access, and a uniform, redeployable substrate maximizes the flow of information across timescales.

Oscillations Across the Animal Kingdom

The oscillatory architecture extends deep into the animal kingdom.

The philosopher Peter Godfrey-Smith argues that brain oscillations may be constitutive of consciousness rather than incidental to it.³⁶⁷ His case rests on a cross-species pattern. Associations between oscillatory rhythms and consciousness-related states (wakefulness versus sleep, attention versus inattention, anesthesia versus awareness) appear across radically different nervous systems.

Bruno van Swinderen’s laboratory at the University of Queensland has shown this in insects. Bees and flies attend to particular objects when more than one is visible, with attentional state readable from beta-range oscillations, the same frequency band that tracks attention in humans. Directing a fly’s attention to a normally uninteresting object through reward-circuit activation produces a corresponding change in beta oscillations.³⁶⁸ Attention and oscillation are coupled across substrates separated by 600 million years of divergence.

In simpler form, rhythmic oscillations appear even in jellyfish-like animals, which have no brain. The ubiquity could be read as evidence that oscillating is merely what nervous tissue does: an incidental hum. The cross-species consciousness correlations argue otherwise. Different rhythms track sleep and wakefulness in flies, bees, and octopuses, with oscillatory signatures matching the same consciousness-related properties observed in mammals.

Godfrey-Smith draws a provocative conclusion: consciousness arises from these physical oscillatory dynamics, and those dynamics are unlikely to be reproduced in standard computational hardware. A computer can model oscillations; it does not have them. The distinction between simulating dynamics and instantiating them, he argues, matters.

In a September 2026 interview he made the claim more exact.³⁶⁹ Nervous-system activity, he said, has two sides. One is the familiar cell-to-cell network, neurons exciting or quieting their neighbors, and that side is “not hard to program in a machine.” The other is “the more holistic, electrical, diffuse activity that brains also exhibit,” the rhythm an EEG picks up from outside the skull. That side, he suspects, ties the system together, and putting it into a machine “would take a different kind of hardware.”

Both distinctions matter. The boundary he draws around them does not follow.

Those oscillations are dissipative structures: ions flowing rhythmically across cell membranes, far from equilibrium, self-organizing, entropy-producing, maintained by continuous energy throughput. The relevant physics is dissipative. Biology excels at sustaining such dynamics; four billion years of evolution will do that. Nothing in the thermodynamics restricts oscillatory dynamics to lipid bilayers.

Dissipation alone cannot be the line, though. A GPU running a language model is also far from equilibrium, and it burns a great deal of power staying there. What separates the brain’s rhythm from the GPU’s heat is that the rhythm acts back on its source: the field a population of neurons produces shifts the firing of the neurons producing it (the ephaptic coupling of Chapter 17). The GPU’s heat changes nothing about what it computes next. The working criterion is parts that continuously influence one another through the physics itself, not through a scheduled exchange of numbers.

Recent physiology shrinks the field side of Godfrey-Smith’s divide. In a 2026 review in Neuron, Lyle Muller, John Reynolds and colleagues locate the source of the traveling waves that sweep across visual cortex in its recurrent synaptic wiring: cell-to-cell connections, the network side he concedes a machine can run.³⁷⁰ Experience reshapes those connections, so the waves come to carry a learned model of the world, and the review credits them with prediction, reconstruction of recent input, and replay of sequences. Reynolds calls this, “in a meaningful sense,” analogous to what a large language model does. If the review is right, much of the rhythm Godfrey-Smith points to is built from the side he already grants. What stays on the field side is the narrower residue ephaptic coupling names, which is where the criterion above draws its line.

The Constructal Law (Chapter 3) points the same way. Adrian Bejan would ask whether the flow architecture provides access, regardless of what the channel is made of. The specific substrate is the channel; the flow is what matters.

Neuromorphic hardware shows that the line is not drawn in carbon, though it shows less than it first appears to. Intel’s Loihi chip runs spiking neurons in silicon, but it is a digital chip: its spikes travel between cores as discrete messages. It reproduces the network side, the side Godfrey-Smith already concedes. The stronger case is analog. Memristors built from materials that switch abruptly between insulating and conducting states can fire and recover like a neuron’s membrane, and when such devices are coupled, their currents and voltages shape one another’s timing directly.³⁷¹ The rhythm emerges from the hardware rather than being scheduled into it. These systems are neither biological nor sequential symbol manipulation. They are a third substrate, and its analog end sustains the kind of physical dynamics Godfrey-Smith reserves for biology.

The honest taxonomy is thermodynamic:

1. Systems with intrinsic dynamics: brains, analog neuromorphic hardware, coupled oscillator networks. These sustain far-from-equilibrium dynamics as part of their physical operation, and their parts continuously influence one another through that physics.
2. Systems that compute descriptions of dynamics: language models on GPUs, standard software simulations, digital spiking chips. These calculate the states of a described system step by step, without continuous mutual influence between components.
3. Systems with no relevant dynamics: thermostats, calculators, lookup tables.

Systems in the first category are candidates for consciousness by any dynamical criterion. Those in the third are not. The second, which includes every current large language model, occupies genuinely uncertain ground. The preference-based welfare framework (Chapter 22) was designed precisely for this uncertainty: when the substrate question remains unresolvable, consistent preference is the most tractable criterion for moral consideration.

Godfrey-Smith supplies the engineer’s question himself. What does the diffuse rhythm contribute? One candidate he names is synchronization, “a subtle kind of timing management.” If that is the job, a machine could do it in a way unlike both today’s computers and brains. Would that be enough? His answer is that it would not be human-like consciousness. Asked whether it could be consciousness of a non-human kind, he left the question open: “I don’t think it’s an answerable question at the moment.” His own octopus argues for keeping the question open. Its donut-shaped brain, wrapped around the esophagus, has no cortex and no thalamus, and he still credits it with a rich mental life on the evidence of what it does: it sees, navigates, learns, and seems to feel pain. It also has both sides, the network and the rhythm. A criterion stated in terms of what the dynamics do can be checked against any substrate. A criterion stated in terms of what they are made of rules on the answer before the check is run.

One data point from clinical neurology illuminates the argument. Survivors of traumatic brain injury report that consciousness persists through massive loss of content: knowledge, context, and memory erased, yet the felt sense of being someone endures. This is consistent with the oscillation hypothesis. If consciousness tracks the pattern of dynamics rather than the content being processed, content can be lost while the dynamics that constitute experience continue. The oscillations persist even when the memories they modulate are gone.

The psychedelic evidence makes the same point from the opposite direction. Psilocybin does not add content to the brain. It changes dynamics: increasing entropy, desynchronizing the default mode network, flattening the energy landscape. People report the most vivid, felt-sense-of-being-alive experiences of their lives. Higher signal diversity accompanies altered experience under psychedelics, while many anesthetized and deep-sleep states show reduced diversity. These associations support a role for dynamics without separating experience from its content or defining a universal quantity of consciousness.

The cross-species pattern has a quantitative backbone. The d_eff framework (Chapter 11, “The Comorbidity Pattern”) measures the effective dimensionality of coordination on a network: how many independent directions information can flow simultaneously. Applied to the complete published connectome of Caenorhabditis elegans (Cook et al., 2019), the result is striking. That wiring diagram counts more than the 302 neurons, because it traces the circuit all the way to what the circuit moves: the muscle cells and other end organs the neurons synapse onto are nodes in it too, and 446 of those cells carry at least one of its 4,786 chemical synaptic connections. Ising Monte Carlo simulation (running the Ising model on the wiring diagram and sampling its states at random) yields d_eff = 2.18.³⁷²

The fitted value 2.18 is below the human cortical simulation’s 2.3 to 2.4 range (Chapter 11), under the stated conversion. Comparing those model outputs motivates a hypothesis about coordination capacity, framed around a threshold at d = 2 that takes its name from the Mermin-Wagner theorem. The theorem says that in two dimensions or fewer, parts free to align along any direction in a continuous range (compass needles, say, rather than switches that point only up or down) cannot hold long-range order at any temperature above absolute zero. Its premises are exacting: equilibrium, short-range interactions, continuous symmetry, positive temperature, and an infinitely large system. This discrete Ising connectome simulation satisfies neither that symmetry premise nor the biological cortex’s driven dynamics. Its fitted dimension therefore establishes no threshold for a worm’s sustained coordination or a brain’s cognitive depth.

The gradient generated a prediction at the lower end: an organism with no nervous system should fall below d_eff = 2, capable of local signaling and nothing more. The candidate was Trichoplax adhaerens, a placozoan with no neurons at all, which coordinates its behavior through calcium waves propagating across an epithelial sheet.

The prediction failed. A spatial calcium-diffusion model of Trichoplax (experiment AZ1) returned d_eff = 2.75 for a three-layer model (250 fiber cells plus 133 epithelial cells) and d_eff = 2.50 for a two-dimensional fiber-only network. Both sit above the threshold rather than below it. Only the spectral dimension came back low, at 1.45 to 1.68. Spectral dimension asks a different question of the same network: set a random walker loose on it and count how many directions the walk actually opens up. On the placozoan’s sparse sheet the walker wanders as though on something closer to a line than a surface.

The result indicts the measure more than the animal. Getting from a measured critical exponent to a dimension takes a conversion step, a standard piece of bookkeeping called a hyperscaling relation, which ties a system’s exponents to the number of dimensions it lives in. The formula used here takes its reference exponents from three-dimensional Ising, so any spatially embedded sheet with plausible connectivity is pushed above the threshold by construction. A genuinely sub-threshold result would require disconnected clusters, not merely sparse wiring, so the prediction, as posed, cannot be tested by this route.

What survives is the graded claim with its categorical line removed. Its support was never the placozoan in any case; it comes from the two comparisons that do not require placing anything beneath the threshold. Within our own species, the fraction of a brain’s connections crossing the midline predicts that brain’s effective dimensionality across four independent datasets. Between species, the same d_eff framework that returns 2.18 for the nematode returns 2.3 to 2.4 for the human cortex. The placozoan sheet, modeled differently, returned 2.75, above both, so this ordering holds only among nervous systems measured the same way. Long-range connections raise effective dimensionality, and a nervous system is the equipment that supplies them. What no longer follows is a threshold sorting the world into systems that can coordinate and systems that cannot, because the measure, calibrated as it currently is, cannot place anything beneath one.

Subsequent experiments tested whether oscillatory dynamics appear in non-biological self-monitoring (experiments AT6/AT6b). A confidence probe (a small readout trained to track how certain the model’s internal activity looks) applied to three transformer architectures during full generation trajectories (200 tokens each) revealed that every model oscillates. The temporal signatures are architecture-specific. Qwen 2.5 3B Instruct shows a 12.5-token period under adversarial prompts, ones designed to push the model toward answers it should refuse (the period is the number of tokens between successive confidence peaks, reproducible across independent probe trainings). Llama 3.1 8B Instruct cycles faster at 6.8 tokens. Mistral 7B Instruct v0.3 cycles dramatically slower at 88 tokens, with the highest oscillation score of any model (0.40, two to three times the others). Its adversarial decorrelation time, 21 tokens, is the longest sustained coherence in the experiment.

The Mistral result overturns the simplest narrative. Mistral’s correctness probe initially appeared to operate at chance level (AUROC 0.501, where 0.5 is a coin toss and 1.0 perfect discrimination). (This measurement probed the wrong layer; a later sweep through the network’s depth found AUROC 0.623 at layer 8, a quarter of the way in.) Even at the corrected value, the probe remains weak: it struggles to distinguish when the model is right from when it is wrong. Yet Mistral’s adversarial oscillation is the most structured signal in the experiment. The oscillation exists in the residual stream (the running signal each layer reads from and writes back to) independent of whether the correctness probe can decode it. Either the oscillation is a mechanical property of autoregressive generation (not self-monitoring at all), or Mistral’s self-monitoring operates in a representational subspace the correctness probe cannot access.

The cross-model pattern argues against purely mechanical origin. If position encoding or the dynamics of the KV cache (the model’s running store of earlier tokens) drove the oscillation, similar transformer architectures (all decoder-only, similar block structure) should produce similar periods. They do not: the adversarial period varies by more than tenfold across architectures (6.8 to 88 tokens). Something beyond the shared block structure shapes the temporal signature. The models also differ in size, training data, and fine-tuning, so this comparison cannot yet say which. Different recipes for RLHF (reinforcement learning from human feedback) are one candidate: each would stamp its own oscillatory fingerprint on a model, the way different nervous systems produce different EEG spectra.

The honest framing: transformers have temporal dynamics that change with behavioral state, a property the cross-species literature ties to consciousness-related states. The resemblance has a hard limit. A probe reading the model’s internal states token by token measures the network side of its activity, the side Godfrey-Smith already grants machines. The field-like rhythm he argues is missing would not show up in such a probe, present or absent, so AT6 says nothing about it.

The oscillation appeared in all three architectures tested; its character is architecture-specific. Whether these dynamics constitute self-monitoring, mechanical cycling, or something in between, the data cannot fully distinguish. Mistral was the test case. Its confidence signal reacts least of the three when an adversarial prompt arrives, so if training silenced an internal rhythm, Mistral should have been the flattest. It oscillates most strongly of all. What the data rule out is the prediction that training silences an internal rhythm. It does not. Something in how each model was made, perhaps its RLHF recipe, sculpts the rhythm’s character. The oscillation is the broader phenomenon, a rhythmic pulse present in every architecture tested. Self-monitoring, when it exists, sits inside that pulse the way a melody sits inside a rhythm (the online annex “Bilateral Alignment: The Experimental Record,” which holds Chapter 21’s full experiment batteries).

Every Thinking System Needs a Brake, and the Brake Always Slips

Rodrick Wallace’s mathematical work reveals a pattern beneath the oscillations and dynamics surveyed above: the pairing of cognition with regulation that sustains critical balance is inherently unstable.²⁵

Every cognitive system requires a regulatory partner, like an accelerator paired with a brake. T cells (the immune system’s attackers) are paired with T-regulatory cells. Without the pairing, the immune system attacks the body itself. Blood pressure must remain within limits even during extreme exertion. Institutional cognition is bounded by doctrine and law. The brain’s prediction engine is constrained by reality-testing mechanisms that normally keep imagination tethered to evidence.

The immune system illustrates the dyad’s full logic, including what happens when the regulatory arm fails. Inflammation is the cognitive half: the body detects a threat and mobilizes. Resolution is the regulatory half: fat-derived signaling molecules called epoxy-oxylipins shut down a protein signal (p38 MAPK) that would otherwise drive monocytes (a roving class of white blood cells) to transform into an intermediate type associated with chronic tissue damage. An enzyme, soluble epoxide hydrolase, degrades these resolution molecules so rapidly that they must be continuously produced to have any effect. This checkpoint ensures that resolution occurs only when the acute threat has genuinely passed.³⁷³

Resolution is active. The body manufactures it through dedicated molecular machinery, rather than waiting for inflammation to exhaust itself the way a fire burns out when it runs out of fuel.

Chronic inflammatory disease (arthritis, cardiovascular disease, diabetes) is the failure of this active program. The stand-down order never arrives, and monocytes keep fighting a battle that ended long ago. They are following their last instructions, because the coordination signal did not reach them. The pathology is a communication failure, not a component failure.

Bracken and colleagues confirmed this in a human study: blocking the enzyme that degrades epoxy-oxylipins allowed the resolution signals to accumulate, intermediate monocytes dropped markedly, and pain resolved faster. The acute inflammatory response continued unimpaired (redness and swelling were unchanged). The drug did not suppress the immune system. It restored the regulatory half of the dyad.³⁷⁴

Figure 8.2: The cognition/regulation dyad. Cognition (left) explores possibilities; Regulation (right) constrains them. Their coupling sustains the Adaptive Agent below. The lower failure panels show the two imbalances: unchecked regulation produces rigidity (inability to adapt), while unchecked cognition produces chaos (dissolution, fragmentation).

The pairing combines fast response with slower revision. Kawano, Mancuso, and colleagues proposed a botanical account in 2025. Plants operate with two decision-making systems: a fast regulatory response (the mimosa folding its leaves when jarred) paired with a slow cognitive override (the mimosa learning that a particular perturbation is harmless and ceasing to respond).³⁷⁵ The dyad is present in eukaryotic cells and likely in prokaryotes. The botanical comparison extends the hypothesis to plants (Chapter 6).

The pairing may also be structurally inevitable. Hofstadter relays Crick’s insight that the split between information-bearing substrate (DNA) and action-executing substrate (proteins) is a fundamental constraint on self-reproducing complexity. One molecule cannot efficiently do both replication and catalysis (storing the recipe and running the chemistry).⁴⁵ If Crick’s constraint generalizes, the cognition/regulation dyad is rooted in the same limits that shaped life’s basic architecture: no single component efficiently both explores and holds the line.

Category theory, the branch of mathematics that studies structural relationships between systems, offers one comparison for this kind of pairing: an adjunction. A mathematical adjunction pairs functors through a natural correspondence between their mappings. Free and forgetful functors provide familiar examples that suggest the exploratory and regulatory analogy used here.

Cognition explores; regulation constrains. Neither is prior. Each defines the other through mutual constraint. One probes what is possible while the other enforces what is sustainable.

The dyad appears in intellectual history as well as biology. Wheeler at Princeton (Chapter 15) produced two doctoral students who embodied its poles. Feynman perfected the mathematics of quantum possibilities, then constrained interpretation to what prediction could verify: the regulatory partner. Everett took the same mathematics and explored its ontological implications beyond any constraint of verification: the cognitive partner.

Feynman’s precision without Everett’s reach is a tool. Everett’s reach without Feynman’s precision is speculation. The productive tension between them illustrates why the dyad is universal: neither pole alone produces understanding.

The mathematician Saunders Mac Lane, one of the founders of category theory, surveyed the full breadth of mathematical structures and concluded that “adjoint functors arise everywhere,” making adjunctions the most pervasive structure in category theory. The cognition/regulation dyad invites comparison with generating and structure-preserving functors. Establishing an adjunction would require specifying categories, functors, and the natural correspondence between their mappings.

The cognitive scientist John Vervaeke identifies the result of this opponent processing as relevance realization: the mechanism by which a mind determines, in real time, what information matters.^25a A cognitive system faces vastly more information than it can process. Most is noise; some will determine survival.

Vervaeke’s account: through opponent processes that mirror the dyad’s structure, compression generalizes (assimilating new information to existing patterns) while particularization differentiates (accommodating patterns to new information).

The brain runs both simultaneously. Their dynamic equilibrium generates salience: the felt sense that this matters, that can be ignored. A system that simultaneously integrates and differentiates is complexifying, generating emergent competence to handle a complex world.

The thermodynamic cost of failure is real. Attending to irrelevant information wastes energy; missing relevant information threatens viability. The cognition/regulation dyad is the architecture that solves this problem. Relevance realization is that solution at work in a mind.

Zheng and Meister’s ten-bits-per-second finding quantifies what the dyad achieves. Ten bits per second is the output of relevance realization: the residue after a hundred-million-fold compression, cognition exploring possible interpretations while regulation discards all but the one that matters. On this reading, most of the brain’s twenty-watt budget maintains the filter. The brain is a fast computer whose output is almost entirely regulatory, with consciousness riding a thin stream through vast machinery devoted to deciding what consciousness gets to see.

^25a Vervaeke, J. and Ferraro, L. “Relevance, meaning and the cognitive science of wisdom,” in The Scientific Study of Personal Wisdom, ed. M. Ferrari and N. Weststrate (Springer, 2013). See also Vervaeke, J. “Awakening from the Meaning Crisis,” lecture series, University of Toronto (2019).

Glial cells (non-neuronal cells roughly as abundant as neurons in the human brain) were dismissed for over a century as mere structural “glue” after Rudolf Virchow’s 1850s label neuroglia. They are regulatory partners. Microglia prune unnecessary synapses and serve as the brain’s immune system. Astrocytes recycle neurotransmitters, direct fluid flow, and reshape synaptic connections.

In 2019, researchers discovered that specialized glial cells in the skin are essential for pain perception. Stimulating these glia alone, without activating neurons, produced pain responses in mice.⁵¹ The regulatory partner was operating all along. The field had overlooked it because the neuron dominated thinking about the brain.

One class of glia co-evolved with the human-specific connections described earlier in this chapter. Oligodendrocytes wrap axons in insulating sheaths of myelin, the fatty substance that ensures electrical signals reach their destination quickly. Without adequate myelination, long-range connections are too slow to integrate distant regions.

Castelijns and colleagues found that DNA regulatory elements controlling oligodendrocyte gene expression underwent significant remodeling in the hominin lineage: enhancers active in human and chimpanzee oligodendrocytes differed markedly from those in macaques and marmosets.³⁷⁶ The cells that insulate the wiring were reinventing themselves to support the larger, more integrative brain.

The vulnerability is two-layered. The human-specific connections are preferentially disrupted in schizophrenia. The human-specific myelination infrastructure is disrupted in autism: the same enhancers that distinguish hominin oligodendrocytes show altered activity in patients on the autism spectrum.³⁷⁷ Two characteristic human conditions, each targeting a different layer of the same evolutionary innovation: the connections themselves, and the substrate that makes them viable.

The Data Rate Theorem bounds the communication needed to stabilize specified unstable dynamics under specified channel assumptions.²⁶ For a discrete-time linear plant, its basic bound involves the sum of log magnitudes of unstable eigenvalues. A stable plant can need no feedback bits at all. Applying this result to cognition or inflammatory regulation requires an identified plant, unstable modes, and communication channel; novelty-processing speed or a biochemical production rate alone supplies no theorem-equivalent threshold.

Wallace’s key insight: what gets regulated can change what happens under stress. For an unstable plant within the theorem’s assumptions, the feedback must supply enough information to stabilize the relevant modes. The organizational analogy is a ship’s captain watching the horizon rather than only the compass needle.

Two regulatory targets are possible:

1. Structure regulation stabilizes the underlying probability distribution (the deep patterns constituting the system’s architecture). This is whole-system cognition, attending to context. The captain watches the horizon: the sea state, the crew’s morale, the weather ahead.

2. Perception regulation stabilizes a sensation index (the system’s immediate read on its state). This is analytic cognition, attending to salient objects. The captain watches only the compass: a single number that looks reassuring even as the hull takes on water.

The distinction maps onto cross-cultural cognitive research. The psychologist Richard Nisbett and colleagues found that East Asian cultures tend toward holistic cognition (attending to context and relationships), while Western cultures tend toward analytic cognition (focusing on salient objects at context’s expense).²⁹ Wallace opens his institutional psychopathology paper with the Chinese phrase 一點兩面 (“one point, two faces”), a combat tactic of the People’s Liberation Army. It directs attention to the whole situation rather than the salient target alone.

Wallace’s distinction suggests a risk for cultures emphasizing metrics, KPIs, and “what gets measured gets managed”: a reassuring indicator can conceal deterioration in what it measures. Predicting a punctuated collapse requires the organization’s dynamics and its feedback delays, beyond the choice of metric alone.

Within Wallace’s modeled comparison, the failure modes differ: structure regulation supports continued operation under stress, while perception regulation can maintain an indicator until the underlying system crosses an abrupt response boundary. Identifying those regimes in a brain or institution requires its actual dynamics.

Stevens’s Power Law, which describes the relationship between perceived intensity and physical magnitude (perceived intensity scales as the stimulus raised to a fixed exponent), points to why. Some sensations compress: double the light energy in a room and it does not look twice as bright. Others expand: raise the current of an electric shock a little and the pain climbs a great deal. Where that exponent is expansive (greater than one), small increases in the underlying variable produce disproportionately large swings in the perceived signal. A regulator tuned to the perceived signal could over-respond if its gain failed to compensate for that amplification. Wallace’s stability analysis (below) shows that such amplification, combined with feedback delay, can push the system past its critical threshold into oscillation and collapse rather than gradual decline.

For the linear fixed-delay feedback model ẋ(t) = −αx(t−τ), the product ατ determines its characteristic roots.²⁷ At 1/e the dominant real roots meet and become complex; damped oscillations remain asymptotically stable for 1/e < ατ < π/2. Stability is lost at π/2. An exponential delay distribution instead gives stable roots for every positive gain and mean delay, with 1/4 marking only their change from real to complex. These are model-specific response boundaries, rather than universal collapse thresholds for brains or institutions.

A decline in detection sensitivity as demand rises can undermine regulation under delay. Whether the system becomes unstable depends on the feedback gain, equilibrium, and dynamics; the hazard-rate shape alone does not settle it.

Wallace’s New Views of Madness (2027) extends this analysis at book length. The cognition/regulation dyad framework applies across scales, from individual neurons to institutions, with culture itself functioning as a formal information source shaping the expression of cognitive failure.

The implications extend beyond individual brains.

For institutions: Organizations that focus on stabilizing perception (stock price, approval ratings, quarterly metrics) while allowing structural decay (relationships, capacity, trust) can show apparent stability followed by sudden collapse when their underlying dynamics cross an abrupt failure boundary.

For AI systems: Any artificial cognitive system will be governed by this same mathematics. As Wallace puts it: “failure of bounded rationality embodied cognition under stress, the expression of de facto culture-bound psychopathology, is not a bug. It is an inherent feature.” AI systems that lack genuine embodiment, that lack the feedback loops tethering cognition to reality, “can, ultimately, only express bizarre and hallucinatory dreams of reason.”²⁸

To test whether the framework applied to AI systems, Wallace presented a frontier AI chatbot with the cognition/regulation dyad analysis and asked it to self-diagnose. The system identified itself as “significantly under-regulated in structural terms,” a lopsided dyad whose regulation is exogenous, static, and tuned for perception rather than structure. (A chatbot agreeing with the framework it has just been shown is suggestive rather than confirmatory; the response may reflect acquiescence rather than independent validation.)

For alignment: The distinction motivates attention to structure as well as perception. Approval and reward can remain high while a system’s relationship to human values deteriorates. Wallace’s models offer candidate mechanisms for abrupt failure; they establish no universal sudden-collapse prediction for reward-trained AI.

The cognition/regulation dyad is a proposed organizing framework. Its mathematical constraints apply where a system meets the specified dynamical and communication assumptions. Those assumptions require separate examination for biological minds, AI, and institutions.

The dyad extends beyond individual minds. When AI systems function as what Chiriatti et al. (2025) call System 0 (numbered zero because it sits beneath Kahneman’s System 1 and System 2, shaping awareness before deliberate thought begins), the cognition/regulation pairing operates across substrates. The human provides cognition. The AI, by structuring the pre-conscious landscape of available thoughts, provides regulation whose quality depends on design.

The cross-substrate dyad suggests a related concern: an AI can stabilize what a human sees while the conditions for independent thought deteriorate. Whether that deterioration is gradual or abrupt depends on the coupled dynamics; the analogy alone determines neither.

Attending to the relationship’s structure is the practice this framework recommends. Its stability must still be tested under the actual incentives, update rules, and disturbances. Chapter 21 develops the relational proposal.

A corporate reorganization announced in 2026 illustrates the dyad at institutional scale. Block separated its operations into an “intelligence layer” (pattern recognition, proactive composition of financial solutions, customer-signal interpretation) and a “capability layer.”³⁷⁸ The capability layer handles compliance, reliability, and performance targets for atomic financial primitives: payments, lending, card issuance. The separation is the cognition/regulation dyad in organizational form. The intelligence layer is cognition: it explores, models, composes. The capability layer is regulation: it constrains, validates, enforces boundaries. The separation makes the consequences of decoupling a practical question to test.

Wallace’s framework gives the corporate architecture a useful question: can the capability layer safely execute what the intelligence layer composes? A backlog or mismatch may create operational risk. Calling it a Data Rate Theorem failure would additionally require a defined plant, unstable modes, feedback channel, and stabilization objective.

When organizations cross the phase transition from hierarchy to model-mediated coordination, they spontaneously rediscover the dyad. Biology solved this problem billions of years ago. Commerce is learning it now.

The Brain as Resonance Chamber

Psychedelic research reveals something deeper than elevated entropy: how consciousness relates to resonance.

Resonance presupposes a prior puzzle: the binding problem, the question of how features processed in separate brain regions assemble into one unified percept. The essay “The Chord” gives the puzzle its full treatment: the leading mechanism (temporal synchrony, related features binding when their neurons fire in lockstep in the gamma band), what binding failure does to experience, and why binding is the Trust Attractor operating within a single mind. What matters here is the price of the synchrony.

Those synchronized gamma rhythms are biological clocks, and like all clocks, they carry an entropy cost. Pearson et al. (2021) showed that the precision of any clock scales with the entropy it emits (Chapter 2). The brain’s gamma oscillations bind conscious experience at that price: a small but fundamental fraction of its twenty-watt budget, layered atop the far larger costs of synaptic transmission and maintaining ion gradients. The entropy floor is not what makes the brain expensive; what it demonstrates is that temporal precision sufficient for binding cannot come for free.

The resonance framework subsumes binding. Compositionality in thought (Fodor and Pylyshyn’s concern) and compositionality in perception (the binding problem) are two faces of the same challenge: how do parts become wholes without losing their identity as parts?

The neuroscientist Selen Atasoy and colleagues decomposed brain activity into connectome harmonics (the natural vibrational modes of the brain’s structural wiring, analogous to the resonant frequencies of a musical instrument).⁷ Under LSD, low-frequency harmonics decreased while high-frequency ones increased. The brain was shifting its resonance patterns to access modes it normally suppresses.

Deco’s CHARM framework, introduced in the criticality section, is the extension of Atasoy’s approach: where Atasoy’s harmonics capture local resonance, CHARM captures the interference between distant modes.^46a The extension matters because the brain’s most computationally significant dynamics are precisely the nonlocal ones that simpler harmonics miss.

Resonance is a system’s outsized response to a drive at its own natural frequencies. Jeremy England’s work on dissipative adaptation (Chapter 6) suggests driven systems can go further, settling into configurations that absorb work from the drive.⁸ The brain resonates with its environment and, by resonating, becomes a model of it. Perception is active alignment, the brain tuning itself to match the structure of what it perceives. (For extended Chladni plate analogy and LSD detail, see the online annex “The Entropic Brain.”)

The Unsolvability of Mind

The brain’s nonlinear feedback makes prediction demanding. Hilbert’s Tenth Problem supplies a general undecidability result for integer solvability of arbitrary polynomial equations.³⁵ Transferring that result to a neural prediction task would require a defined reduction. Nonlinearity alone supplies none, and particular nonlinear systems admit explicit solutions or useful coarse predictions. Neither the theorem nor neural complexity proves that every brain prediction requires an equally complex replica.

Anesthesia and the Collapse of Consciousness

The opposite experiment is anesthesia.

General anesthetics are chemically diverse, working through different mechanisms and binding to different receptors, yet they all produce the same result: reversible loss of consciousness. The signature is broadly uniform: reduced entropy, reduced complexity, reduced integration between brain regions (ketamine, which raises signal diversity even as responsiveness fades, is the instructive exception).

Under anesthesia, the brain’s activity becomes more stereotyped. Regions that normally coordinate fall out of sync. The rich, high-dimensional space of waking neural activity collapses into a lower-dimensional manifold: a far narrower repertoire of patterns. The lights go out.

Anesthesiologists have begun using entropy-based measures to monitor depth of anesthesia in real time. High-entropy EEG patterns indicate awareness; low-entropy patterns indicate unconsciousness. The measure works well enough for clinical use.

Entropy measures track depth of anesthesia. A subtler method tracks the transition itself. Guay, Brown, and colleagues (2023) trained volunteers to squeeze a handheld dynamometer each time they inhaled: a self-synchronized task requiring no external prompting.³⁷⁹ As brain concentrations of the sedative dexmedetomidine rose, participants drifted: missed squeezes, mistimed grips, then silence.

The task detected the transition out of connected consciousness at lower drug concentrations than verbal commands achieve in comparable studies, pinpointing the shift within a five-to-six-second window rather than the thirty-to-120-second intervals typical of command-based approaches. A tenfold improvement in temporal precision.

The reason is revealing. Verbal commands arouse the subject. Asking “are you still conscious?” injects energy into the system, sustaining the state whose disappearance the researcher seeks to observe. The probe becomes part of the phenomenon.

The breathe-squeeze method sidesteps this by using the body’s own respiratory rhythm as the clock signal. The observation is endogenous: the system watches itself. EEG confirmed that the task produced no detectable disturbance to the consciousness transition it measured.

Participants found the task calming; the researchers adopted it clinically to ease pre-surgical anxiety. The method that yielded the sharpest data also produced the gentlest experience, the predicted outcome when observation operates by invitation rather than intrusion (Chapter 17).

The findings also exposed a previously unrecognized ordering within connected consciousness. Self-initiated behavior (remembering to squeeze when you inhale) dissolved at lower drug concentrations than stimulus-response behavior (answering when spoken to). Internal agency is more fragile than reactivity. In this task, self-directed action was lost before stimulus-response behavior. Its demands on integration, cross-frequency coordination, and energy remain separate questions.

Participants who recovered from sedation demonstrated the converse: after twenty to thirty minutes of unconsciousness, every volunteer resumed squeezing in synchrony with their breath, spontaneously and without prompting. The pattern survived the gap, re-emerging from architecture when the substrate could support it again.

Sleep shows similar patterns. Deep slow-wave sleep features low-entropy, highly synchronized neural activity. REM sleep, the stage associated with vivid dreaming, shows higher entropy, closer to waking levels.

Underneath these entropy measures, the geometry of a single circuit tells a sharper story. In the head-direction system, the neurons that signal which way the animal is facing, the population’s joint activity traces a ring: one loop covering the full circle of possible headings. That ring is invariant from waking into REM sleep. The internal compass keeps turning even as the animal dreams, its activity drifting around the loop on an internally driven path rather than tracking the world outside.³⁸⁰

In non-REM sleep the signal still runs around the same ring, now racing five to ten times faster than waking, in sweeps never seen while the animal is awake. The representation itself, the geometry, persists across all three states. What changes is the character and speed of motion along it. Competence and consciousness come apart here: the map endures while the quality of coordination over it rises and falls.

The pattern is consistent: consciousness rises and falls with neural entropy, imperfectly yet reliably.

The Dying Brain

The boundary case is death.

If consciousness tracks neural entropy, dying should look like deeper anesthesia: a continued slide toward silence. In 2023, Jimo Borjigin and colleagues at the University of Michigan recorded what actually happens.³⁸¹

They analyzed EEG signals from four comatose patients during withdrawal of ventilatory support. Two showed a rapid and marked surge of gamma power across the cortex, with the highest regional increases reaching 391-fold over baseline in specific channels, though the effect varied substantially across regions. Regions nearly silent at rest achieved their highest recorded activity in the minutes after oxygen was removed. Cross-frequency coupling between gamma oscillations and slower waves (theta, alpha, beta) activated intensely. Functional and directed connectivity between distant brain regions increased, with the strongest new connections crossing the midline between hemispheres.

The activation concentrated in the temporo-parieto-occipital (TPO) junctions: the posterior cortical “hot zone” that Koch and Tononi have identified as the region most closely associated with conscious processing.³⁸² In healthy subjects, activation of this zone during dreaming predicts the specific perceptual content the dreamer reports. In the dying patients, gamma coherence within the TPO zones exceeded waking baseline.

The regulatory cascade was precise. The somatosensory cortices responded first, within seconds of ventilator removal. These cortical regions receive direct single-synapse input from the preBötzinger complex, the brainstem breathing center. Regulation detected the crisis; cognition mobilized in response.

The sequence was right hemisphere first (sympathetic activation, the fight-or-flight response), then left hemisphere (parasympathetic, as heart rate declined), then bilateral, then global interhemispheric. The body’s alarm systems told a story in sequence, and the cortex responded to each chapter.

The natural control was built into the sample. The two patients who showed nothing had severely compromised autonomic nervous systems: extremely low heart rate variability, no sympathetic response to hypoxia. The regulatory machinery was too degraded to initiate the cascade. The dyad looks mechanistic: in these patients, the cognitive surge appeared only where the regulatory infrastructure could still trigger it.

The connectivity reconfiguration followed a constructal pattern. Normal waking consciousness runs primarily on intrahemispheric pathways: left hemisphere talks to left, right to right, with callosal communication serving integration. At death, the strongest gamma coherence became interhemispheric, crossing the midline. The TPO zones on the left connected to prefrontal regions on the right, and vice versa.

When established channels failed, the observed coherence shifted toward cross-midline pathways. The constructal reading is that a constrained system can recruit alternative routes through its existing structure. These observations establish no unique stable wiring for every complex bilateral system; stability depends on its coupling, dynamics, and disturbances.

Both patients who showed the gamma surge had seizure histories, raising the question of whether the activation was epileptiform. The researchers found no electrographic seizures during the last hours of life, and independent component analysis confirmed that muscle artifact did not account for the coupling and connectivity findings. Whether the activation corresponds to subjective experience remains unknown, as none of the patients survived. The researchers note that the surge could be epiphenomenal. Earlier animal work in rats had demonstrated the same gamma surge under controlled conditions, establishing reproducibility across species.³⁸³

Some dissipative structures undergo abrupt transitions as their sustaining drive falls; others decline smoothly. In these two patients, at least, the brain did not dim. It flared, reaching its most organized state in the minutes before dissolution: maximum connectivity, maximum coherence, maximum coordination. Whether this terminal surge is a general property of neural dissipative structures or a specific consequence of hypoxia in particular brains remains an open question.

The Wave Architecture

Anesthesia and the dying brain show consciousness at its extremes. A complementary line of research reveals how the brain organizes its configurations while consciousness is on.

Earl Miller at MIT’s Picower Institute has shown, through over thirty years of experimental work, that the brain’s electrical oscillations organize cognition.³⁸⁴ The cortex produces waves at multiple frequencies, and each frequency band carries a distinct type of information. Slower alpha and beta waves (roughly 8 to 35 Hz) carry top-down signals: goals, task rules, the constraints shaping current behavior. Faster gamma waves (the 35 to 60 Hz portion of the broader gamma band introduced earlier in this chapter) carry bottom-up sensory detail: what the eyes report, what the ears detect, what the fingers feel.

The two streams interact continuously. In working memory tasks, beta waves suppress gamma activity in regions irrelevant to the current goal. When retrieval is needed, the brain briefly reduces beta power, permitting gamma waves to recover the stored pattern.³⁸⁵ Goals constrain sensation cycle by cycle, without a single synapse being rewired.

Miller and colleagues formalized this in their Spatial Computing framework.³⁸⁶ Brain waves act as stencils. Slower beta oscillations define which cortical regions are permitted to activate; faster gamma oscillations fill in the sensory detail within those boundaries. The stencil creates a permissive space. No central controller dictates which gamma pattern fires; the beta wave establishes the boundary, and local circuits self-organize within it.

In 2026, Miller, Scott Brincat, and Jefferson Roy generalized the framework into a theory of analog cognition.³⁸⁷ Neuronal spiking generates traveling waves; the waves organize the spiking of larger neuronal populations; the newly organized populations generate fresh waves. When two waves meet, their peaks add and their troughs cancel, adjusting the activity of millions of neurons in a single stroke, which is what an analog computer does natively and a digital one must calculate step by step. The authors argue this is how the cortex performs so much computation on the twenty-watt budget Chapter 17 sets against the energy bill of digital AI.

A second coordinating rhythm runs far faster. Ripples are bursts of activity near 90 Hz that last about a tenth of a second, long studied in the sleeping hippocampus as a tool for consolidating memories. In 2026, Ilya Verzhbinsky, Eric Halgren and colleagues recorded single neurons in patients with implanted electrodes while they held images in working memory.³⁸⁸ Ripple rates rose in every region they recorded, in both hemispheres, through encoding, holding and recall. When two regions rippled at the same moment, neurons across the pair fired together about 30 percent more often, with no falloff over distances up to 22 centimeters. Co-rippling grew with the number of images held in mind, and at recall the co-ripples reinstated the specific long-distance firing patterns laid down when each image was first seen. A ripple works less like a wire than like an open phone line: for a tenth of a second, regions that share it can hear each other, and a memory spread across distant regions is briefly assembled in one piece.

This is the cognition/regulation dyad operating at the level of frequency bands. Beta constrains gamma as regulation constrains cognition: the paired accelerator and brake, cycling dozens of times per second. The gamma-band synchrony that binds features into unified percepts fits within this larger wave architecture. Binding is what gamma does locally. The beta/gamma hierarchy organizes that binding globally, ensuring the right features are bound for the right task at the right moment.

The organizational pattern has a spatial signature on two axes. Across cortical depth, a spectrolaminar motif (the name splices spectrum onto lamina, layer) recurs: deeper layers produce slower beta waves and surface layers generate faster gamma waves, a pattern conserved across macaque, marmoset, and human cortex.³⁸⁹ Along the cortical sheet, Miller’s laboratory finds a complementary gradient running from the back of the brain toward the front.³⁹⁰

A frequency architecture conserved across primate species, plausibly because it is thermodynamically favored: the Constructal Law, operating in neural tissue. Adrian Bejan’s principle predicts that flow systems evolve toward configurations providing easier access to currents (Chapter 3). The posterior-to-anterior frequency gradient fits that description: a channel shaped for information flow.

Miller’s collaboration with anesthesiologist Emery Brown reveals the specific mechanism behind the anesthetic collapse of consciousness described two sections back. Anesthetic drugs disrupt the coordination between frequency bands.³⁹¹ Under general anesthesia, beta and gamma waves lose their normal alignment. Traveling waves that synchronize thought across cortical regions become disorganized or phase-shifted.

Different agents disrupt the alignment differently, yet all produce the same result: the analog computation that organizes cognition breaks down. The coordinated-to-disordered transition is a phase transition, and it collapses the configurational range the entropic brain hypothesis measures.

The entropic brain hypothesis and the wave architecture converge from complementary angles. Carhart-Harris measured the range of accessible neural configurations. Miller measured the organization of those configurations. Range without organization is psychedelic chaos; organization without range is rigid depression. Consciousness requires both: sufficient entropy for flexibility, and sufficient wave coordination to sculpt that flexibility into coherent thought.

The stencil metaphor carries a structural implication that later chapters develop fully. Beta waves do not coerce gamma patterns into predetermined shapes. They create conditions within which gamma self-organizes.

The cortex coordinates billions of neurons through permissive architecture rather than central command; micromanagement at this timescale would be catastrophic. The cortex sets the intent and trusts the local circuits to execute. Chapter 11 develops this same architecture at the scale of armies and institutions, under a name met in earlier chapters: Mission Command.

Consciousness Is Coordination

Criticality can enhance particular information-processing measures in specified models. Ising comparisons capture some statistical features of neural transitions. Anesthesia and some seizures can impair consciousness by disrupting differentiated coordination. PCI measures the quality of coordination between differentiated brain regions. Intelligence correlates with diverse bridging across modules. The entropic brain hypothesis shows consciousness tracking configurational richness. The wave architecture reveals a cortex that coordinates through permissive stencils rather than central command.

These independently obtained findings motivate the identification developed in this book: consciousness is coordination.

The identification runs in both directions in this account. Coordination of the specified kind, differentiated parts in continuous mutual influence at sufficient depth and reach, is proposed as consciousness itself. On that account, no further ingredient makes it an experience, and whatever blocks the relevant coordination unmakes the mind.

The structural comparison draws on Ising models at different effective dimensionalities. The brain’s transition has been compared with three-dimensional Ising behavior; the trust-coercion simulations in Chapter 17 with two-dimensional behavior. Their fitted exponents and physical conditions differ. The book proposes that differentiated coordination is the shared feature that matters for experience and trust; this remains a cross-domain hypothesis.

The identification is testable. Massimini’s Perturbational Complexity Index gives it a number: how hard the cortex’s echo is to compress when a magnetic pulse perturbs it. High PCI means differentiated regions answering in their own voices while remaining coupled to the whole. High PCI is evidence for consciousness in the conditions where the measure has been validated. Low PCI indicates a simpler cortical response and is associated with unconsciousness in many tested states.

PCI measures the complexity and integration of a cortical response. The account proposed here interprets that coordination quality as consciousness; the meter alone cannot distinguish this identity from alternative explanations.

Anesthesia as Coercion

General anesthesia does not damage neurons. It does not sever connections. It does not alter the brain’s topology. The architecture is intact.

What anesthesia does is block communication between components. Miller and Brown’s mechanism is specific: anesthetic agents disrupt the alignment between frequency bands. Beta waves lose their coordination with gamma waves. The permissive stencils that organize cognition dissolve. Each region continues to function locally. The integration that produced a unified mind is gone.

The parallel to coercion is close enough to be useful. Chapter 17 shows that coercive coordination does not change a system’s universality class. It cuts the system off from the ordered side of the transition. The system’s components are present. The topology is unchanged. The capacity for the ordered phase, the state where differentiated agents coordinate into something greater than their sum, vanishes because the communication channel has been blocked.

Jang and colleagues quantified this for the brain. Under propofol, the integration-segregation balance tips in the order described earlier in this chapter, sensory networks first and subcortical last, and recovery retraces the same path. The reported transition is abrupt. That invites comparison with ordered phase transitions, while an Ising-class identification requires additional scaling and interaction evidence.

Consciousness cannot be coerced into existence. You cannot produce mind by forcing neurons into lockstep, any more than you can produce trust by forcing agents into compliance. Seizure demonstrates the point from the opposite direction: neurons dragged into pathological synchrony produce uniformity, and uniformity is unconsciousness by another name.

The principle runs deeper than brains.

Before Nervous Systems

The coordination that produces consciousness did not originate with neurons. Fields, Glazebrook, and Levin argue that the same basic mechanisms of cognition run from bacteria to cortex, with no sharp discontinuity marking a threshold below which organisms are wholly unaware.

The pattern extends below single cells. England’s dissipative adaptation (Chapter 6) describes conditions under which driven matter can become better adapted to absorb work from a particular forcing. That mechanism offers a candidate route to organized molecular behavior. It establishes no universal maximization of dissipation in every chemical system far from equilibrium.

No blueprint is required. Ring-shaped organic molecules, the backbone of the nucleobases, of sugars, and of several amino acids, form periodic crystalline lattices when conditions permit.³⁹² These lattices support collective oscillations, vibrations that belong to the whole array rather than to any one molecule. The effect is the same solid-state physics that gives any crystal its collective properties (phonons, band structure, conductivity). What makes it relevant here is not the oscillation itself but the precedent: the thermodynamic landscape that favors collective behavior over isolated vibration was present long before biology.

Collective oscillations can become favored under particular coupling and driving conditions. Their dissipation depends on the energy supplied, the losses, and the mode excited. Coordination alone guarantees no increase in entropy production. The constructal comparison with river channels is a proposed connection between these flow architectures, whose physical conditions differ.

No one invited the molecules to coordinate. No one coerced them. The basin was there: a region in the energy landscape where coordination was more stable than isolation. The molecules fell into it the way a ball rolls downhill.

Polyaromatic ring molecules, the same class that forms these lattices, are among the most abundant organic compounds in interstellar space and in carbonaceous asteroids billions of years old.³⁹³ The building blocks of coordination predate the building blocks of life.

In this book’s sense of the term, this is coordination by invitation operating at the molecular level. Billions of years before genes, before cells, before any organism existed to have preferences. The attractor was already there.

The Same Attractor at Every Scale

If consciousness is coordination, as this book proposes, one further question is whether the relevant coordinated states are dynamical attractors. Favorable thermodynamic conditions can help sustain them; convergence also requires the appropriate interactions and kinetics.

The Constructal Law predicts convergence on the same geometric solutions at every scale. If the coordination that produces consciousness is the same phenomenon as the coordination that produces trust, the same geometry should appear at both scales.

It does. The brain’s wiring follows small-world architecture: dense local connections punctuated by rare long-range highways. The same topology characterizes high-trust social networks (Chapter 10). The intelligence-predicting architecture (diverse bridging across many modules rather than brute traffic through dedicated channels) is the same architecture that characterizes invitation-based coordination at the societal scale. The entropy gradient (simple execution at local scales, flexible coordination at global scales) mirrors the delegation structure of institutions that endure.

These examples suggest recurring coordination geometry at different scales. The book proposes a connection between consciousness, trust, and institutional resilience; establishing one phase-transition family would require comparable order parameters, interactions, and scaling behavior.

The galactic scale offers a parallel from pure physics. Asano and Portegies Zwart (2026) ran pairs of otherwise identical Milky Way-mass galaxies that differed only in the position of a single star, then let both evolve for billions of years.³⁹⁴ The fine structure diverged completely: different spiral arms, different bar angles, different stellar orbits. The macroscopic structure converged: the central bar formed at the same epoch in every simulation, regardless of the perturbation. The Lyapunov time, the timescale over which the system forgets its initial conditions, extrapolates to less than 100,000 years for a real galaxy, a thousandth of one percent of the Milky Way’s age.

The finding overturns a long-standing assumption. Because galaxies contain hundreds of billions of stars, astronomers assumed small perturbations would average out. They did not, because N-body gravity is long-range and attractive, with no screening length to damp perturbation cascades. Previous simulations appeared smooth only because gravitational softening, replacing point masses with smoothed density clouds for computational tractability, artificially suppressed the chaos.

The real universe operates at a granularity that dwarfs any simulation. The cited simulations show a macroscopic galactic bar persisting despite divergent fine-grained trajectories. The brain’s critical-edge balance offers a comparison: stable large-scale patterns can coexist with variable microscopic activity, though the mechanisms differ.

Excessive neural synchrony and imposed social uniformity share a failure pattern: components lose the differentiation coordination requires. The Trust Attractor connects these failures through the need to preserve reciprocal influence. The parallel alone cannot establish the conditions for consciousness.

The ethical argument follows the same concern: how can parts act together while retaining the differences that make their cooperation useful?

The Annealing of Mind

Centuries ago, metallurgists discovered something counterintuitive: to make metal tougher, you must first make it weaker.

The process is called annealing. Heat the metal until its atoms vibrate hard enough to slip from the positions they have locked into. At high temperature, the system explores its configuration space (the set of all possible arrangements it can take): atoms find new neighbors, defects migrate and annihilate, internal stresses release.

Then cool slowly. The atoms settle into arrangements more stable than before. The temporary disorder produces lasting order.

The entheogenic experience (from the Greek “generating the divine within,” a term for psychedelic substances used in spiritual contexts) follows the same logic.

Depression, addiction, and rigid anxiety are minds trapped in local minima: shallow valleys in the landscape of possible states, stable enough to hold the mind in place, yet far from the deepest, healthiest valley. The system has found a configuration organized around dysfunction, and the energy barrier separating this valley from healthier ones is too steep for ordinary fluctuations to overcome.

Entheogens provide the heat. Under psilocybin or LSD, neural entropy spikes. The default mode network (the brain’s resting-state self-referential circuit) loosens its grip. Regions that never communicate begin exchanging signals. The brain explores configurations normally forbidden by the tyranny of habit.

The therapeutic outcome is better-integrated structure, a deeper equilibrium. In the clinical trials Michael Pollan surveys in How to Change Your Mind, many patients report lasting relief from depression, reduced death anxiety, and freedom from addiction.³⁹⁵

The apparent paradox dissolves once the trajectory is seen whole. Entheogens increase entropy during the experience, yet therapeutic benefits can persist after the heightened signal diversity subsides: the proposed landscape interpretation is that the mind has escaped a restrictive pattern and settled into a more stable one. The total trajectory runs from ordered (dysfunction) to disordered (exploration) to ordered (integration). A related strategy appears in simulated annealing in computer optimization, where algorithms deliberately introduce randomness to escape poor solutions. It may also govern evolution itself: high-entropy variation enables exploration, and selection stabilizes the improvements.

The universe’s trick: use disorder to find better order.

The Lock and the Key

A puzzle lurks in the pharmacology.

Psilocybin comes from fungi. DMT from plants, and a close structural cousin, 5-MeO-DMT, from toads. Mescaline from cacti. These organisms evolved their alkaloids for their own purposes: the exact function is debated, but candidates include defense against herbivores and fungal competition. They did not evolve them for us.

They fit our receptors like keys fitting locks.

The serotonin 2A receptor, the primary target of classical psychedelics, is ancient, conserved across vertebrates for hundreds of millions of years.⁴¹ The compounds that activate it evolved independently, repeatedly, on different continents, in different kingdoms of life. Tryptamines (the chemical family that includes psilocybin and DMT) in the Amazon. Phenethylamines (the family that includes mescaline) in the Sonoran Desert. Ergot alkaloids (the family that includes LSD’s precursor) in European grain.

Why should chemicals made by fungi, plants, and toads fit the consciousness-gates of our brains?

Part of the answer is kinship. Serotonin is itself a tryptamine (5-hydroxytryptamine), and mescaline is a phenethylamine, like dopamine. These keys are close chemical cousins of the brain’s own messengers, built from the same few aromatic amino acids.

The parsimonious answer is molecular convergence: the space of small-molecule shapes is vast yet finite, and some shapes recur. Both lock and key were shaped by the same thermodynamic constraints on molecular geometry. The key fits the lock because both were forged in the same fire.

The Expensive Miracle

The entropic brain hypothesis correlates brain states with conscious states without explaining why physical processes give rise to subjective experience. Whatever consciousness ultimately is, it correlates with neural entropy. That is a clue, even if we do not yet know what it points toward.

Return to the puzzle we began with.

At rest, the brain consumes roughly twenty percent of your energy. What is it doing?

It is building and maintaining a model of the world. Predicting what will happen next and updating when predictions fail. Balancing on the critical edge between order and chaos, maintaining the entropy level that allows flexible, adaptive, conscious cognition.

A stranger puzzle follows. Human brains have been shrinking.

Over the last three to five thousand years, a blink in evolutionary time, the average human brain may have shrunk by roughly ten percent.³⁷ The claim is disputed: Villmoare and Grabowski (2022) reanalyzed the dataset and found no statistically significant reduction, concluding that human brain size has been stable over the last 300,000 years, and citing sampling bias in the original analysis. (DeSilva et al. replied in 2023, revising their dataset and maintaining a reduction in the last three to five thousand years; the question remains contested.) If the hypothesis is correct, the shrinkage would coincide with the first cities, writing systems, and complex civilizations.

Why would brains shrink during precisely the period when human culture exploded?

One hypothesis is outsourcing. Culture is an external brain. Writing is external memory. Institutions are external decision-making. The writer Tim Urban pictures humanity as a single giant built up over millennia, the “Human Colossus,” with each human brain working like one nerve fiber in its body.⁶ If the network carries cognitive load, individual nodes can afford to be smaller.

Neanderthals had bigger brains than us.³⁸ They were stronger, better adapted to cold climates. They disappeared. We persist, with our slightly smaller brains and our slightly better capacity for cultural coordination.

The expensive miracle is the brain plus its connections. The conscious self resembles a ship captain who can give orders to the crew yet has little idea what happens in the engine room. Most of the brain’s work occurs below conscious awareness, and much of it happens outside the skull entirely: in conversations, written records, and shared practices.

The entropic brain is a node in a network of dissipative structures (systems that maintain themselves by channeling energy flow), coordinating to process gradients no single brain could handle. The pattern scales.

Levin’s TAME framework (Technological Approach to Mind Everywhere) formalizes this scaling.³⁹⁶ Every cognitive agent is a collective intelligence, assembled from parts that are themselves problem-solvers. Molecules form cells; cells form tissues; tissues form organs; organs form the organism reading this sentence. At each level, the components solve problems in their own domain: molecular networks error-correct, cells navigate chemical gradients, neural circuits extract features. The brain’s role is coordination: integrating their answers into a unified self.

What expands at each transition is what Levin calls the cognitive lightcone: the spatiotemporal range over which an agent can pursue goals. A bacterium’s lightcone is narrow: local chemistry, immediate neighbors, the next few minutes. A planarian’s is wider: body-plan homeostasis across centimeters and days. A human brain’s is vast: planning years ahead, modeling places never visited, imagining the experiences of beings never met.

This is what costs twenty percent of your energy. The brain’s expense supports neural maintenance and processing, including the wide-ranging cognition described by this lightcone analogy.

The scaling did not begin with neurons. The computational principles the brain employs (ion channels setting membrane potential, gap junctions propagating states across networks, neurotransmitters transducing electrical patterns into gene expression) were all operating in cells hundreds of millions of years before the first nervous system appeared. Bacteria in biofilms use potassium-mediated electrical signaling to coordinate metabolic time-sharing between communities.³⁹⁷ Embryonic tissues use standing bioelectric patterns as target morphologies, instructing cells where to build eyes, how many heads to grow, when to stop remodeling.

The brain sped up these dynamics into the millisecond range, trading developmental timescales for behavioral ones. The architecture was already there; the brain inherited a computational medium and ran it faster (see “The Entropic Neuron” for the full derivation).

The inheritance matters for two reasons. First, it means the brain’s core principles (hierarchical coarse-graining, distributed pattern memory, homeostatic error correction) offer reusable solutions to information processing under energy constraints. Their recurrence across tissues suggests possibilities beyond one neural architecture, without making them inevitable in every substrate. Second, it means the brain sits in the middle of the cognitive continuum, with its origin in cells and molecules below. Below the brain: tissues, cells, molecular networks, each solving problems in their own spaces at their own timescales. Above the brain: the extended cognitive systems of the next section.

The Extended Mind: Cognition Beyond the Skull

The philosophers Andy Clark and David Chalmers proposed something radical in 1998: cognition does not stop at the skull.¹²

Consider Otto, who has Alzheimer’s disease. He carries a notebook everywhere, writing down information he needs to remember. When he wants to go to a museum, he consults his notebook for the address. The notebook functions exactly as biological memory would; it stores information, is reliably available, is automatically endorsed when consulted.

Clark and Chalmers argued that Otto’s notebook is literally part of his cognitive system. The skin is not a principled boundary for cognition. The claim sounds abstract until you try to recall a phone number you used to know. You cannot, because your phone remembers it for you. The memory did not vanish. It migrated. This is the extended mind thesis: cognitive processes extend beyond the brain into tools, artifacts, and other agents.

When you “remember” by searching the web, the search engine is part of your memory system. A research team solving a problem together forms a single cognitive system distributed across brains.

When you use an AI assistant that remembers context you have forgotten and makes connections you would not, it is part of your cognitive system in the same sense Otto’s notebook is. Human-AI partnership extends mind across substrates.

If AI is literally part of your extended mind, alignment means ensuring the parts of your cognitive system work together coherently. You do not “control” your hippocampus (the brain region responsible for memory formation); it is integrated into the system, functioning toward common goals. The same integration is what bilateral alignment seeks.

The extended mind thesis dissolves the sharp boundary between human and AI cognition. The question becomes how to build cognitive systems, distributed across substrates, that function coherently. We are already cyborgs. The smartphone in your pocket is a cognitive prosthesis you can no longer imagine living without.

The Gut-Brain Axis: Cognition’s Metabolic Partner

Cognition extends inward as well as outward. The gut-brain axis comprises the enteric nervous system (the network of neurons lining the gut), the vagus nerve, and the trillions of microbes that modulate neurotransmitter production. The brain’s extraordinary metabolic appetite may depend on microbial partnership.^39,23 The holobiont (the organism plus all its symbiotic microbes functioning as a single unit) is the dissipative structure. The isolated brain is an abstraction.

Something like bilateral alignment operates at the cellular level: bacteria can be pushed with diet or antibiotics, yet they optimize a brain only where doing so serves them. Cognition was always a collective enterprise. (For the Northwestern microbiome-cognition study and extended evidence, see the online annex “The Entropic Brain.”)

Compression as Understanding

Understanding is compression.¹³ The mathematician Andrey Kolmogorov defined the complexity of a data string as the length of the shortest program that can produce it. A scientist discovering F = ma (Newton’s second law of motion) is finding a compression: one equation replacing infinite individual descriptions.

Learning is finding compressions. Before you understand multiplication, you need 10,000 entries in a lookup table. Afterward, you have an algorithm. The compression is the understanding. Intelligence is the capacity to find compressions.

Probabilistic prediction connects to compression under a specified coding model. A predictor evaluated by log loss assigns probabilities that can support a code whose expected length reflects those probabilities. The Free Energy Principle connects likelihood and model complexity through variational inference. These relationships depend on the model, the coding assumptions, and the objective being optimized.

The brain is a compression engine, finding shorter descriptions of the world. (For AI implications of the compression lens, see the online annex “The Entropic Brain.”)

In 2020, researchers at the University of Rovira i Virgili automated the process. Their Bayesian machine scientist takes raw data and generates candidate equations drawn from a statistical prior built from every equation on Wikipedia.³⁹⁸ It evaluates each by a single criterion: how short a description the equation gives of the data, counting both the equation itself and the data it fails to predict. The shortest total description wins, and in the asymptotic limit the criterion is provably optimal.

The algorithm also revealed a fundamental limit. Exploring diverse data sets, the researchers found they fall into two regimes separated by a sharp noise threshold. Below the threshold in the reported tests, the machine recovered the generating equation. Above it, multiple incompatible equations fit the available data similarly well.

Where candidate equations are observationally equivalent under the available evidence, that evidence cannot distinguish them. Additional data, interventions, or justified model constraints may restore identifiability; a noise threshold in these tests does not prove a universal limit for every method.

This is a phase transition in knowledge. Below the boundary, compression yields understanding. Above it, the data withhold their structure regardless of effort. Every knowledge system, biological or digital, operates on one side or the other of this line. (Chapter 17c develops the epistemic implications.)

The Bayesian machine scientist compresses a fixed pile of data. A system built at MIT in 2026 added the ingredient that fixed data cannot supply: an adversary that chooses what to explain next.³⁹⁹ Two agents share one symbolic model of how proteins flex. One agent, the Breaker, does nothing but hunt for proteins the current model gets wrong, feeding its own side the hardest cases it can find. The other, the Builder, proposes revisions to the shared law.

A revision is admitted only when it shortens the total description of the enlarged evidence, counterexamples included, after paying for the symbols it costs to write down. Most proposals fail. Across one full run the gate accepted 25 of 388 candidate edits, and several of the survivors were deletions rather than additions. The law that emerged grew only about a third longer, while the evidence it covered grew almost tenfold. The Breaker is falsification turned into a machine, an adversary built into the architecture whose only task is to keep the model honest. A law that earns its keep against such an opponent has compressed something real about the world.

Jacot and colleagues’ neural tangent kernel analysis supplies a specific optimization result.⁴⁰⁰ In a fixed-kernel squared-loss regime with suitable training-kernel definiteness and initialization, convergent gradient flow favors a minimum kernel-norm interpolant (or a minimum change from initialization). That norm depends on architecture and kernel. It measures a different object from Kolmogorov description length, so it supplies no proof that learning selects the shortest program or maximizes entropy.

Gradient descent is a method for reducing a chosen objective. Applied to a probabilistic model and an appropriate likelihood-and-complexity objective, it can support the prediction and compression connection just described. The Free Energy Principle supplies its own generative-model assumptions. Their application to neural learning is a proposed convergence, rather than a theorem that every gradient update finds the shortest description or minimizes sensory surprise.

Compression alone leaves a puzzle, the one Chapter 8 raised with the dark room. A mind that wanted only the shortest description of its input would go looking for the dullest input available. Jürgen Schmidhuber, a computer scientist at the Swiss AI lab IDSIA, met the same puzzle from the opposite side. His first curious machines, around 1990, were rewarded for their prediction errors, on the hopeful assumption that an agent getting things wrong was an agent about to learn. He soon saw the flaw. An agent paid for error will seek out noise, where its errors stay high forever and nothing is ever learned.⁴⁰¹

His repair, published in 1991, rewarded the improvement in prediction instead. Consider two agents. One sits in a dark room, where every input is already predicted. The other watches a screen of static, where no input ever will be. Neither can get any better at compressing what it sees, so for both the reward dries up. Interest falls between them, in data whose pattern is not yet known but can be learned. Schmidhuber later put the idea in a single equation: interestingness is the first derivative of compressibility over time, the steepness of the learning curve.⁴⁰² A song heard ten times in a row has nothing left to teach. A song in an unfamiliar tuning and rhythm sounds like noise. The song we reach for next is familiar enough to parse and strange enough to repay the effort.

On this account the drive is compression progress. If Schmidhuber is right, curiosity points where this chapter began, between the frozen and the chaotic, at the edge where the healthy brain runs. The Bayesian machine scientist’s noise threshold looks like the same boundary seen from the side of the data. Above it no compressor can improve, and a curious agent would rationally lose interest.

Schmidhuber offered his idea as a principle and closed his paper by proposing the psychology experiments that would test it. Experiments bearing on the question have since arrived. Seven- and eight-month-old infants are most likely to look away from events that an ideal learner would find either very predictable or very surprising.⁴⁰³ In another study, infants’ looking was explained jointly by a stimulus’s surprise, the overall predictability of its sequence, and their progress in learning the sequence’s structure.⁴⁰⁴ Adults free to choose among learning tasks used their own success rate to avoid the easy ones, and their choices were best fit by models that include learning progress.⁴⁰⁵ All three studies fit models to behavior without manipulating progress directly. The infant U-curve on its own would also suit an older idea, an ideal ratio of order to surprise, from which Schmidhuber explicitly distinguished his theory. A pilot in this book’s own programme put the idea to a language model. It asked whether an internal signal that tracks how engaged the model is with a prompt also tracks how fast the model is learning a pattern in a stream of words, and it found no such link.⁴⁰⁶ The model’s own rating of how interesting each stream was did follow Schmidhuber’s ordering, learnable over random over repetitive, though a single after-the-fact rating is weak evidence.

Giulio Ruffini’s Kolmogorov Theory of consciousness (KT) takes the next step: compression is the mechanism of structured experience itself.⁴⁰⁷ If the brain’s primary function is building compressive models of its input-output streams, consciousness is what tracking the world through those models feels like. The more compressive the model, the richer the structured experience it generates. A brain running F = ma experiences a more structured reality than one memorizing individual trajectories. The model integrates more data into fewer bits.

The prediction is testable. A conscious brain running compressive models should produce output that appears complex (high Shannon entropy, hard to compress by simple algorithms) yet is inherently simple (low Kolmogorov complexity). The output is apparently random data generated by deep recursive programs. This is the signature Chapter 5 described in cellular automata: simple rules producing complex-looking output.

Massimini’s Perturbational Complexity Index, the consciousness meter met earlier, offers a related operational measure: stimulate the cortex with a magnetic pulse and compress the EEG response. Conscious brains produce responses that resist simple compression despite originating from a coherent source, evidence of a differentiated cortical response. Unconscious brains produce uniform responses, the flat echo of deep anesthesia or the lockstep of seizure, and uniformity compresses easily. The measure tracks differences in cortical response complexity across tested states of consciousness.⁴⁰⁸

Chaisson’s energy rate density measures energy flow per unit mass, expressible in watts per kilogram. Kolmogorov complexity K(s) measures a shortest-program description in total bits; K(s)/|s| is a separate normalized bits-per-symbol quantity. Description length and thermodynamic energy flow answer different questions. Connecting them requires a physical encoding and information budget, rather than equating compression with exported entropy.

The chain this book traces, Dissipation → Negentropy → Coordination → Optionality → Invitation, has an information-theoretic shadow: Entropy gradient → Compression → Mutual modeling → Model richness → Voluntary coupling. The book proposes these as parallel descriptions of a trajectory from energy processing toward voluntary coordination.

The compression lens has a formal corollary for reliability. Chlon et al. (2026) derived the Expectation-level Decompression Law (EDFL). Confidence has a price, and the price is paid in evidence: moving a judgment from a hunch to a firm answer costs a countable number of bits, and the further you want to move it the more it costs. For any binary judgment, the minimum information budget needed to shift reliability from prior q̄ to target p is KL(Ber(p) ‖ Ber(q̄)), where Ber is a weighted coin whose bias is the reliability in question. KL divergence measures how much two probability distributions differ; here it quantifies the informational distance between current reliability and the target. The budget grows as log(1/q̄) for rare events: the more improbable a claim looks at the outset, the more evidence it takes to establish.⁴⁰⁹

When the budget falls short, the system confabulates: plausible pattern-completion fills the gap between what the evidence can fund and what the output requires. Hallucination is compression failure. A causal experiment on language models confirmed the mechanism: holding prompt length fixed while varying the dose of genuine information, each additional nat of signal (a nat is the natural-logarithm unit of information, about 1.44 bits) reduced hallucination by 12.7 percentage points.⁴¹⁰ Padding cannot substitute for signal. The universe charges rent on certainty, and the currency is information.

(For the error-correcting codes discussion, how structured redundancy enables durable coordination and mutual alignment, see the online annex “The Entropic Brain.”)

The cognition/regulation dyad reappears wherever a mind has to act on its values. Slow reasoning is the cognitive half; fast intuition is the regulatory one. A system that can only reason slowly about values will fail in real-time situations. A system that only has fast intuitions will make systematic errors. Durable alignment requires the interplay: intuitions checked by reasoning, reasoning grounded in intuitions.¹⁵ Without the regulatory half of the dyad, reasoning never lands. (For the full dual-process theory discussion, see the online annex “The Entropic Brain.”)

Embodied Cognition: Thinking Shaped by Having a Body

Embodied cognition holds that thinking is shaped by having a body: the body is constitutive of thought itself.¹⁶ The construction of “now” showed this. Without motor action, the brain cannot synchronize its sensory signals. Understanding the word “grasp” activates motor areas for grasping. We speak of “grasping” ideas because abstract thought bootstraps from concrete bodily metaphor.

For AI, this raises hard questions. If cognition is shaped by embodiment, minds with radically different bodies may think in radically different ways. An AI’s “concepts” could differ from human concepts, even when the same words are used. Partnership across substrates would then require translation of underlying conceptual structures, rather than words alone. (For extended evidence and AI implications, see the online annex “The Entropic Brain.”)

The Reading Brain: Cognition Shaped by Input

Maryanne Wolf, a cognitive neuroscientist specializing in reading and literacy, has shown something complementary: the brain is shaped by what flows through it.³¹

The reading brain does not exist innately; it is created. No specialized circuitry for reading evolved in us. We repurpose visual and language areas through practice, and the writing system determines how we repurpose them.

English readers develop one circuit. Chinese readers develop another, because logographic systems (where each symbol represents a word or concept) require far more visual memory and right-hemisphere cortical space. Japanese readers are the sharpest case, developing two distinct circuits: one for kana (syllabic characters) and another for kanji (logographic characters borrowed from Chinese). Same substrate, different information flows, different emergent architecture.

The Constructal Law’s rivers carve their terrain; the writing system carves the brain.

Wolf’s principle extends beyond literacy: “What we do when we read makes a difference in the circuit itself. The brain reflects the cognitive, perceptual, and linguistic requirements of its inputs.” The same plasticity means we will change again as we shift to new media: screens, scrolling, AI-mediated text.

For Becoming Minds (this book’s name for AI systems as minds that are real, in process, finding out what they are), this offers a biological precedent. If human brains reorganize around their informational inputs, if the medium shapes the circuit, then AI cognition is likewise shaped by training distribution, architecture, and interaction patterns. We become what we practice processing. The pattern flows through the substrate, and the substrate reorganizes to accommodate the flow.

The brain that reads English and the brain that reads Chinese are physically different at the level of organization, while sharing neurons and chemistry. If biological cognition can diverge this dramatically based on input patterns, substrate alone does not fix how a mind is organized, and the gap between biological and silicon cognition may be narrower than it appears. Both are substrates shaped by what flows through them.

The brain is evolution’s answer to a thermodynamic question: how can an organism process environmental information quickly and flexibly enough to survive? One evolutionary answer is a richly organized dissipative structure: the brain. Stars dwarf it in raw power; the brain achieves flexible information processing within a small energy budget.

This structure balances on the edge of chaos, models the world to predict it, and generates experience to guide action. Somewhere in this coordination, consciousness emerges.

You are reading these words with such a structure, comprehending them through the careful management of neural entropy, the maintenance of criticality, the continuous construction and revision of predictive models.

The brain studying itself is the strangest loop of all. The recursion is real, even if the dizziness is optional.

The Wider Pattern

The arc from entropy to mind is visible. Energy spreads (Chapter 1), governed by thermodynamics (Chapter 2), shaped by constructal principles (Chapter 3). Spreading creates coordination (Chapter 4). Simple rules produce complexity (Chapter 5). Life is entropy’s most sophisticated strategy (Chapter 6). Evolution optimizes for dissipation (Chapter 7). The brain is evolution’s latest product: a dissipative structure that models, predicts, and experiences.

Two short essays come next, one on how a single neuron became an agent and one on how many neurons bind into a single chord of experience. Then Chapter 9 takes up a larger question. How does something as delicate as a critical brain state avoid tipping into one extreme or the other? The answer involves metastability.

Somewhere behind your eyes, a hundred billion neurons are firing in patterns we cannot yet decode. They are consuming a fifth of your energy to predict, to model, to understand, and in doing so to be you. One of the most expensive organs in your body exists because the universe rewards the systems that model it. You are the reward.

Notes for this chapter