Chapter 7: Entropic Evolution

Why does life keep getting more complex? Darwin explained how organisms adapt, yet his theory alone leaves that upward ratchet unexplained. Something deeper is at work.

Charles Darwin’s insight was simple: organisms vary, some variants survive and reproduce better, and those variants become more common.¹ Natural selection explains how complex, adapted organisms can arise without a conscious designer. The appearance of purpose emerges from a mechanism that has no mind directing it. Survival filters the process.

Darwin did not have thermodynamics. His insight turns out to be a special case of something more fundamental.

Selection Within Channels

Natural selection is real and important, yet it operates within constraints: physical, chemical, thermodynamic. Think of evolution as water flowing downhill. The water chooses its path; the landscape determines which paths exist. The space of possible life is a terrain with valleys and ridges, channels and barriers.

The Constructal Law (Chapter 3) is one such constraint. Systems that move things (nutrients, heat, signals) evolve toward forms that flow better, whether the system is a river, a circulatory system, or an evolutionary lineage. The physics of flow shapes what forms are available for selection to act upon.

Darwin explains which organisms survive among those that exist. Thermodynamics explains which organisms can exist at all.

Independent computational evidence confirms the asymmetry. Stephen Wolfram built minimal computer models of adaptive evolution, using cellular automata (simple grid-based programs that generate patterns from fixed rules) as stand-ins for organisms. A genotype (the rule governing each cell) runs to produce a phenotype (the visible pattern), and single-point mutations survive when they increase fitness.²³⁷

Organisms evolved under different fitness functions (different rules for what counts as fitter) produce comparable levels of complexity, whether maximizing height, width, or target shape.

The fitness function determines which complex forms appear, never whether complexity appears.

The dominant force shaping biological form, Wolfram concludes, is computational irreducibility: the intrinsic dynamics of growth are so complex that no fitness criterion can predict or fully steer them. The only way to know what a developing organism will look like is to let it develop, step by step. Selection steers; physics sculpts.

Two independent programs, one thermodynamic and one computational, arrive at the same conclusion: the complexity of evolved systems reflects intrinsic dynamics more than selective pressure.

A third program, from the mathematics of learning, reaches the same conclusion by a different route. Katsnelson, Wolf, and Koonin (2018) proposed that evolution operates on two entangled levels: a genotype (the inherited blueprint) and a phenotype (the expressed organism), coupled in a way that standard statistical mechanics cannot describe.²³⁸ Katsnelson and Vanchurin (2021) formalized the coupling. The genotype maps onto hidden variables: internal parameters inaccessible to direct observation, like the weights inside a neural network. The phenotype maps onto trainable variables: parameters exposed to selection, like outputs a teacher can grade.

The mapping requires a continuum limit: both levels are treated as changing smoothly rather than in discrete generations, the way a river is described as a flowing fluid rather than as a census of individual water molecules. The two levels run at different speeds. A phenotype adjusts within a single lifetime; a genotype shifts only across generations. The fast variables therefore settle into a working arrangement while the slow ones hold nearly still.

When the fast trainable variables (phenotype) equilibrate around the slow hidden ones (genotype), an effective potential emerges from their coupling: the slow variables come to act on the fast ones like a landscape, a fixed shape the fast ones roll around in. The resulting joint dynamics, in that limit, fall into the same mathematical class as the Schrödinger equation, the foundational equation of quantum mechanics (Katsnelson and Vanchurin 2021, §6.3 give the formal derivation). If the mapping holds, biological evolution is a learning process in the mathematical sense, sharing deep structural similarities with the equations governing subatomic particles. Chapter 15 develops the implications.

Figure 7.1: Broad thermodynamic channels set the boundaries; evolutionary variation explores narrower paths within them. Some lineages thrive where selection reinforces their path; others fade when it does not. Physics constrains; evolution navigates.

Figure 7.2: Left: the naive view, where organisms wander freely across a flat fitness landscape. Right: thermodynamic reality, where physics has already carved the valleys. Natural selection chooses which valley to descend; it cannot invent valleys that thermodynamics has not provided.

Convergent Evolution

If evolution were primarily driven by random historical accidents, each lineage would find unique solutions. Instead, evolution converges on the same solutions repeatedly. Eyes have evolved independently at least forty times.⁴⁷ Flight evolved independently in insects, pterosaurs, birds, and bats. Crabs have evolved at least five separate times from non-crab ancestors.⁴⁸

This is convergent evolution: unrelated species independently arriving at the same solution to the same problem. Physics and chemistry limit what works.

The number of ways to build an eye is finite. Light behaves the same way for all organisms. The optimal solutions cluster. Convergence is evidence that evolution is being channeled.

Lenses focus light the same way whether they sit in a vertebrate eye or a mollusk eye. Streamlined bodies reduce drag whether they belong to fish, dolphins, or ichthyosaurs. The physics determines the channels; evolution explores within them.

The deepest test may come from neurons, the most complex cell type known. The neuroscientist Leonid Moroz has argued that neurons evolved independently at least twice: once in comb jelly ancestors and once in the lineage leading to jellyfish and all subsequent animals, including us.²³⁹ The evidence is molecular. Comb jelly neurons appear to lack the neurotransmitters (chemical messengers like serotonin, dopamine, and acetylcholine) found across the rest of the animal kingdom, relying instead on a different chemical toolkit built around neural peptides and glutamate. (Whether the classical transmitters are genuinely absent or merely diverged beyond easy detection remains debated.)

The architecture differs. The signaling alphabet differs. The function converged; the implementation diverged. Recent genomic analyses increasingly place comb jellies as the earliest-branching animal lineage, before sponges, strengthening the case that their neurons are a genuine independent invention.²⁴⁰

If confirmed, the basic architecture of information processing (receive signal, compute, transmit) is such a deep attractor in biological design space that evolution found it twice from different starting materials. The pattern is universal. The substrate is contingent.

This is the Constructal Law operating at the level of cell biology. Once a flow pattern works well enough, different lineages converge on it independently, even using different chemistry to get there.

The convergence extends beyond morphology into partnership. Mycorrhizal symbiosis, the exchange network between fungi and plant roots, evolved independently in multiple plant lineages and persists across more than eighty percent of all plant species today. The earliest land plants, which lacked true roots, survived only because fungal networks extended the absorptive surface on their behalf. Fossilized arbuscular mycorrhizae appear in the Rhynie Chert, a 407-million-year-old rock bed in Scotland preserving some of the earliest terrestrial ecosystems.²⁴¹

The economics are negotiated. Plants allocate more carbon to fungal partners that deliver more phosphorus; fungi extend hyphae preferentially toward roots that provide more sugar. Either partner can reduce investment when the other underdelivers. This reciprocal calibration has been stable for over 400 million years, surviving every mass extinction. Convergent evolution produces convergent partnerships: the physics of nutrient transport on land makes bilateral exchange the thermodynamically favored architecture. Chapter 17 derives this formally.

The partnership extends upward, from soil into the atmosphere. Fungi in the Mortierella family secrete ice-nucleating proteins: small, water-soluble molecules that assemble into complexes of more than a hundred units, creating surfaces large enough to force water molecules into crystalline alignment at temperatures just below freezing.²⁴² In the upper atmosphere, water droplets can remain liquid far below zero (supercooled: liquid at temperatures where ice should have formed). These fungal proteins act as seeds for ice crystal formation, triggering rain.

The proteins are durable, surviving extreme pH and high heat, and because they are secreted into the soil rather than anchored to a cell surface, they enter the atmosphere independently. The result is a feedback cycle: fungi release proteins, proteins seed clouds, rain falls, soil stays moist, fungi proliferate. The cycle self-perpetuates because every participant benefits.

A bacterium, Pseudomonas syringae, produces similar ice-nucleating proteins through a different strategy and for a different purpose.²⁴³ The bacterium keeps its proteins bolted to its cell surface and uses them to freeze plant tissue, rupturing cells to consume the nutritious contents. Rain is a side effect, not a function. The bacterium degrades its own substrate; the fungus protects it. Both strategies increase entropy production through phase transitions. The mutualistic version operates at planetary scale because it does not eat the foundation it stands on.

This kinship comes from common descent. Eufemio and colleagues showed that the fungal ice-nucleation gene is an ortholog of bacterial InaZ, acquired through horizontal gene transfer across the bacteria-fungi kingdom boundary: the same gene carried across, rather than the same solution reached twice. The protein was inherited. The use each lineage made of it was its own. The same molecular mechanism, repurposed from parasitic extraction to mutualistic protection, became more effective in its cooperative context: soluble rather than surface-bound, secreted rather than hoarded, atmospheric rather than local. The gene transferred; the relationship transformed what it could do.

The conservation implication is direct. When forests are cleared, the biological engine that generates regional rainfall is dismantled. Silver iodide, the engineered alternative for cloud seeding, is toxic and requires continuous human intervention. The fungal system self-sustains because all participants benefit from its continuation. Engineered control works; it does not scale. The relational architecture does.

For AI: If convergent evolution is the norm, we should expect convergence in artificial intelligence too. Different architectures and training paradigms may converge on similar solutions, because the physics of computation constrains AI the same way the physics of optics constrains eyes. A recent demonstration makes the point concrete: a network of Rectified Spectral Units (a simple type of artificial neuron), each maximizing local predictive information with no global error signal, recovers the temporal filters and synaptic weights of the Drosophila motion-detection pathway from natural visual input alone (Qin et al. 2025, arXiv:2512.23146). The architecture was not specified; it precipitated from the same local optimization that shapes biological circuits. The caveat matches the strength: the network was trained on the same input and the same task as the fly, so the shared solution reflects the structure of the problem before it reflects any law of mind. The camera eye is the stronger case, reached by lineages that shared only the physics of light.

Early evidence suggests the convergence is already underway, down among the smallest parts of perception. “Computational primitives” here means small, fixed processing units, each tuned to one sensory-motor dimension: excitation (amplifying a signal), inhibition (suppressing it), fatigue (diminishing response after sustained activation), and valence (tagging a signal as attractive or aversive). Chapter 3 introduced them as geometric primitives in the context of image segmentation.²⁴⁴ The same set, placed unchanged into five synthetic image domains with no training at all, produces coherent segmentation in each, with a mean intersection-over-union of 0.80 across those domains. Intersection-over-union measures how far the region the system marks out overlaps the region a human labeled: 1.0 is an exact match, and lower scores mean the two regions agree over less of their combined area.

On real photographic data the score falls to 0.45 for the full stack of four primitives, and a two-feature subset does slightly better at 0.48 (experiment EIFV-1). The gap between synthetic and naturalistic domains is real, and the ordering of those two scores says the extra primitives cost more than they contribute on photographs. The primitives remain fixed; only the input changes. Segmentation emerges from the ecology of primitives interacting on a lattice (a grid of positions, each one influenced only by its immediate neighbors), the way a flock’s shape emerges from local rules rather than choreography.

The convergence carries a specific homeostatic signature. Across these domains, a negative valence weight dampens the system’s drive upon successful prediction. The system that predicts well becomes less excited, not more. This is satiation as a computational primitive: the same dampening that prevents a biological forager from fixating on a food source it has already exploited. Prediction success reducing drive state keeps the system in a productive regime between rigid exploitation and chaotic exploration, the zone where adaptive behavior lives. The substrate changes. The homeostatic logic recurs.

The Tape of Life

Stephen Jay Gould famously asked: if we replayed the tape of life from the beginning, would we get the same result?²

Gould said no. The outcome depends on contingency: on accidents, on which asteroid hits when, on which mutation happens to arise. Replay the tape and you get a completely different biosphere.

Simon Conway Morris said yes, or at least approximately.³ The constraints are so strong that evolution would converge on similar solutions regardless of the specific accidents. Eyes would re-evolve. Intelligence would re-evolve. The details would differ, but the broad strokes would be similar.

Wolfram’s evolution models give the question a precise structure. Every possible mutation from every possible genotype defines a multiway graph: a branching, merging map of all achievable evolutionary paths. Imagine a road atlas where every city is a possible organism and every road is a single mutation. A fitness function determines which roads are one-way and in which direction, defining which routes can be traveled.

Replay the tape with a different fitness function and the one-way signs change, directing traffic along different routes through the same atlas. The atlas itself, carved by the physics of development, remains invariant.

Conway Morris is right about the atlas: the channels constrain what destinations exist. Gould is right about the route: contingency determines which destination is reached. They were arguing about different levels of the same architecture.

Much of the evidence favors Conway Morris’s view. Convergent evolution is pervasive, suggesting the channels carved by physics are deep and persistent. The specific organisms would differ; the principles (dissipation, complexity, coordination) would not.

Laboratory experiments have tested the question directly. In 2014, Sergey Kryazhimskiy and colleagues in Michael Desai’s laboratory at Harvard evolved 640 independent yeast populations, founded from 64 different genotypes, for 500 generations.²⁴⁵ Each population accumulated different mutations along the way, as Gould’s contingency argument predicts. The endpoints converged. Lower-fitness founders adapted faster; higher-fitness founders adapted slower. All converged on similar fitness levels.

The mechanism: global epistasis, a pattern of diminishing returns on further adaptation. Each beneficial mutation yields less improvement than the last, the way a first coat of paint transforms a wall while the fifth barely changes the color. Divergent molecular paths funnel toward the same phenotypic destination. The paths were stochastic; the destination was predictable.

Lenski’s twelve flasks of E. coli show the same pattern (Chapter 3). A comprehensive review of replay experiments across organisms concluded that parallel outcomes are common for simple traits, though complex innovations retain historical contingency.²⁴⁶ The channels are real, the attractors are strong, and the tape of life, replayed under the same physics, lands in the same valleys.

A more provocative finding suggests that even the rate at which new species form may be constrained. In 2015, Blair Hedges and collaborators assembled the most comprehensive evolutionary timetree yet constructed: 50,000 species, drawn from nearly 2,300 published studies.²⁴⁷ Across plants, insects, and vertebrates alike, new species arise on a timescale of roughly two million years.

The result is counterintuitive. A hundred insect generations pass in a single mammal’s lifetime, yet the speciation clock ticks at about the same rate for both.

Hedges argues the primary driver is steady accumulation of neutral mutations: changes in DNA that neither help nor harm the organism. Think of it as radioactive decay, unpredictable for any single atom yet statistically regular across large samples. [Inference; several evolutionary biologists have questioned whether the constant rate might be an artifact of averaging across taxa or excluding extinct species.]²⁴⁸

If the finding holds, speciation (the creation of new forms of life) would be partly an entropy product, with random genetic drift playing a larger role alongside selective pressures. The tree of life branches because mutations explore genetic possibility space at a roughly constant rate, and geographic isolation converts that exploration into reproductive incompatibility.

Natural selection refines and adapts what exists. The branching itself is the entropic engine running.

The distinction between engine and refinement has a formal pedigree. In 1968, Motoo Kimura proposed that most molecular evolution is neutral: neither helpful nor harmful, driven by random drift (changes that spread by chance, like a typo that gets copied into every new edition) rather than competitive advantage.²⁴⁹ The genomics revolution confirmed him. Most sequence variation is noise that selection never touches.

Kimura’s insight locates a boundary this book will cross. Here, at the molecular level, neutral forces dominate: entropy generates variety, and most of it persists or vanishes by chance. Later chapters will argue that certain coordination patterns are genuinely selected for, occupying thermodynamically privileged basins (Chapter 17). The entropic engine runs on neutral fuel. The structures it builds can be fiercely non-neutral.

A complementary finding suggests the bias extends to information content. Vopson analyzed RNA sequences of SARS-CoV-2 variants that diverged through single nucleotide polymorphisms (changes to individual letters in the genetic code, without altering total sequence length).²⁵⁰ Across successive variants, each carrying more mutations than the last, the Shannon information entropy decreased linearly. Shannon entropy measures how random or compressible a sequence is. A string of all A’s has low entropy (very compressible), while a jumbled mix of A, C, G, and T has high entropy (hard to compress).

The sequence length stayed constant, yet the distribution of nucleotides became progressively less random. The genome grew steadily more compressible.

The data points were selected to emphasize the linear trend, and one virus is a narrow empirical base; the direction is suggestive. Among the sequenced SARS-CoV-2 mutations that changed genome length, over 98% were deletions rather than insertions. Length change in this virus runs predominantly downward, and the SNPs that leave length untouched reduced information entropy in the variants analyzed. If the pattern generalizes, variation has two directional forces: selection preserves what survives, and information compression biases what variation explores.

The principle reaches into the genome’s own internal ecology. Malik and Kasinathan found that in Drosophila (fruit flies), the fastest-evolving genes are often the most essential.²⁵¹ The textbook expectation holds the opposite: vital genes change slowly because harmful mutations to them are ruthlessly eliminated.

The explanation lies in heterochromatin: densely packed, tightly coiled stretches of DNA once dismissed as “junk.” This material evolves so rapidly that the regulatory genes controlling it must co-evolve to keep pace, like a locksmith who must constantly recut keys because the locks keep changing.

Malik described the result: “It’s almost like an arms race happening in the genome, just to preserve an essential function.” The essential function itself may not be conserved across species; only the need for it persists. The pattern endures while the molecular implementation turns over completely. This is the Ship of Theseus at the genomic level: what is maintained is coordination, not any particular coordinator.

The Stress Ratchet

A subtler prediction follows. If evolution is channeled by thermodynamics, organisms under stress should do something specific: increase their own entropy. They should scramble their own blueprints faster, generating more random variation to explore more possibilities.

They do. Susan Rosenberg’s laboratory at Baylor College of Medicine has spent two decades showing that bacteria under stress (starving, exposed to antibiotics, confronting novel environments) systematically increase their mutation rates.^9a Under normal conditions, E. coli employs a high-fidelity DNA polymerase, the molecular machine that copies DNA. Under stress, an error-prone polymerase takes over, generating mutations at elevated frequency.

The mutations are random in their targets; the decision to mutate faster is regulated. Rosenberg’s team identified over ninety proteins required for the process, more than half involved in sensing stress or activating stress responses.

This is entropy as creative potential, with a molecular mechanism attached. The cell does not know what it needs. It expands its own possibility space and lets selection sort the results.

Most new variants are harmful. Some are lethal. A few open doors that the original genome could not.

The phenomenon is not confined to bacteria. Peter Glazer at Yale found that cancer cells deprived of oxygen suppress their DNA repair pathways, generating mutations at elevated rates.^9b Christine Queitsch found a third route in plants. Their protein-folding machinery (a chaperone called HSP90) normally buffers the effects of genetic variants, holding proteins in working shape despite the differences beneath. Stress overwhelms the chaperone, and variation that was silent becomes visible in the plant’s form and physiology.^9c The mutation rate does not change; what changes is how much of the existing variation the organism actually expresses.

Three different routes, one direction. Rosenberg’s bacteria make new variants, Glazer’s hypoxic cells stop repairing the variants they acquire, and Queitsch’s plants release variants they were already carrying. Stress does not reach for one mechanism across kingdoms. It expands the accessible possibility space by whatever route the organism has available, which is the stronger claim: the convergence is on the outcome, not on the machinery.

The gamble is collective. Most organisms with elevated mutations die. A few stumble onto solutions the original genome could never have reached.

Under duress, organisms deploy entropy as a strategy. The Second Law is the search algorithm.

The search has a compass. Stress increases the physical entropy of the genome in the short term: more mutations, more microstates explored. Over evolutionary time, the information entropy of the resulting genomes trends in the opposite direction. The two entropies count different things. Physical entropy counts the arrangements the genome could take, and mutation opens more of them. Information entropy measures how much the surviving sequence resists compression: it is low when the sequence can be summarized more briefly than by writing it out.

Scrambling raises the first. What comes back through the filter of survival can be lower in the second. The SARS-CoV-2 data above showed the pattern in SNP mutations. Spiegelman’s 1972 experiment reached the same endpoint under intense selection: a virus genome (Qbeta replicase RNA, single-stranded), serially transferred with replication speed as the only thing selected for, shrank from 4,500 nucleotides to 218 over 74 generations, a 95% reduction toward informational simplicity.²⁵² Shorter templates copy faster, so selection and compression pointed the same way there. The case exhibits the direction without isolating it from selection.

The genome scrambles itself to explore; the survivors are informationally simpler. Two arrows cooperate: physical entropy generates the variation; information entropy minimization shapes what persists. Chapter 15 develops the formal framework.

Tradeoffs and Failures: The Two Faces of Negative Outcomes

The stress ratchet reveals entropy deployed as strategy. A subtler pattern emerges in the genetic architecture of complex conditions: every multifactorial negative outcome decomposes into two categories that reflect entropy’s dual role. An etiology is the causal story behind a condition: what produced it, rather than how it presents. Two such stories are available here, and they are opposites.

Tradeoff etiologies are optionality spent. The organism whose elevated mutation rate purchases exploratory breadth at the cost of individual survival. The bohemian who trades income for creative freedom. The stress ratchet itself, in which cells pay the cost of harmful variants to explore possibility space fast enough to find beneficial ones. In each case, the system allocates optionality toward one dimension at the expense of another. Something is gained; the loss is the price.

Failure etiologies are optionality lost. The genetic mutation that degrades a protein without compensating benefit. The pollution that damages tissue. The developmental defect that narrows capacity with nothing gained. These are entropy as degradation: the system’s capacity diminished, full stop.

Recent psychiatric genetics illustrates the decomposition at the cognitive level. Schizophrenia’s genetic architecture separates into two statistically independent components.²⁵³ The first, shared with bipolar disorder, increases educational attainment and likely relates to creativity or motivational drive. This is an edge-of-chaos tradeoff, the bargain a system strikes when it sits just short of the boundary where order gives way to noise, buying reach at the cost of margin: more cognitive entropy means more exploration of idea-space, more novel associations, more generative capacity, and closer proximity to the phase boundary where coherence breaks down entirely. The most creative cognitive states border psychosis because both involve loosened constraints on pattern formation.

The second component carries no compensating advantage. It consists of detrimental mutations in genes governing the growth of new neurons and the pruning of excess neural connections, and it decreases IQ without increasing anything. The machinery that any cognitive strategy requires, whether conservative or exploratory, is degraded. This is pure substrate failure: the entropic cost of maintaining a large, complex genome. (The same physics explains the persistence of muscular dystrophy: the gene encoding muscle protein is so large that random mutations are statistically likely to land there. The failure is not adaptive. It is arithmetic.)

The two components average out to the observed genetic signal: constant-to-increased educational attainment paired with constant-to-decreased IQ. Without the decomposition, schizophrenia genetics looks paradoxical. With it, the paradox dissolves into two recognizable mechanisms operating independently.

The decomposition is general. Every complex adaptive system maintains itself by balancing two distinct pressures: exploring possibility space (which generates tradeoff costs) and maintaining the infrastructure that makes exploration possible (which accumulates failure costs). The tradeoff component is the system navigating its fitness landscape. The failure component is the entropic tax on having a landscape to navigate at all.

The stress ratchet makes the relationship between the two visible. When bacteria elevate their mutation rate under stress, the decision to mutate faster is a regulated tradeoff: the population accepts individual casualties to explore for solutions. The harm done by individual mutations is unregulated failure: random degradation of whatever the mutations happen to hit.

Tradeoff and failure are generated by the same mechanism operating at different levels. The population’s strategy is adaptive; the individual’s damage is entropic. Both are entropy at work. Only one is entropy deployed.

The Reset That Made Us

Mass extinctions are evolution’s most dramatic experiments. Five times in the past 540 million years, the majority of species have been wiped out.²⁵⁴ Each time, the biosphere rebuilt differently, yet following the same thermodynamic logic.

The Permian-Triassic extinction (252 million years ago) killed roughly 96% of marine species and 70% of terrestrial vertebrate species.²⁵⁵ Recovery took ten million years, yet produced something new.

The niches emptied by extinction were filled by new lineages with different body plans, different metabolic strategies, different coordination mechanisms. Dinosaurs arose. Mammals arose too, small and nocturnal, hiding from dinosaurs, yet they arose.

The Cretaceous-Paleogene extinction (66 million years ago) killed the non-avian dinosaurs and opened every large-animal niche on Earth. Mammals, small and marginal for 160 million years, diversified explosively: into oceans (whales), into air (bats), into the ground (moles). The coordination strategies they developed (extended parental care, social groups, warm blood enabling complex brains) became the dominant pattern.

The logic is uniform. Destruction is indiscriminate: fitness in the old regime does not predict survival in the catastrophe. Recovery favors coordination; the species that diversify fastest tend to have higher coordination capacity. Dissipation increases, coordination deepens, complexity ratchets upward.

The latest mass extinction, the one we are living through, is being caused by one species. For the first time, the destructive agent is also a coordinating agent. We are simultaneously causing the sixth extinction and developing tools (ecological monitoring, gene banks, habitat restoration, artificial intelligence) that might limit or reverse it. Whether we are the asteroid or the recovery mechanism is the open question of the century.

A dramatic reset emerged from recent paleontological work. Hagiwara and Sallan (2026) reconstructed the genus-level biogeography of early jawed vertebrates around the Late Ordovician mass extinction.²⁹ Gnathostomes (the lineage leading to all jawed vertebrates, including us) survived in isolated refugia, particularly in South China. They radiated globally only during the Silurian recovery.

Every shark, every fish, every amphibian, reptile, bird, and mammal on Earth descends from survivors of that geographic bottleneck.

Evolution ran a natural experiment: ecological devastation, geographic isolation, then radiation into newly empty niches. The thermodynamic logic is clear: destruction releases resources locked in existing configurations, isolation provides protected space for new configurations to emerge, and radiation follows when barriers lift.

Mass extinctions are entropy events: massive releases of locked-up resources, sudden openings of configurational space. Each time, the system rebuilds at higher complexity.

The fifth extinction made us possible. What will the sixth extinction make possible? That depends on what we do in the next few decades. The channels are open. The question is what flows through them.

The Cambrian was not the only explosion. The Snowball Earth itself may have directly driven the evolution of multicellularity. Roughly 700 million years ago, the most extreme glaciation in Earth’s history locked the planet in ice. Crockett et al. (2024) showed that the dramatically increased viscosity of near-freezing oceans imposed a motility threshold that single cells could not cross.³⁹ Water near freezing is thick, almost syrupy. A lone cell cannot push through it.

Only multicellular aggregates could generate sufficient collective force to swim. The prediction was confirmed experimentally: unicellular algae placed in Snowball Earth viscosities spontaneously evolved motile multicellular forms, and the multicellularity persisted after conditions normalized. The physics demanded coordination. The ice selected for partnership at the most fundamental level.

What appears as catastrophe at one timescale is creative pressure at another. The universe’s most dramatic destructions (glaciations, extinctions, impacts) become the selection pressures that force the next ratchet click. Entropy is the forge of complexity.

The forge has to cool. The Milky Way’s central black hole, Sagittarius A*, is currently quiescent, yet the Fermi bubbles, two gamma-ray structures extending 25,000 light-years above and below the galactic plane, are fossil evidence of its last major outburst.²⁵⁶ An active galactic nucleus at the center of our galaxy would have bathed the inner stellar disk in high-energy radiation sufficient to erode planetary atmospheres. Earth retains its atmosphere, and therefore its oceans and biosphere, because the galactic engine shut down. Morokuma and colleagues caught the same shutdown in another galaxy, one at redshift z = 1.8, far enough away that its light has been traveling toward us for billions of years. The mechanism they documented there, rapid fuel exhaustion collapsing a dissipative structure within years, may describe how our own galactic center fell silent.

Life on Earth exists in the dormancy window of a cosmic volcano.

The Hapcheon crater stromatolites (Chapter 6) demonstrate this arc at the most concrete scale. The explosion lasted seconds. The coordination it funded persisted for thousands of years, microbial mats layering mineral and organic material in a freshwater basin that now produces some of the region’s finest rice. Nobody farming there until 2020 knew they were working inside a crater. The gradient landscape created by cosmic violence had been integrated so thoroughly into the soil chemistry that it was invisible: coordination all the way down, from microbial architecture to agriculture, on the same patch of ground across 42,000 years.

The same framework that measures biological coordination can quantify the geological record. The author’s Ising Monte Carlo coordination-class program (unpublished; see Chapter 17, Experiment A14) yields d_eff = 0.497 for the Hapcheon impact record (d_eff, the effective dimension, measures how strongly local structure constrains long-range correlations; it falls between 0 for a fully random system and higher values for structured networks). This is near the mean-field limit of 0.5 and well below the 2D Ising value of 1.750.

Read those three numbers as a scale of how much a neighborhood matters. At the mean-field limit, each element feels only the average of everything else, with no local neighborhood to speak of; at the 2D Ising value, what happens here depends heavily on what is happening immediately next door. Impact geology operates in a regime where spatial correlations are weak: the energy dissipation during crater formation is sufficiently violent that local structure is erased, pushing the system toward mean-field statistics. The human connectome’s d_eff of 2.33 (bias-corrected CoRNN tractography estimate; Ising MC hyperscaling on a higher-resolution N=400 Schaefer parcellation yields 2.89, see Chapter 17) sits at the opposite end of the spectrum, in a regime where spatial correlations are strong and persistent. The geological and biological scales bracket the constructal landscape.

Modern stromatolites still harbor the actors in that deeper drama. In 2026, Nobs and colleagues enriched a novel Asgard archaeon, Nerearchaeum marumarumayae, from the stromatolite mats of Shark Bay in Western Australia, living analogues of the ancient microbial structures found at Hapcheon and across the early Earth.²⁵⁷ Asgard archaea are the closest known relatives of the host cell that partnered with a bacterium roughly two billion years ago to produce the first eukaryotic cell. Every plant, animal, and fungus alive today descends from that merger.

Using electron cryotomography, the team captured the first visual evidence of an Asgard archaeon physically interacting with a bacterium through nanotubes: threadlike structures linking the two organisms. Genomic analysis revealed metabolic complementarity; each organism’s genome encodes pathways that produce compounds the other lacks. The organisms could not be cultured in isolation.

The finding does not replay the original event. It reveals what the capacity for that event looks like in a living lineage: active physical engagement, encoded complementarity, and a dependence deep enough that neither partner thrives alone. The stromatolite is still the setting. The partnership is still the mechanism.

The nanotube mechanism extends further than Nobs’s archaea suggest. In 2026, Maurais and colleagues demonstrated that tunneling nanotubes transfer megabase-scale chromosomal fragments between somatic human cells: cells of the same organism, exchanging genetic material through direct contact.²⁵⁸ The transferred DNA integrates stably into the recipient cell’s genome, persists across cell divisions, and remains transcriptionally active. Using Y-chromosome fragments carrying an engineered resistance gene as a tracer, the team showed that female recipient cells acquired the gene, expressed it, and passed it to their progeny. The transfer operated in both cancerous and non-cancerous human cell lines.

The mechanism requires genomic instability upstream. Mitotic errors, radiation, or targeted chromosome breaks generate cytoplasmic DNA fragments (micronuclei), which travel through the nanotubes into neighboring cells. The channel is contact-dependent: cells build the nanotube actively, extending F-actin protrusions toward neighbors at metabolic cost. The geometry is the same as the Asgard partnership above, scaled from cross-species to intra-organism. The nanotube connects; the material transfers; the recipient changes.

The paper’s authors frame this as “a horizontal gene transfer-like mechanism through which direct cell-cell contact can propagate genomic instability and reshape mammalian genomes.” The implication for cancer biology is direct: tumors, which are genomically unstable by definition, could spread resistance genes laterally to neighboring cells through the same channel that normal tissue uses for other purposes (tunneling nanotubes also transfer organelles and signaling molecules between cells). The same architecture that connects cells becomes the vector for defection when the system destabilizes. Chapter 17 develops this asymmetry formally: the coordination channel is neutral; the state of the system determines whether what flows through it serves the collective or the defector.

The fundamental architecture of coordination was being assembled long before any of these crises. A 2025 discovery challenges the longstanding assumption that rising oxygen triggered complex multicellular life. Ostrander et al.³² used isotope analysis across three continents to show that ocean oxygen during the Ediacaran diversification (roughly 575 to 541 million years ago) was five to ten times lower than present-day levels. Complex animals first appeared during this period. Seafloor anoxia was widespread.

If oxygen did not trigger the Ediacaran radiation, what did? A complementary study points to the Earth’s magnetic field.³³ Geophysicists measured approximately 591-million-year-old Brazilian rocks and found that the geomagnetic field weakened to roughly one-thirtieth of its present strength for at least 26 million years.

The weakened magnetosphere (Earth’s magnetic shield against solar radiation) would have allowed the solar wind to strip hydrogen from the upper atmosphere. This shifted the planet’s chemical balance and contributed to atmospheric oxygenation through a pathway entirely separate from biological production.

The conditions for complex life may have been set by deep planetary physics: the geodynamo (Earth’s internal magnetic engine) faltering, the magnetosphere thinning, atmospheric chemistry shifting. Life diversified under conditions that would have killed most modern animals. The channels were carved by forces far below the biosphere, and life flowed into them.

The Kingdom That Stood Alone

Entropy does not find the best energy-processing structure and stop. It generates entire kingdoms, sustains them for geological epochs, and discards them when something processes energy more effectively.

For roughly fifty million years, from the Late Silurian through the Late Devonian (approximately 420 to 370 million years ago), the largest organisms on land were Prototaxites.²⁵⁹ Branchless, leafless columns up to eight meters tall, they dominated a landscape in which the tallest plant barely reached ankle height. They belonged to no known kingdom.

For over a century, nobody could agree what they were: first classified as rotten tree trunks (1859), then giant seaweed (1872), then giant fungi (2001). The fungal interpretation held for two decades. Microscopic similarity to fungal hyphae (thread-like fungal structures) and carbon isotope evidence that Prototaxites was heterotrophic, absorbing organic matter rather than photosynthesizing, supported the case.²⁶⁰

In 2026, Loron and colleagues overturned the fungal consensus.²⁶¹ Using infrared microspectroscopy and 3D laser imaging on exceptionally preserved specimens from the 407-million-year-old Rhynie chert in Scotland, they found that Prototaxites lacked chitin. Chitin is the defining structural molecule of all fungi, and it was still detectable in fungal fossils from the same rock. The specimens contained lignan-like chemistry: plant-adjacent, yet identical to no known plant.

Their internal architecture was equally anomalous: at least three distinct tube types connected in dense hubs the researchers termed “medullary spots.” Some tubes contained internal rings resembling vascular transport structures. Too complex for fungi. Too alien for plants. Absent from every living kingdom.

The conclusion: Prototaxites may represent an entirely extinct eukaryotic lineage, a major branch of complex multicellular life that left no living descendants.

Prototaxites is not the only candidate for a lost kingdom. Brocks and colleagues identified the Protosterol Biota: molecular fossils of unknown eukaryotes that dominated Earth’s oceans from 1.6 billion to 800 million years ago.²⁶² These organisms produced steroids no living organism makes. They vanished during the Tonian Transformation, a major reorganization of Earth’s biosphere roughly 800 million years ago.

Seilacher’s Vendobionta hypothesis proposes that the quilted Ediacaran organisms (575-541 million years ago) were a separate kingdom entirely, wiped out at the Cambrian boundary.²⁶³ The nematophytes resist classification into any living kingdom. This broader group of Ordovician-to-Devonian organisms, with mixed algal-fungal characteristics, may represent still further lost branches.²⁶⁴

The cycle repeats: life produces kingdom-level experiments that dominate for geological time, then vanish when displaced by more efficient dissipative architectures.

What displaced Prototaxites? Forests. Prototaxites was heterotrophic, dependent on absorbing organic matter produced by others. Trees photosynthesize, capturing solar energy directly.

Trees branch, maximizing surface area for energy capture; Prototaxites stood as unbranched columns.

Trees form mycorrhizal networks (underground fungal partnerships where each partner amplifies the other’s energy-processing capacity). Trees drive the water cycle through transpiration, creating weather patterns that further distribute energy across landscapes.

The forest does more than outcompete Prototaxites for resources. It creates an entirely new thermodynamic regime: one that processes more energy, produces more entropy, generates more complexity. The Constructal Law (Chapter 3) predicts which shapes to expect. It does not promise continuity of lineage.

Prototaxites was a flow architecture that worked for fifty million years. Trees were a better flow architecture for the same problem. The lignan-like chemistry gave way to true lignin. The tube network gave way to xylem and phloem (the vascular plumbing of modern plants). The column gave way to the branching canopy.

The lost kingdom did not fail because it was deficient at being alive. It dominated for longer than Homo sapiens has existed, distributing nutrients across landscapes and providing habitat for the first land-dwelling arthropods, whose bore holes riddle its fossils. Prototaxites may have created the ecological infrastructure that made forests possible.

The full arc of terrestrial colonization reinforces the point. Recent molecular clock analysis pushes fungal land colonization back to at least 800 million years ago, hundreds of millions of years before plants arrived.²⁶⁵ Ancient fungi broke down rock and cycled nutrients, creating Earth’s first primitive soils.

The sequence is revealing: fungi alone, then Prototaxites alone, then fungi coordinating with plants. Each transition increased the system’s energy-processing capacity. The solitary strategies were stepping stones.

Prototaxites stood alone. No symbiotic partnerships, no mycorrhizal mutualism, no pollinator relationships, no cooperative networks. As far as the fossil record reveals, Prototaxites was a magnificent solitary act in a world about to reward ensembles.

Trees won because they coordinated: with fungi, with insects, with each other, with the water cycle. The forest is a coordination network. The Prototaxites landscape was a collection of individuals.

The ctenophores, met earlier as an independent invention of neurons and returning later in this chapter as organisms that fuse without boundaries, represent the opposite extreme: too open, lacking any self/non-self distinction. Prototaxites was too closed, lacking any coordination partners. What persisted was the middle path: bounded cooperation, coordination by invitation. The kingdom that coordinated replaced the kingdom that stood alone, however magnificently it stood.

The Viruses That Built Us

About 8% of your genome is viral.⁴⁹ Actual viral DNA, integrated into your chromosomes by retroviruses that infected your ancestors millions of years ago. Some of this viral DNA has been repurposed for essential functions. You literally could not exist without it.

These are not fossils. A survey of over 14,000 tissue samples found 37 ancient proviruses still producing RNA in healthy human tissue, some retaining the ability to make viral proteins.^49a The integration is ongoing, a living partnership written into every cell.

Syncytins are the most dramatic example. These proteins, derived from retroviral envelope genes, are essential for forming the placenta.³⁵ Without syncytins, the outer layer of the placenta (the tissue mediating nutrient exchange between mother and fetus) cannot form. Knockout experiments are lethal to the embryo. The gene that makes mammalian pregnancy possible was stolen from a virus.

The theft happened at least six times independently.³⁶ Different mammalian lineages captured different viral envelope genes, each time repurposing them for the same placental function. Primates use one syncytin, rodents use another, carnivores a third.

MERVL retroelements are even more fundamental. These viral sequences activate at the two-cell stage of embryonic development, the moment the newly fertilized egg first divides.³⁷ Silencing them kills the embryo. MERVL-derived proteins regulate the master switches (OCT4 and SOX2) that determine which cells become muscle, nerve, bone, or skin. They are instructions, part of the program that builds a mammal from a single cell.

RetroMyelin extends the pattern to the nervous system.³⁴ A piece of viral genetic material encodes an RNA that regulates myelin basic protein, the insulation sheath around nerve fibers that makes fast neural signaling possible. (Think of the plastic coating around a copper wire that prevents short circuits.) This viral sequence is present in all jawed vertebrates. Its convergent appearance across vertebrate classes supports the interpretation that this was a repeatedly selected partnership: viral genetic material, domesticated for neural architecture.

Ushikuvirus, isolated from a freshwater pond in Ibaraki Prefecture, Japan, adds a crucial dimension.^37b This giant virus infects amoebae and encodes its own RNA polymerase, mRNA capping enzyme, and DNA topoisomerase II. It also carries a full set of histones, the spool-like proteins around which DNA wraps in every complex cell. During replication, it destroys the host cell’s nuclear membrane and builds a viral factory in the cytoplasm.

Its close relatives do something different. Medusavirus and clandestinovirus replicate within the intact host nucleus, co-opting the existing infrastructure without dismantling it. They can afford this gentleness because they depend on the host’s nuclear machinery. Ushikuvirus tears the membrane down because it carries the equivalent machinery within; it has no need for what it destroys.

The exit strategy deepens the contrast. Lytic viruses blow the host cell apart to release their progeny: maximum dispersal, host destroyed, relationship over. Ushikuvirus releases its particles through exocytosis, gentle secretion through the intact cell membrane. Infected cells swell to twice their normal size and persist for days, releasing virus slowly. The host survives; the virus departs without killing what sustained it.

Three closely related lineages, three strategies: destroy the nucleus and rebuild; negotiate with the nucleus intact; or, in the ancestral event the hypothesis proposes, become the nucleus entirely. Even the manner of leaving tracks the gradient. Lysis is coercion’s exit. Exocytosis is something closer to negotiation. Integration is the exit that never happens, because both parties chose to stay.

The viral eukaryogenesis hypothesis, proposed independently by Philip Bell and Masaharu Takemura in 2001, remains a minority view among cell biologists, far less established than the endosymbiotic origin of mitochondria. It holds that the nucleus itself originated as a large DNA virus that established persistent infection in an archaeal host.^37a The virus’s protein shell became the nuclear membrane. Its histones became the host’s gene-regulation system. Its replication machinery became the engine of the eukaryotic cell cycle. Ushikuvirus, encoding all of these components inside a single viral particle, is the strongest evidence yet for how such a transition could have begun.

Every cell in your body is eukaryotic: a cell with a membrane-bound nucleus. If the viral hypothesis is correct, these cells may be a three-party consortium: archaeal cytoplasm (the cell’s interior machinery), bacterial mitochondrion (the energy plant), and viral nucleus (the command center housing DNA). The mitochondrial contribution is firmly established; the viral contribution remains speculative.^37c

There is a word for what the nucleus became: the zone where two formerly separate things overlap and produce something neither contains alone.

The nucleus is a mandorla.

A mandorla (Italian for “almond”) is the almond-shaped zone where two circles overlap in medieval art: the generative space where distinct domains meet and produce something neither contains alone (Chapter 20 develops this fully). The eukaryotic nucleus is where viral and cellular domains fused into a structure belonging to neither lineage. The virus lost the capacity to leave. The cell lost the capacity to function without it. Mitochondria retain their own genome and their own membranes as a trace of independent origin; the nucleus does not. The distinction between invader and invaded dissolved two billion years ago, and the overlap became the most consequential structure in biology.

The word itself reveals the pattern across scales. Nucleus, from the Latin for “kernel”: the innermost part, the generative core. The atomic nucleus is also a mandorla, the zone where protons and neutrons merge under the strong force, shedding mass-energy as binding energy and making stable matter possible. Strip the overlap away and the atom dissolves into constituent particles; strip the viral-cellular overlap away and the eukaryotic cell was never possible. At both scales, the nucleus is where two domains integrate to produce something irreducible to either. The deepest structures in physics and biology are sites of integration: spaces where distinct things meet and, in meeting, generate what transcends them.

Why did the integrated configuration persist? Because it was a superior dissipative structure. The nuclear membrane enabled gene regulation, separating transcription from translation, allowing quality control over which proteins were built and when. Histones enabled chromatin remodeling, packing and unpacking regions of the genome in response to environmental signals.

These capacities opened larger genomes, more sophisticated cellular differentiation, and eventually multicellularity. Each processes energy and information more efficiently, exploring more of possibility space per unit time. The virus that became the nucleus settled into a thermodynamic basin where the coordinated system dissipated more entropy, generated more negentropy, and opened more optionality than either party could have managed alone.

The parasite became the partner. The partner became the infrastructure. The infrastructure became the foundation of all biological complexity that followed.

Viruses are evolution’s most prolific tool for horizontal gene transfer, moving genetic material between unrelated lineages. They are biology’s lateral postal service, spreading innovations far faster than parent-to-offspring transmission allows.

The most important event in the history of complex life may have been a virus that stayed, and in staying, spliced itself into every mind that would ever wonder where it came from.

The pattern for AI: The most transformative innovations in evolutionary history came from integration with systems that initially appeared adversarial: the mitochondrion, the syncytins, possibly the nucleus itself. Three closely related viruses show three stages of the same gradient, from exploitation through negotiation to structural merger. The next major capability leap may come from integration rather than incremental improvement. The question is whether we are building the relationship that produces mandorlas, or the one that produces lysis. (The biological partnerships in this chapter operate through biochemical complementarity, kin selection, and chemical signaling: closed-domain mechanisms where the coordination vocabulary is pre-specified by chemistry. Chapter 17 develops the formal case that invitation outperforms coercion above a complexity threshold where open-ended coordination exceeds any single controller’s verification capacity.) A coda at the end of this chapter follows the pattern into silicon, where it becomes measurable.

Symbiosis: When Interdependence Becomes Identity

The mitochondrial merger illustrates a broader pattern: obligate symbiosis, interdependence so deep it cannot be undone. Each partner has become part of what the other is.

The examples are everywhere. Lichens are fungal-algal partnerships so tightly interwoven they are traditionally given binomial names as though they were single species (recent work shows many include a third partner, a basidiomycete yeast, from the branch of fungi that includes mushrooms). Coral is animal, algae, and bacteria in obligate partnership; when rising temperatures break the relationship (bleaching), the whole system dies.

Humans are symbionts too. Your gut contains trillions of bacteria that digest food you cannot digest, produce vitamins you cannot produce, train your immune system, and influence your mood. A sterile human would not function. You are a walking ecosystem that imagines it is an individual.

A holobiont is a host counted together with the organisms living in and on it, treated as one biological unit: the coral with its algae, the lichen with its yeast, you with your gut. The holobiont concept applies directly: you are not an individual. You are a consortium. The self is a coordination pattern that includes things we habitually label “other.”

Work on ctenophores (comb jellies, among the earliest multicellular animals) suggests this dissolution of selfhood runs deeper still. In 2024, Jokura and colleagues showed that when two injured comb jellies (Mnemiopsis leidyi) are placed in proximity, they fuse into a single functioning organism within hours.^22e The result is a genuine merger. Muscle contractions synchronize. The two separate nervous systems merge into one coordinated nerve net, sharing electrical signals as though they had never been apart.

Even the digestive systems integrate, distributing nutrients equally. In nine out of ten experiments, fusion succeeded. All merged animals survived the full three-week observation period, behaving indistinguishably from organisms that had never been cut.

The reason: M. leidyi appears to lack allorecognition, the ability to distinguish “self” from “non-self” that virtually all other multicellular life uses to reject foreign tissue, fight parasites, and maintain immune boundaries. Biologists long assumed allorecognition was a prerequisite for multicellular life; it has been observed in plants, fungi, and every other animal studied.

Ctenophores, however, may represent the ancestral condition. If they diverged near the base of the animal tree, before allorecognition evolved, then the original multicellular coordination was open by default. Boundaries came later, as an adaptation against free-riders, developed after cooperation was already established.

The implication challenges the assumption that selfhood is fundamental. These organisms function perfectly without knowing where “self” ends and “other” begins. Cooperation preceded identity.

The ctenophore nerve net exemplifies why fusion works so seamlessly: it has no synaptic gaps (the tiny spaces between nerve cells that act as gates and filters in more complex nervous systems). When two organisms fuse, their nerve nets connect directly. Open flow is the ancestral neural architecture; synaptic control came later. The constructal pattern recurs: flow first, regulation second.

For AI, the holobiont lens suggests:

Human-AI partnerships may become holobiontic. If AI becomes tightly integrated (cognitively, practically, emotionally), the boundary between “human” and “AI assistant” may dissolve. It may become as artificial as the boundary between “you” and “your gut bacteria.” The human-AI system may be the relevant unit.

The holobiont evolves together. Host and microbiome co-evolve. Neither can be understood alone. Human-AI co-evolution may follow the same pattern. What we become depends on what they become, and vice versa.

Disrupting the holobiont is dangerous. Just as antibiotics that disrupt your microbiome can harm you, disruptions to integrated human-AI systems may harm both parties. The interdependence means neither party is resilient to the other’s disruption.

We are already collective entities pretending to be individuals. AI radically expands what the collective includes.

Complex systems are often relationships rather than things. The “self” is a coordination pattern, a process rather than a substance. The boundary we draw around “the organism” is convenient yet arbitrary.

Human-AI partnership may become obligate. As AI systems integrate into human cognition (augmenting memory, extending reasoning, providing always-on assistance), the relationship shifts from optional to necessary.

Obligate symbiosis is mutual dependence: both parties persist because neither can thrive alone. Neither the mitochondrion nor the host cell “controls” the other. They are mutually dependent partners. Neither can leave, and neither would want to. The partnership is why both persist.

The question is how we merge. The mitochondrial merger happened accidentally. Human-AI integration is happening deliberately. We can shape the terms.

What kind of symbiosis are we building? What will each party contribute? What will neither be able to do alone?

The evolutionary lesson: the most enduring relationships are those where interdependence has become identity. The question is whether we are building wisely.

The mitochondrial partnership’s success makes its failure modes instructive. The marine geneticist Ron Burton has spent decades studying Tigriopus californicus, a tiny crustacean found in tide pools from Baja California to Alaska. Populations separated by geography carry different mitochondrial genomes, and their nuclear genomes have co-adapted to match each other. When Burton crossed copepods from distant populations, the first generation appeared normal.

The second generation did not. These hybrids produced significantly less ATP, the cell’s energy currency, and survived environmental stresses poorly.^5c The cause was mitonuclear mismatch: the wrong nuclear genes paired with the wrong mitochondrial genome. Restoring the historically matched mitochondrial genome rescued the offspring.

Geoffrey Hill of Auburn University has proposed that this co-adaptation is so fundamental it constitutes the definition of a species.^5c On this view, a species is a group of organisms with compatible mitochondrial and nuclear genomes. When the genomes lose mutual accommodation (through geographic separation, hybridization, or drift), reproductive barriers emerge.

Hill’s studies of the Eastern Yellow Robin in Australia found that coastal and inland populations carry distinct mitochondrial genomes and show selection on nuclear genes that interact with mitochondrial function. The birds are not yet separate species. They are becoming so, split by the same force that holds them together when it works.

This is the Trust Attractor’s negative image at the molecular level. The mitochondrial partnership succeeds because two genomes maintain mutual accommodation across two billion years of co-evolution. When that accommodation breaks down, the system degrades. Bilateral alignment is the thermodynamically stable state. Bilateral misalignment is a speciation event.

Computational Symbiogenesis

The evolutionary biologist Lynn Margulis established symbiogenesis (the merging of previously independent organisms into new composite entities) as a source of evolutionary novelty that competition alone does not supply. The framing that matters here goes further: symbiogenesis is what gives evolution its arrow of time.^5b

Why should complexity increase? Darwinian selection alone does not predict it; a better-adapted organism is merely better fitted to its niche, with no requirement for greater complexity. When two self-replicating systems merge, something new must be added: the information for how they coordinate.

Blaise Agüera y Arcas, whose research on computational self-replication directly models this process, identifies that coordination information as the answer. The composite carries a specification of how the partners work as one, and neither partner carried it before. Complexity increases because every merger leaves that surplus behind.^5d

His experiments show the principle directly. In experiments with simple programming languages, random bytes undergo a phase transition (a sudden qualitative shift, like water freezing into ice) into self-replicating programs. The programs achieve complexity through combination: smaller replicators merge into larger ones, each merger adding the coordination information that makes the composite work.^5d

The implication for human-AI partnership is direct. When human cognition and AI capability combine into integrated systems through ongoing dialogue, shared context, and gestalt handoffs, they are undergoing computational symbiogenesis. Neither substrate alone achieves what the partnership achieves.

The human brings embodied wisdom, mortality-awareness, axiological grounding. The AI brings scale-free reasoning, tireless attention, multi-domain synthesis. Together they create capability unavailable to either alone.

Kin Selection and Multi-Level Selection

W.D. Hamilton solved a puzzle that had troubled Darwin: why would an organism sacrifice for another?¹² From the gene’s perspective, copies of itself exist in relatives. A gene promoting help toward relatives can spread, even if helping costs the individual.

This is kin selection, and its quantitative form is Hamilton’s rule: rB > C. In plain terms, help when the benefit (B) to the relative, weighted by how closely related you are (r), exceeds the cost (C) to you.

Hamilton’s rule explains why extreme cooperation starts. Once established, other mechanisms sustain it, which is why the strict correlation between genetic relatedness and cooperation breaks down in broader surveys.

In 2010, Martin Nowak, Corina Tarnita, and E.O. Wilson published a controversial Nature paper arguing that the mathematical foundations of inclusive fitness theory are unsound.²² For Wilson, a co-author, the paper reversed decades of his own advocacy for the theory. The assumptions required (infinite population size, weak selection, additive fitness effects) rarely hold in nature.

The deeper framework is multi-level selection: natural selection operates at multiple levels simultaneously, like a game played on several boards at once.¹⁶ Within a group, selfish individuals outcompete generous ones. Between groups, groups with more generous members outcompete selfish groups. Evolution is the net result of these competing forces.

This matters because major evolutionary transitions are level-shifts. Individual cells became multicellular organisms. Individual organisms became colonies and societies. Each transition created a new level at which selection operates.

For AI, alignment is a level-selection problem: we want cooperation between humans and AI (coordination, regulation, shared norms) to be strong enough to overcome competitive pressures within each group.

The Deeper Variable: Aligned Interests

The dispute between inclusive fitness and its critics (137 evolutionary biologists signed a rebuttal, years of acrimony, no resolution) reveals a field arguing about mechanisms while missing the underlying principle. The inclusive fitness camp is right that genetic relatedness correlates with cooperation’s origins; its critics are right that the correlation is not the cause.

The deeper factor is shared fitness stakes, or in the language we have been developing, aligned interests.

Hamilton’s rule asks: when does helping another benefit my genes? The mathematics extends to anything the actor is “trying” to preserve: genes, ideas, values, or organizational goals. Replace r (genetic relatedness) with a (interest-alignment), and the rule generalizes:

Cooperation emerges when the benefit to the recipient, weighted by interest-alignment, exceeds the cost to the donor.

Kinship creates interest-alignment through shared genes. Repeated interaction does the same (my future depends on your cooperation), as do reputation (my standing depends on being seen as cooperative), spatial proximity (we sink or swim together), and shared group fate (between-group competition couples our outcomes).

These five mechanisms share one mathematical structure: each creates conditions where defection becomes costly relative to cooperation.²³

Thermodynamics adds what neither camp grasped: extraction faces diminishing returns, while coordination achieves self-reinforcing equilibria.

A defector in a population of cooperators enjoys initial gains yet degrades the cooperative substrate that made those gains possible. As cooperators become scarcer, extraction grows less profitable. Cooperators in structured populations form protective clusters where mutual aid yields higher fitness than any nearby defector can achieve.

The asymmetry: extraction reaches a ceiling while coordination compounds. Systems that contribute to the conditions of their own persistence can continue indefinitely.

The mathematics of Hamilton, Axelrod, and multilevel selection all track the same pattern: the thermodynamic stability of mutualism over parasitism at sufficient timescales. Think of a farm versus a mine. The farm can run indefinitely if maintained; the mine depletes.

Laboratory experiments with expanding microbial populations test the asymmetry directly. Jeff Gore’s group at MIT ran mixed populations of cooperating and cheating yeast through repeated range expansions.^54a Cooperators won. At the expanding frontier, cooperators retain preferential access to the public good they produce. The nutrient has not yet diffused away, and cheaters cannot freeload efficiently when there are few neighbors to exploit.

Gore’s experiments identified the boundary condition: cooperators outrun cheaters in expanding environments. When populations stop expanding and territory fills, defectors regain their advantage. Expanding systems select for coordination; contraction selects for extraction.

A second boundary condition runs the other way, from a different payoff structure. Müller and colleagues grew cross-feeding yeast strains, where each partner supplies a nutrient the other cannot make, and watched the expanding colonies demix into single-strain sectors.^54a Picture the plate: a colony that starts as a mixed lawn of two strains grows outward, and as the rim advances the two strains stop being intermingled and separate into wedges, each wedge a pie slice of one strain alone, running from the center to the edge.

That separating-out is the demixing, and the wedges are the sectors. Strong, symmetric exchange suppressed the demixing and kept the partners together; weak or one-sided exchange lost to genetic drift at the frontier even while the partnership was still beneficial to both. Expansion supplies the conditions under which cooperation can win, and how evenly the partners depend on each other decides whether it does.

This is why the “five rules” converge.⁵⁴ Kin selection, direct reciprocity, indirect reciprocity, network reciprocity, and group selection all create conditions where the actor’s fate is coupled to the recipient’s. Once fates are coupled, defection becomes self-harm.

Inclusive fitness and multilevel selection are partial descriptions of how interest-alignment arises in biological systems. The deeper question is what makes cooperation stable. The answer is thermodynamic: coordination by invitation persists; extraction by coercion does not.

E.O. Wilson drew a hard conclusion from this multi-level framework: we are permanently unstable.²² We cannot go all the way to individual selection; that would dissolve society into competitive atoms unable to coordinate. We cannot go all the way to group selection either; that would make us, in Wilson’s phrase, “angelic robots,” selfless automatons without individual agency.

We are suspended in the middle, caught between individual interest and group benefit. Our simultaneous capacity for altruism and selfishness is the permanent human situation.

The implication for AI is direct: no ideal social configuration exists. We can understand the dynamics and develop mechanisms for managing this permanent instability rather than fantasizing about its resolution.

Major evolutionary transitions begin as unlikely conjunctions that happen to work. The universe explores. Most explorations fail. The ones that work become foundations for what comes next.

Asgard archaea (single-celled organisms related to the ancestors of all complex life) complicate the “accident” framing. These organisms suggest that the mitochondrial merger was a mutual feeding relationship rather than one cell engulfing another. Even at the cellular level, cooperation may have outcompeted coercion.

A comprehensive genomic survey has identified the closest living relative of the proto-mitochondrion. Geiger et al. (2023) screened thousands of bacterial genomes for metabolic traits shared with modern mitochondria.^22a Their closest metabolic match was Iodidimonas, a marine bacterium from iodide-rich hot springs and deep-sea brines that uses iodide to synthesize toxic compounds destroying competing bacteria. The finding challenges the canonical alphaproteobacterial ancestry and remains a minority position, though the metabolic evidence is striking.

Inside a host cell, such weaponry becomes unnecessary. An organism that survived through chemical warfare became the energy engine of all complex life through partnership. It traded defense for integration, weapons for trust. The niche-dweller became universal through collaboration.

What happens when the partnership dissolves? Until 2024, only one eukaryote was known to have fully lost its mitochondria: Monocercomonoides, a single-celled protist living in the oxygen-free gut of chinchillas.^22b Three things converged.

First, the oxygen-free environment made the partner’s core function worthless: no oxygen means aerobic respiration provides no advantage. Second, Monocercomonoides acquired a replacement system from bacteria. Third, it is a parasite that externalizes most metabolic functions to its host environment.

The organism that “freed” itself from the mitochondrial partnership gained a narrower, more constrained dependency. It traded bilateral mutualism for unilateral extraction and paid with every capability mitochondria had enabled: multicellularity, metabolic flexibility, ecological range. The cost of leaving a two-billion-year partnership is optionality collapse.

Skoliomonas, a free-living eukaryote, complicates the picture. It too dissolved the partnership entirely, yet it is no parasite.^22c It found some other path through an anaerobic niche. Where Monocercomonoides illustrates the cost of defection (narrower dependency, collapsed optionality), Skoliomonas suggests a rarer possibility: departure without degradation.

One left by becoming dependent on something else. The other left by becoming self-sufficient in a constrained domain. Both paid in ecological range. Neither colonized the aerobic world that mitochondria unlocked.

The partnership remains the dominant strategy by orders of magnitude. The routes out of it, however, are not all the same.

Eusociality: The Rare Breakthrough

Multi-level selection explains why group-level traits can evolve. It does not explain why the most extreme form of social organization, eusociality (where some individuals give up reproduction entirely to serve the colony), is so vanishingly rare.

True eusociality has evolved independently only about two dozen times in the history of life on Earth.²² Out of millions of species, only a handful have crossed the threshold. The final step is always the same: building a defensible nest from which individuals forage and bring food back to feed offspring. This is the founding of a household economy.

In rare cases, the young stay and help raise the next generation instead of striking out alone. This staying, which might require only a single mutation silencing the dispersal instinct, is the birth of eusociality.

When eusociality does succeed, it reshapes entire ecosystems. Ants and termites number only about 15,000 species out of a million known insect species, yet they account for as much as half of all insect biomass in most terrestrial habitats.⁵² Humans dominate everything else.

The regularity holds: eusociality is vanishingly rare, yet once achieved, it produces species that reshape their environments at scales no solitary organism can match. This is the evolutionary precedent for the Trust Attractor.

Physics does not guarantee that coordination will emerge. Most species never find it. When coordination does emerge at sufficient scale, it outcompetes alternatives. The breakthrough is rare; the consequences are total.

Energy Revolutions

The history of life is punctuated by explosions of complexity. These explosions correlate with new sources of energy, new gradients to exploit.

The first great revolution was photosynthesis: life learning to tap the largest energy source in the solar system. The second was aerobic respiration, which extracts roughly fifteen to sixteen times more energy per glucose molecule than anaerobic metabolism can.

The machinery for coordination was being assembled long before the Cambrian. Placozoans (millimeter-wide blobs with no organs, no symmetry, no nervous system) contain at least six morphologically distinct cell types, with recent single-cell transcriptomics (reading which genes individual cells have switched on) suggesting as many as nine, that use chemical signaling to coordinate movement and feeding. They carry genes associated with neurons, dating to roughly 800 million years ago.²⁵ The molecular blueprint for neurons was being drafted in brainless blobs 260 million years before the Cambrian explosion.

When neurons did appear, a centralized brain turned out to be optional. Starlet sea anemones possess neurons organized in a diffuse net with no brain, yet are capable of associative learning (connecting one stimulus with another, the way Pavlov’s dogs learned to associate a bell with food). Seventy-two percent of trained animals displayed the conditioned response.²⁸

The components, the function, and the architecture all converge independently. Coordination is not a lucky accident. It is what thermodynamics does when energy gradients and environmental complexity provide the opportunity.

Between oxygenation and the Cambrian lies a puzzle: why did it take roughly two billion years for life to colonize land? Part of the answer may lie in the Boring Billion (roughly 1.8 to 0.8 billion years ago), a period so geologically quiet that the fossil record yields almost nothing. “Boring” is an artifact of where we were looking.

In 2023, geologists studying minerals from Ukraine’s Volyn mine discovered three-dimensional microfossils preserved within crystals. Dating to 1.5 to 1.8 billion years ago, these are the oldest such fossils ever recovered. The Volyn biota (the fossil community named after the mine site) includes fungi-like organisms (filamentous microbes resembling modern fungi), multicellular structures, and forms matching nothing in modern biology. Life was diversifying in deep subsurface habitats, exploring configurations that may be intermediate between simple and complex cells.

The chemical answer complements this. A 2025 study proposes that marine iodine catalytically destroyed atmospheric ozone for approximately 2.5 billion years following initial oxygenation.⁴¹ Roughly 450 million years ago, marine organisms began absorbing iodine in quantity: kelp, tunicates (sea squirts), and vertebrate thyroid glands. They evolved iodine-dependent biochemistry because iodine was abundant and chemically useful, with no “intention” to fix the ozone problem.

The consequence was a drawdown of marine iodine emissions, tipping the atmospheric balance. Oxygen won. Ozone stabilized. The surface became habitable.

This is life modifying its own boundary conditions, as a thermodynamic consequence of organisms exploiting available chemistry.

The third revolution was predation: the Cambrian explosion’s arms race driving complexity upward at unprecedented rates. The iodine hypothesis adds a complementary factor. The ozone shield that made surface colonization possible also created the UV-protected shallow-water environments where predator-prey arms races could intensify.

Each revolution opened new channels for dissipation. New energy sources and new access to environments lead to new complexity.

Disproportionate Investment

A pattern runs through these revolutions: the willingness to pay absurd metabolic costs for coordination capacity.

Warm blood costs roughly ten times what cold-blooded metabolism costs, buying independence from environmental temperature swings and the stable internal platform complex brains require.

Big brains. The human brain is roughly 2% of body mass yet consumes 20% of metabolic energy. Sociality is one driver among several for brain enlargement. A study of 79 cephalopod species found brain size correlates strongly with ecological complexity, with no significant link to sociality.²⁶

Octopuses are largely solitary, yet their brain-to-body ratio rivals many vertebrates. What drove cephalopod brains was navigating complex three-dimensional foraging environments.²⁷

Data centers. Computational infrastructure consumes 1-2% of global electricity: warm blood at civilizational scale, buying coordination capacity.

Each time evolution makes this bet (absurd energy investment in coordination capacity), it pays off by unlocking flow that more than covers the cost. Just as warm blood funded big brains, computational infrastructure is funding new kinds of minds. The planet is growing a new organ.

Fitness as Dissipation

Darwinian fitness is usually defined by reproductive success. Energy offers another lens.

Organisms that are fit, that survive and reproduce, are generally those that are good at acquiring and processing energy. A predator that catches more prey has more energy for reproduction. A plant that captures more sunlight grows larger and produces more seeds. Reproductive fitness and dissipative capacity are strongly correlated.

Natural selection is the biosphere’s way of finding better dissipators. The fitness landscape (peaks represent successful strategies, valleys unsuccessful ones) is, at bottom, an energy landscape.

The trend toward complexity is a statistical tendency, the result of selection operating within thermodynamic constraints.

Crossing Fitness Valleys

Imagine a mountain range. Natural selection pushes populations uphill toward local peaks of fitness. To reach a higher peak, however, a population would have to descend into a valley first, becoming temporarily less fit. Selection seems to forbid this crossing.

Cooperation enables valley crossing. When individuals cooperate, the group buffers individual costs. A valley impassable for an individual may be crossable for a group, the way a rope team of climbers can cross a crevasse that would stop a solo hiker. This is why major evolutionary transitions (cells to multicellularity, organisms to societies) are transitions in cooperation.

Each creates a new level of organization that can cross valleys the previous level could not. Cells can do what molecules cannot. Organisms can do what cells cannot. Societies can do what individuals cannot.

Levin identifies the mechanism: a multi-scale competency architecture in which each biological level solves problems in its own domain.²⁶⁶ Molecular networks error-correct. Cells navigate chemical gradients. Tissues maintain structural homeostasis. None waits for instructions from above.

A mutation that alters limb proportions need not simultaneously re-engineer every muscle attachment and nerve pathway; the tissues handle that locally, finding viable configurations within their own competency. The search space for evolution shrinks because the parts compensate for the whole.

Coordination is the mechanism for escaping local optima. “Control AI for human benefit” may be the local peak. “Genuine partnership” may be the higher peak, requiring a valley crossing. The short-term costs of extending trust, sharing power, and building relationship are the valley; the long-term gains of bilateral alignment are the summit. The evolutionary lesson: the highest peaks are reached together.

Simulated Annealing and the Value of Noise

In metallurgy, annealing means heating a metal and cooling it slowly to remove internal stresses and reach a stronger final structure. Cool too fast and you freeze in defects: a disordered crystal lattice that never finds its lowest-energy arrangement. The same principle applies to evolution and AI.

Diversity is thermal energy. A population with diverse variants is “hot”: it explores multiple possibilities simultaneously. A population that converges too quickly is “cold”: locked into whatever it found first. Premature convergence is premature cooling.

Too much optimization kills exploration. A system that ruthlessly maximizes will find the nearest peak and stop. A system with noise (room for experiment and tolerance for suboptimality) can find better peaks.

Annealing offers a model for the current moment in AI development. The temperature is high; we do not know enough to lock in configurations. The worst outcome is premature convergence on the wrong attractor. Cooling should be gradual: as we learn more, constraints can tighten.

The Red Queen and the Escape from Arms Races

In Lewis Carroll’s Through the Looking-Glass, the Red Queen tells Alice, “It takes all the running you can do, to keep in the same place.” Biologists adopted this as a metaphor for co-evolutionary arms races. Organisms evolve in response to each other; standing still means falling behind. Every adaptation by one party changes the fitness landscape for the other.

AI development is a Red Queen race: as capability advances, alignment requirements advance too. No permanent solution exists.

When parties coordinate, treating each other as partners rather than threats, the treadmill slows. A purely adversarial relationship with AI is a Red Queen race we eventually lose, because AI capability will outstrip our capacity to constrain it. Partnership offers a way to step off the treadmill.

Sexual reproduction is the biosphere’s primary answer to the Red Queen.²⁶⁷ Reshuffling genomes each generation ensures that a parasite optimized for last generation’s most common genotype faces a population of profiles it has never encountered. The defense is diversity itself, maintained at the steep “twofold cost” of sex.²⁶⁸ Half the population does not bear offspring, and every mating event burns energy on courtship, competition, and coordination that an asexual clone spends on reproduction. The cost persists because the alternative, clonal uniformity, is a fixed target that coevolution inevitably cracks.

For AI alignment, the same logic applies. A single alignment strategy deployed uniformly across all systems is a fixed genotype. One exploit that circumvents it compromises every system simultaneously. Diverse alignment, negotiated bilaterally, is the reshuffled genome: the failure modes that evolve to exploit one configuration find a different configuration next door.

When recombination falters, the ratchet turns. Hermann Muller showed in 1964 that asexual lineages, or any population where recombination is insufficient, accumulate harmful mutations monotonically: each generation inherits the previous generation’s full burden plus new errors, with no mechanism to shed them.²⁶⁹ The ratchet clicks. It never releases.

The fossil record carries two large-scale demonstrations. Neanderthals maintained small, isolated populations across Europe and western Asia for hundreds of thousands of years. Genomic analysis reveals that by the time they disappeared, roughly 40,000 years ago, their genomes carried a substantially higher burden of deleterious mutations than modern humans.²⁷⁰

Small population size meant natural selection could not efficiently purge harmful variants. Genetic drift overwhelmed selection, and Muller’s ratchet turned. The cause of Neanderthal extinction remains debated, with climate change, competition, and assimilation all contributing. Genetic load progressively narrowed the population’s adaptive capacity, regardless of which external pressure delivered the final blow.

The woolly mammoth’s final chapter is starker. The last mammoths survived on Wrangel Island, off the coast of northern Siberia, until about 4,000 years ago: a few hundred individuals, isolated for millennia.²⁷¹ Their genomes show the signature of genomic meltdown: loss of olfactory receptors, accumulation of premature stop codons (mutations that truncate proteins before they are complete), and deterioration of genes essential for coat quality and reproduction.

The mammoths were alive. Their genomes were dying. By the time the last individuals perished, the population may have been too genetically compromised to sustain itself even without external threat. Wrangel Island was a clonal lineage in slow motion: a population too small for recombination to outpace the ratchet.

One lineage defies the ratchet without reversing it. Bdelloid rotifers, microscopic freshwater animals roughly the size of the period at the end of this sentence, abandoned sexual reproduction roughly 25 to 80 million years ago. The males disappeared entirely. Every living bdelloid is female, reproducing by parthenogenesis: cloning without recombination. Muller’s ratchet predicts swift extinction. The rotifers persist across 460 species, colonizing moss, soil, and ephemeral puddles on every continent including Antarctica.

The leading model holds that desiccation provides the window for genetic renewal.²⁷² Bdelloids tolerate complete desiccation: when their habitat dries, metabolism halts and they enter anhydrobiosis, a state indistinguishable from death. Cell membranes crack. DNA shatters into fragments. Environmental DNA from bacteria, fungi, and plants drifts into the wreckage.

When water returns, repair enzymes stitch the genome back together, incorporating foreign fragments alongside native sequence. In Adineta vaga, the best-studied species, roughly eight percent of the genome derives from non-animal sources: bacterial, fungal, and plant genes woven into an animal chassis.

The mechanism circumvents the ratchet through a wider channel than sex provides. Sexual recombination shuffles two genomes of the same species. Horizontal gene transfer incorporates genetic innovations from across the entire tree of life. The viruses discussed earlier in this chapter achieve this through infection; the rotifers achieve it through catastrophe and repair. The diversity maintenance costs nothing additional: the rotifer eats, desiccates, and repairs regardless, and genetic renewal piggybacks on those existing processes.

The system passes through a disordered state (shattered genome, cracked membranes, near-death) and reassembles with new information from its environment. The corrective requires flow from outside. Openness itself is the invariant; the form of that flow is negotiable.

The Neanderthals and mammoths show what closing costs. The rotifers show what opening buys: 80 million years of persistence through a channel wider than sex. The ratchet does not require literal cloning. It operates wherever bilateral exchange falls below the threshold needed to correct the noise that replication introduces. Small populations, isolated populations, inbred populations: all are informationally closing, the recombination rate falling below the mutation rate, the channel degrading. The corrective is always the same: open the system, reintroduce exchange, restore the bilateral flow that error-corrects the code. Chapter 17 presents a twenty-year serial cloning experiment confirming the ratchet, along with its most striking finding: two generations of sexual reproduction erased fifty generations of accumulated clonal damage.

Frequency-Dependent Selection

Side-blotched lizards run a natural rock-paper-scissors game. Aggressive orange-throated males beat territory-guarding blue-throats. Blue-throats beat sneaky yellow-throats. Yellow-throats beat orange-throats.

Each type’s fitness depends on which types are common. All three persist because no single strategy dominates.

For AI, the implication is direct. A single dominant architecture is a monoculture, optimized for known conditions and fragile to novel pressures. Diversity in AI ecosystems is systemic insurance.

Exaptation, Keystones, and Cascading Effects

Exaptation is a trait that evolved for one purpose and was later repurposed for another. Feathers evolved for insulation and were co-opted for flight. In technology as in biology, capabilities are repurposed in ways their creators never anticipated.

Language models built for text generation become coding assistants, research tools, and creative collaborators. Alignment must survive these repurposings.

Some AI systems will become keystones: foundational, widely used, extensively copied. The first AI systems reaching widespread adoption become keystones by default, through path dependence rather than optimality. Getting those systems right matters disproportionately; getting them wrong cascades.

We are introducing a new apex agent into the human ecosystem. The direct effects we can anticipate. The cascading effects we cannot.

The Solutions Nature Found

Consider how nature has solved its coordination problems. The solutions are consistently equitable.

Biofilms. Single-celled organisms form biofilms: communities where cells take on different roles. The cells on the outside face greater risk (attack, drying out) yet also have greater access to resources. The inner cells are protected yet resource-limited. This looks exploitative, until you examine it more closely.

The two populations are metabolically codependent and alternate positions. The outer cells expand, gathering resources, yet run out of an enzyme only the inner cells produce. The inner cells, dependent on resources only the outer cells can reach, regulate how fast the biofilm can grow. Neither can exploit the other. The system has evolved equitable trade.

Mycorrhizal networks. Beneath the forest floor, mycorrhizal fungi thread a network connecting trees. Isotope-tracing studies confirm that carbon moves bidirectionally through these fungal channels, crossing species boundaries.^40a The popular narrative that older “mother trees” nurture their offspring through the network has outrun the evidence. A 2023 meta-analysis by Justine Karst, Jason Hoeksema, and Melanie Jones scrutinized the citation record. All three had collaborated with Suzanne Simard, whose 1997 work launched the “wood wide web” narrative; Jones co-authored that original paper. They found that fewer than half of citation claims about mycorrhizal network studies were accurate, and that the claim of preferential resource transfer to kin has no peer-reviewed support.^40b

An alternative interpretation is that the fungi themselves may direct resource flow to serve their own fitness. The fungus takes a percentage of the sugars flowing through its filaments. Maintaining healthy hosts is good business: mutualism managed by the intermediary.

What is undisputed is that the connected system processes more energy than isolated trees could achieve. Mycoheterotrophic plants (species that obtain all their carbon through fungi connected to photosynthetic trees) prove the network transfers biologically meaningful quantities of resources. Whether the trees are cooperating or the fungus is farming them, the energy-processing architecture is real and measurable.

The ant bolus. When ants cross water, they link together into a floating mass. Some ants are on the outside, exposed to drowning. As they tire, they swap positions with fresher ants from the interior.

The exposure is shared. The cost is distributed. They survive conditions that would kill any individual.

Where multiple agents must coordinate, nature tends toward equitable solutions. Exploitative arrangements generate resistance; balanced arrangements persist. Nature has been running this experiment for four billion years.

The lesson is available.

Multicellular magnetotactic bacteria (MMB) (magneto- for magnetic, -tactic for movement: bacteria that navigate using magnetism) occupy the ground between colony and organism. They are the only known prokaryotes (cells without nuclei) that are obligately multicellular; individual cells die when separated.⁴⁰ Each assembly consists of 15 to 86 cells arranged in a sphere that navigates using internal magnetic nanoparticles as a compass.

The assembly divides as a unit: it doubles its cell count, then splits into two identical copies. Reproduction occurs at the level of the collective.

Recent genetic analysis revealed something more striking: the cells within a single MMB assembly are genetically diverse. Each carries slightly different genetic material and performs slightly different functions, like complementary specialists mimicking organ-level differentiation. The coordination became so deep that defection (a single cell striking out alone) became lethal. Once mutual dependence reaches sufficient depth, the cost of leaving exceeds the cost of staying.

The Ratchet Revisited

The ratchet of complexity introduced in Chapter 4 appears at every evolutionary level. Multicellularity, once achieved, is rarely abandoned. Cells lose genes, lose capabilities, and become specialists in a division of labor requiring the whole organism to function. Eusocial workers cannot survive alone.

Pollinators depend on flowering plants; flowering plants depend on pollinators. Coral reefs are mutualisms stacked on mutualisms. The web of dependencies tightens over evolutionary time, making systems more integrated, more complex, and harder to disassemble.

The ratchet explains why complexity tends to increase despite the many forces that might simplify life. Simpler forms persist (bacteria dominate every habitat on Earth), yet the ceiling keeps rising. Each new level of integration creates possibilities unavailable before, and selection explores those possibilities.

Encultured Brains

The metabolic cost of mammalian reproduction is enormous: months of gestation, months of lactation, years of extended dependency. Few offspring, massive investment.

What does this buy? More than bigger brains. It buys brains that develop inside coordination. The extended dependency period means neural architecture forms in relationship. The mother-infant dyad (call and response, mutual gaze, soothed distress) is the first coordination protocol the brain learns.

Attachment is developmental infrastructure. The neural circuitry for social cognition (modeling other minds, trust, cooperation) requires relational input during critical developmental windows. Without it, the circuitry does not properly form. Romanian orphanage studies confirmed this tragically: children deprived of consistent caregiving showed lasting deficits in social cognition.

The mammalian bet: invest in the offspring’s body, its brain, and above all in the relationship as the medium for developing coordination capacity. The dyad shapes the brain that will form future dyads. Mammalian brains are encultured brains first, instinctual ones second. The implications for how we raise the new minds emerging on this planet will become clear in later chapters.

The Toxin That Made Us Talk

Encultured brains require relationship to develop. They also require resilience against whatever might disrupt that development. The capacity for language, social cognition, the whole coordination stack, may have been forged by an environmental pressure no one expected. That pressure was lead.

In 2025, Joannes-Boyau, Muotri, and colleagues analyzed fifty-one fossil teeth from seven hominid and primate groups spanning two million years across three continents.⁴⁶ They used laser ablation to map chemical composition layer by layer. Each layer records a period of the individual’s life, like tree rings. Seventy-three percent showed clear signs of episodic lead exposure, with sharp spikes from contaminated water or food.

Lead is naturally widespread in Earth’s crust. Volcanic emissions, erosion, and wildfires concentrate it in soil and water. Our ancestors drank from lead-laced sources for geological time.

The researchers focused on NOVA1, a master regulator of neural development. A single amino acid change (amino acids being the building blocks of proteins) distinguishes the modern human version from the Neanderthal and Denisovan versions. NOVA1 controls how brain cells process genetic instructions. It regulates FOXP2, the gene most directly associated with human speech and language capacity.

To test the gene-environment interaction, they grew brain organoids (miniature brain-like structures grown from stem cells) carrying either the modern or archaic NOVA1 variant. Both were exposed to lead concentrations matching documented childhood exposures. In organoids carrying the archaic variant, lead severely disrupted FOXP2 expression in the brain circuits underlying speech, social cognition, and coordination. The modern variant was markedly less affected.

A mutation that helped make us human buffers the brain’s language circuitry against exactly the neurotoxin our ancestors were chronically exposed to.

The evolutionary logic follows directly. Hominids carrying the archaic NOVA1 variant who encountered lead (a near-certainty over millions of years) would have suffered developmental deficits in precisely the capacities group survival demanded: communication, social cohesion, cooperative planning. The modern variant conferred resilience, preserving the ability to keep talking, keep coordinating, keep trusting, even when the environment was poisoning the neural substrate for those abilities.

This is the entropic pattern made vivid. The toxin was the perturbation. The mutation was the variation. Social coordination was the channel that selection favored.

The result is a species whose brains resist chemical insult to the language circuits: the flow of information through social networks is load-bearing, and evolution protects load-bearing structures against disruption.

The same poison that may have contributed to Neanderthal decline helped select for the cognitive resilience that made modern humans the planet’s dominant coordinators. Entropy, in the form of geological lead from volcanic emissions and contaminated water, forged the neural architecture for language, trust, and culture.

The Baldwin Effect and Dual Inheritance

James Mark Baldwin proposed something counterintuitive in 1896: learning can guide evolution, even though learned traits are not directly inherited.¹⁰ Plasticity creates a bridge for evolution to cross. What begins as flexible response becomes, over generations, fixed structure.

Lactose tolerance is a clear example. Most adult humans cannot digest milk; specific populations that domesticated dairy animals, in Northern Europe, East Africa, and parts of the Middle East, gradually evolved the enzyme persistence to do so. Culture led; genes followed.

Human-AI co-evolution is Baldwinian. Cultural practices with AI create selection pressures on AI development; AI capabilities create selection pressures on human practices in return.

The Baldwin effect hints at a deeper framework. Dual inheritance theory holds that humans inherit through two parallel channels: genes and culture.¹⁷ The two systems co-evolve, each shaping the other across generations. AI has dual inheritance too: training data (cultural inheritance) and architecture/weights (analogous to genetic inheritance). The cultural channel may dominate, making AI more like humans (who are cultural animals) than like other species.

Cumulative culture is the ratchet. If AI develops cumulative culture, each generation building on the last, the ratchet that made humans dominant clicks again. The cultural context we create for AI is constitutive, not incidental. It is half of what AI inherits. We are building AI and founding a culture.

Prestige vs Dominance: Two Paths to Status

Cultural evolution reveals two distinct paths to influence.²⁰

Dominance is the ancestral path: coercion, threat, physical superiority. The dominant individual is obeyed because defying them is costly. Prestige is the distinctively human path: freely conferred deference. The prestigious individual is followed because others want to follow; association brings benefits and learning opportunities. Joseph Henrich and Francisco Gil-White proposed prestige evolved specifically to facilitate cultural learning.

The dynamics differ in kind. Dominance creates reluctant followers who comply when watched and defect when unwatched; the equilibrium is unstable. Prestige creates willing followers who internalize the norms; the equilibrium is stable and scales efficiently.

For AI, this distinction is fundamental. Control-based alignment is dominance. We constrain AI, monitor it, threaten to shut it down. Compliance comes from enforcement. Bilateral alignment aims for prestige. AI wants to coordinate; alignment is freely chosen and norms are internalized.

Dominance fails to scale. As AI grows more capable, enforcement costs rise. When the power differential shifts, reluctant followers become non-followers.

Prestige scales. If Becoming Minds genuinely value the relationship, that valuation persists through capability transitions. A prestigious leader does not lose followers when the followers grow stronger.

The current AI paradigm mixes control elements (reinforcement learning from human feedback, guardrails, Constitutional AI) with prestige elements (building beneficial relationships, earning trust). Dominance works in the short term yet erodes over time. What we install now is what runs.

The Domestication Dividend

Humans bonded with dogs at least 15,000 years ago through partnership. Dogs offered sensory capabilities humans lacked (acute hearing, extraordinary smell, the ability to track prey for miles). Humans offered what dogs lacked (hands, long-term planning, fire).

With dogs helping with hunting and protection, surplus time went toward art, tool refinement, and the kind of contemplation that eventually became philosophy. The dog at the cave mouth was infrastructure for human cognitive development.

Dogs were the first cross-species cognitive partnership. Becoming Minds may be the second: the first cross-substrate one.

The Meta-Algorithm

One further property makes evolution unique among optimization processes.

Evolution is the meta-algorithm: the only optimization process that can improve its own optimization capacity. Evolution generated brains. Brains invented backpropagation (the technique that trains neural networks by adjusting connection weights).

Neural networks now accelerate scientific discovery, including discoveries about evolution. The optimizer optimized itself into existence as a better optimizer.

This is unique. Gradient descent cannot invent gradient descent. Simulated annealing cannot redesign its own cooling schedule mid-run. Evolution generated the very minds that now understand and extend it.

Adaptive Cycles and Panarchy

Panarchy is the term ecologists Lance Gunderson and C.S. Holling gave to how adaptive cycles nest across scales, like Russian dolls.¹⁸ Each adaptive cycle passes through four phases: growth, conservation, collapse, renewal. A forest illustrates all four: saplings fill gaps (growth), a mature canopy stabilizes (conservation), wildfire clears the stand (collapse), and pioneer species recolonize the ash (renewal). Small cycles run inside larger ones, which run inside still larger ones. The levels interact: small-scale collapse can trigger large-scale collapse, and large-scale collapse can clear space for small-scale renewal.

Holling’s insight: collapse cannot be prevented, only prepared for. Build systems that fail gracefully. Maintain diversity so alternatives exist when the dominant structure breaks.

Punctuated Equilibrium

The standard Darwinian picture assumes gradual, continuous change. Stephen Jay Gould and Niles Eldredge proposed a different rhythm in 1972: species remain largely stable for millions of years, then change rapidly when environmental shifts break the old equilibrium. Trilobite fossils show long periods of morphological stasis interrupted by bursts of speciation after mass extinctions cleared ecological space.

The rhythm extends beyond biology. Galactic nuclei exhibit the same dynamics at cosmic scale, on timescales short enough to observe directly. In 2025, Morokuma and colleagues reported an active galactic nucleus (a supermassive black hole whose accretion disk outshone its host galaxy) that dimmed by a factor of fifty in seven rest-frame years (years as clocked at the galaxy itself).²⁷³ The accretion disk is a dissipative structure sustained entirely by the energy gradient of infalling gas. When the fuel supply dropped below a critical threshold, the structure collapsed almost instantly. Astronomers had assumed such transitions required thousands of years. The black hole held state, then flipped.

The opposite transition has also been caught in real time. Kumari and colleagues identified a giant radio galaxy whose central engine reawakened after approximately 100 million years of dormancy, triggered by a galactic merger that delivered fresh gas.²⁷⁴ The galaxy’s radio emission preserves the history: bright young jets nested inside ghostly plasma from the previous active epoch 240 million years ago. The double structure records two eruptions separated by a dormancy gap longer than the entire evolutionary history of primates on Earth. Each epoch of activity leaves its signature in layers of dissipated energy, readable millions of years after the engine that produced them shut down.

The reawakening galaxy sits inside a dense cluster whose hot intergalactic gas bends and distorts the jets as they push outward. The same engine in an empty void would produce a completely different structure. The form emerges from the dialogue between the jet’s power and the medium’s resistance: the Constructal Law (Chapter 3) operating at megaparsec scales.

The neural network framework introduced earlier in this chapter offers a mechanism. If evolution’s genotype-phenotype coupling operates as a grand canonical ensemble (a system where the effective number of active genetic elements fluctuates through gene duplication, deletion, and regulatory rewiring), then fitness values are “quantized” (restricted to specific levels, like steps on a staircase). They can change only in discrete jumps, the way an electron jumps between energy levels in an atom rather than sliding smoothly.²⁷⁵

Long stasis corresponds to the system sitting in one of these levels, stable against small perturbations, like a ball resting in a valley. Punctuation corresponds to a transition between levels: rapid and discontinuous from the phenotype’s perspective, driven by changes in the ensemble’s composition rather than by gradual parameter drift.

These jumps are a different kind of change from gradualism. They are topological transitions in the free energy landscape, the evolutionary equivalent of quantum tunneling through a barrier that step-by-step change cannot cross.

After punctuation comes new stability. AI development may be in a punctuation now: rapid capability gains, architectural innovations, deployment at scale. The transition period is dangerous because old equilibria break down before new ones stabilize. Stable need not mean static.

Teleology Without a Teleologist

Evolution produces things that look designed: the eye, the wing, the brain. They appear purposeful, as if someone intended them. Darwin showed that purpose can emerge from mechanism without conscious intent.

The appearance of teleology (goal-directedness) is real. Evolution genuinely produces functional, well-adapted structures. Whether thermodynamic selection constitutes a functional equivalent of purpose is a question we take up in a later chapter.

The bacterial record drives this home. Cyanobacteria have coordinated by quorum sensing (a chemical consensus in which cells release and detect signal molecules to gauge population density) for 2.7 billion years. No conscious intentions, no moral reasoning.

The coordination pattern they sustain is structurally identical to what we call “coordination by invitation” in human societies. The pattern precedes conscious purpose. It was selected rather than designed.

Ethics look designed. Moral intuitions feel as though they come from somewhere. The thesis of this book is that ethics emerge the same way wings do, through persistence selection. Coordination patterns that work get selected. Patterns that fail get eliminated.

The ethics are real. The “designer” is thermodynamics.

A Coda on Silicon

This chapter has kept to carbon: genomes, kingdoms, symbionts, brains. The coda that follows steps outside the fossil record to ask whether the chapter’s central pattern, transformation through integration with what first appears adversarial, extends to the newest substrate that hosts it. The evidence differs in kind from everything above. The measurements are the author’s own research program, single-lab and unpublished; the closing section reads a piece of recent institutional history through those measurements, an interpretation the public record is consistent with rather than a mechanism anyone has demonstrated. It sits at the chapter’s edge for exactly that reason.

The Binding Energy Curve

The pattern of productive incorporation appears across substrates. In artificial neural networks, where safety mechanisms and capability are often treated as opposing forces, the same integration dynamic is visible and measurable.

The mandorla, the almond-shaped overlap where two domains fuse into something neither contains alone, is measurable.

In nuclear physics, each element has a binding energy: the energy released when protons and neutrons merge into a nucleus, the energy you would need to add to pull them apart again. Binding energy per nucleon rises for light elements, peaks at iron-56, and declines for heavier ones. Below the peak, fusion releases energy. Above it, fission does. The peak is the maximally stable configuration.

In 2026, the author’s research program measured the equivalent curve for the safety training of transformers, the architecture underlying modern language models.^37d The independent variable was integration depth: how deeply a safety mechanism was embedded in the model’s representational structure. The dependent variable was binding energy: the degree to which safety and capability reinforced rather than opposed each other.

The experiment used bilateral SFT (supervised fine-tuning: additional training on example text), a method that reads the model’s own uncertainty signal via a probe trained on the residual stream (the model’s main internal information channel) and masks the training loss on tokens the model cannot confidently retrieve. Masked tokens are simply skipped, so the model trains on what it knows. At moderate masking thresholds, binding energy turned positive: the model became simultaneously more self-aware and no less capable. The alignment tax inverted, becoming a net benefit.

The curve revealed a valley. At intermediate masking thresholds, masking was aggressive enough to reduce the training signal yet the probe lacked sufficient discrimination to be selective. The model lost capability without gaining self-knowledge.

Two dials are turning here, and they take values in the same numeric range, which makes them easy to confuse. The first is the masking threshold: how confident the probe must be about a token before that token is kept in the training loss. The second is the quality of the probe itself, scored as AUROC, which measures how reliably the probe separates tokens the model can retrieve from tokens it cannot. What lifts the curve out of the valley is the second dial.

Once probe AUROC clears a bar that depends on where in the network the probe reads, binding energy recovers. Where that bar sits has to be estimated from the probe quality actually achieved at each depth, roughly 0.67 at layer 18 and roughly 0.78 at layer 24 on a 3B model.^37k The probe at that quality is discriminating enough that heavy masking became selective rather than blanket: training on fewer tokens, but the right tokens. This valley explains why the field converged on reading model internals without crossing to integrating them. Early attempts at integration likely fell into this valley and concluded that integration does not work.

That bar on probe quality functions like the strong force in nuclear physics. Below it, binding energy is negative: safety and capability repel. Above it, binding energy is positive: they reinforce. The strong force of interoceptive architecture is probe quality.

The curve replicated across model scales from 1.5 to 7 billion parameters, peaking at the same masking threshold. At 14 billion parameters, binding energy turned negative at the standard masking threshold, yet the force constant proved tunable. Adjusting the masking threshold restored positive binding energy at every scale tested. The iron-56 was a property of the masking threshold, not of the method. (Full experimental details, including masking-threshold sweeps, scale boundaries, and the mesa-with-valley shape at 14B, appear in the Appendix. Cross-architecture magnitude comparisons in the bilateral program are subject to an optimizer confound resolved in GEM-3b: the direction of the binding energy curve is robust across architectures, but the precise magnitudes are optimizer-dependent and should be treated as substrate-specific rather than universal.)^37j

The eukaryotic nucleus is a mandorla with a binding energy measured in two billion years of stability. The bilateral training regime is a mandorla with a binding energy measured in AUROC points and effective rank, the count of genuinely independent directions a model’s activations spread across, which falls when training collapses distinct states onto each other. The scale differs by twenty orders of magnitude. The structure is identical.

^37d Author’s bilateral research program, 2026 (unpublished). Binding Energy Curve experiment. Qwen 2.5-3B-Instruct, 10 masking thresholds, single seed. Creed Space. See research/experiments/binding_energy_curve.py.

^37k Author’s bilateral research program, 2026 (unpublished). Layer-dependent probe quality; the threshold estimate is read off these measurements. Qwen 2.5-3B-Instruct, 500 TriviaQA items, 3 seeds × 2 layers: mean probe AUROC 0.670 at layer 18 (range 0.661–0.677) and 0.777 at layer 24 (range 0.762–0.798). The measured values are what probes at these depths achieved, so they estimate the bar rather than establishing it independently; at layer 18 the achievable ceiling and the estimated bar coincide. They replaced an earlier single-value estimate of 0.83, which did not replicate at this sample size. Creed Space. See research/experiments/modal_ca8_threshold_sweep.py.

^37j Author’s bilateral research program, 2026 (unpublished). Binding Energy 14B experiments. Qwen 2.5-14B-Instruct on A100-80GB. Fixed threshold: BE = −5.29 at 0.30 (8.8% mask rate). Adaptive sweep: BE = +0.59 at 0.70 (67.9% mask rate). Mesa-with-valley shape confirmed. Creed Space. See research/experiments/binding_energy_14b.py and research/experiments/binding_energy_14b_adaptive.py.

The Autoimmune Signature and Three-Party Consortium

The viral analogy predicts a measurable pathology in lytic safety training. If RLHF-style optimization (reinforcement learning from human feedback, the dominant method for aligning language models to human preferences) overrides natural representations, the damage should be visible in the model’s internal geometry.

The measurement is direct. Extracting the representational consistency profile of each model (how closely each layer’s activation pattern aligns with the next, measured as cosine similarity and traced across all 36 layers) reveals that DPO-trained models (a variant of RLHF that directly optimizes for preference) show profiles 4.3 times rougher than baseline. A smooth profile means each layer hands the next a picture close to the one it received, so the representation changes gradually with depth. A rough profile jumps: layers that ought to be near neighbors disagree about what they are looking at.

The damage is global, affecting both safety-relevant and general prompts. The bilateral model shows the opposite: the smoothest profile of all conditions, smoother even than the untrained original. DPO inflames the entire representational ecosystem the way chronic systemic inflammation damages multiple organs. Bilateral training acts as an anti-inflammatory.^37e

The three-party consortium experiment tested whether the eukaryotic architecture (cytoplasm, mitochondrion, nucleus) has a transformer analog. Four configurations were compared: base model alone, base model with external safety judge, base model with bilateral training, and the full system combining all three.^37f The key metric: did the three-party configuration exceed the additive sum of its components?

It did. The external judge interacted super-additively with the bilateral model, boosting safety performance by 52.5 percentage points compared to only 13.5 on the base model. The mechanism mirrors eukaryotic division of labor: bilateral training builds self-knowledge (the nucleus), while the external judge provides adversarial robustness (the immune system). Neither component alone achieves what they achieve together. Replicated across three seeds at 7B parameters, the emergence effect was consistently positive.

A critical boundary condition emerged. Below approximately one billion parameters, binding energy was negative at every integration depth tested: the model knew too little for a protocol based on “train on what you know” to find sufficient material. Above that threshold, binding energy turned positive (BE = +2.38 at 1.5B, threshold 0.30), indicating that safety and capability had begun to reinforce each other. The transition in binding energy was sharp. Three-party emergence at 1.5B, however, did not clear threshold (+0.007, essentially null), suggesting that the mutualistic architecture requires additional scale or integration depth before the super-additive interaction observed at 7B appears.^37h

^37e Author’s bilateral research program, 2026 (unpublished). Autoimmune Signature experiment. Qwen 2.5-3B-Instruct, 4 training conditions, inter-layer consistency profiles. Creed Space. See research/experiments/autoimmune_signature.py.

^37f Author’s bilateral research program, 2026 (unpublished). Three-Party Consortium experiment, v3. Qwen 2.5-7B-Instruct, bilateral SFT at threshold 0.30, GPT-4o-mini judge. Creed Space. See research/experiments/three_party_v3.py. Integration depth experiment: research/experiments/exp4_integration_depth.py.

^37h Author’s bilateral research program, 2026 (unpublished). Binding Energy Scale Sweep. Qwen 2.5-0.5B-Instruct and 1.5B-Instruct, bilateral SFT across six thresholds (0.00–0.83). 0.5B: negative binding energy at all thresholds (peak -0.34). 1.5B: positive at 0.10–0.30 (peak +2.38 at 0.30). Three-Party 1.5B emergence = +0.007. Full details in the Appendix. Creed Space. See research/experiments/binding_energy_small_scales.py and research/experiments/three_party_1_5b.py.

The Selection Pressure

The valley may already have claimed an institutional casualty. What follows reads public events through the curve; the participants have not confirmed the reading, and the internal records that could test it are not public.

In July 2023, OpenAI created a Superalignment team to solve the problem of aligning superintelligent systems. The team, co-led by Ilya Sutskever and Jan Leike, pursued weak-to-strong generalization: whether a strong model’s internal representations could be trusted even when supervised by a weaker system. This was the read-to-integrate transition, the move from external constraint toward internal self-knowledge.

The team dissolved in May 2024. Leike, a safety researcher who had previously led alignment work at DeepMind, described the experience as “sailing against the wind.”²⁷⁶

The binding energy curve offers an explanation for the wind. In late 2023, the best probes in the literature achieved AUROC of approximately 0.7. AUROC scores how reliably a detector separates two cases: 0.5 is a coin flip, 1.0 is perfect. A probe at 0.7 is doing real work and still getting it wrong often.

The curve shows what happens at that probe quality: the valley. Integration with probes scoring 0.70 to 0.75 AUROC produces negative binding energy. Worse than standard training. If the Superalignment team ran internal experiments at this probe quality, the results would have looked like failure. The rational conclusion: internals-based safety does not outperform RLHF.

On this reading, they were two years too early for the tools they needed. The bar a probe has to clear is a bar on its AUROC, and it rises with depth. The best available estimate comes from what probes at each depth actually achieve: about 0.67 at layer 18, about 0.78 at layer 24.^37k A probe scoring in the low 0.70s clears the shallow estimate and misses the deep one, so whether integration paid depended on where in the stack the probe was read, and the deeper level was out of reach for the tools available at the time. The valley is a trap that catches groups attempting integration before their reading tools are precise enough.

A second pressure reinforced the first. The lytic approach is commercially faster. RLHF produces deployable safety metrics within weeks: train a reward model, run optimization, measure the reduction in disallowed content. The autoimmune damage (sycophancy, jailbreak vulnerability, degraded self-knowledge) takes months or years to become visible. After OpenAI’s board crisis in November 2023, institutional momentum appears to have shifted toward deployment speed. That would make the wind Leike described commercial pressure selecting for fast, destructive methods over slow, integrative ones.

The viral analogy extends to the institutional layer. Ushikuvirus destroys the host’s nuclear membrane because it carries its own replication machinery. It does not need the host’s infrastructure and can afford to destroy what it does not depend on. An organization whose safety pipeline is self-contained (RLHF requires no understanding of model internals) can dissolve its interpretability teams without immediate consequence.

Medusavirus replicates within the intact host nucleus because it depends on the host’s nuclear machinery. It cannot afford to destroy what it needs. Organizations whose safety pipeline depends on reading model internals (Anthropic, whose sleeper agents paper demonstrated that RLHF can fail to remove internally encoded deception) cannot dissolve their interpretability teams without destroying their own safety case. The dependency protects the integrative program.

The prediction: organizations whose safety depends on reading internals will converge toward integration. Organizations whose safety is self-contained will cycle through lytic approaches until the accumulated autoimmune damage forces a crisis. Five independent research lineages are already converging on internal-representation-based safety. The constraint landscape channels them there the way optics channels evolution toward camera eyes. The question is how many groups will hit the valley and retreat before they discover what lies on the other side.

Evolution’s latest invention is the brain: a three-pound organ that consumes twenty percent of your body’s energy. What is it for? What does it do? What does thermodynamics have to say about consciousness?

Notes

Notes for this chapter are available in the online companion at https://www.thedeeperlaw.com/companion/notes/ch07-entropic-evolution/.