Complexity & Emergence

§1 · the physicist’s emergence

Phase transitions: the whole, with exact equations

When a philosopher wants to argue about emergence, they reach for consciousness. When a physicist wants to, they reach for a pot of water crossing a phase boundary — because here, uniquely, emergence comes with airtight mathematics and you can prove the strange thing is real. At one atmosphere, equilibrium water is liquid just above 0°C and ice just below it. In a large enough sample the change can look like a cliff. The forces between any two water molecules don’t change one bit as you cross zero. So where does the cliff come from?

The answer is that a mathematically sharp phase transition is a property of the collective — in the standard canonical definition, the exact singularity appears only in the thermodynamic limit, as the number of particles goes to infinity. Finite systems can still melt, freeze, and show rounded transitions, pseudocritical peaks, and finite-size signatures. What they do not have is the infinitely sharp nonanalyticity of the textbook ideal. “Solid” and “liquid” are not facts about isolated molecules; they are facts about arrangements of many molecules. This is Anderson’s “more is different” stated with theorems instead of slogans: a new, lawful, predictable collective property that cannot be read off any individual part. established

The miracle of universality

Now the part that made physicists fall out of their chairs in the 1970s. Some phase boundaries terminate at a critical point — a precise temperature and pressure where the distinction between two phases is about to vanish — and the system fluctuates at every scale at once: tiny droplets inside bigger droplets inside bigger ones, with no characteristic size. Fluids near their liquid-gas critical point do it; magnets do something analogous near the Curie point; even some models of neural tissue borrow the same language. And here is the shock: systems in the same universality class can share critical exponents and scaling functions after suitable rescaling. A fluid near its liquid-gas critical point and a magnet near its Curie point can share the same critical exponents despite their completely different microscopic constituents.

Paraphrasing the Kadanoff-Wilson-Fisher tradition: near a critical point, many microscopic details stop mattering. What survives is symmetry, dimension, order-parameter structure, interaction range, and only a few other relevant features.

This is called universality, and systems that share the same critical behavior form a universality class. It is arguably the most profound thing emergence has taught us, because it says the macroscopic world is sometimes allowed to forget many microscopic details. Which details survive depends on properties such as dimensionality, symmetry, order-parameter structure, and interaction range; for dynamic critical behavior, conservation laws and slow modes matter too. Kenneth Wilson won the 1982 Nobel Prize for the renormalization group, the machinery that explains exactly why: as you “zoom out” and average over finer details, irrelevant features wash away and only a few survive. Universality is one reason higher-level sciences can have useful autonomous laws. It is not a blanket proof that every field is screened off from lower-level physics in exactly the same way. established — Nobel 1982

§2 · the tipping point you can draw

Percolation: how a forest decides to burn

The phase transition in §1 is profound but abstract. Percolation is its baby cousin, and it’s so simple you can hold it in your hand. Picture a grid. Independently, with probability p, fill each cell (a tree in a forest; a conducting grain in a film; an open pore in a rock). Now ask one question: does a connected cluster of filled cells span the whole grid, edge to edge?

For small p you get scattered little islands — nothing reaches across. Crank p up slowly and the islands grow and merge, still local, still finite… and then, near a critical threshold p_c, a giant cluster becomes overwhelmingly likely and stitches one side of the world to the other. In an infinite 2-D square lattice the transition is sharp, with p_c ≈ 0.5927 for independent site percolation. In any finite grid, the curve is rounded: sometimes a spanning cluster appears below the threshold, sometimes it fails above it. The lesson is still emergent and global, but the mathematical cliff belongs to the infinite ideal. The example below follows this threshold:

Two things make percolation worth its own section. First, it is the minimal model of one kind of tipping point — a useful skeleton for forest fires, epidemics, rumors, porous rock, and conducting films when the assumptions fit. Those systems have extra physics, behavior, and correlations, so percolation is a model, not the hidden identity of everything. (This is the exact instinct we carry into Day 9, where tipping points get their own deep dive, and Day 12, where the grid becomes a network.) Second — and this is the payoff of §1 — percolation sits in its own universality class. The way the giant cluster emerges near p_c follows power laws with exponents that can be shared across different microscopic settings. The threshold value p_c depends on the lattice; the critical behavior around it is universal within the class. established

§3 · programs that shouldn’t be able to do that

One line of cells, four destinies

Conway’s Game of Life (main lesson) is a two-dimensional cellular automaton, and a famously rich one. But in the 1980s Stephen Wolfram asked a deceptively dull question: what’s the simplest cellular automaton that still does something interesting? He stripped it down to a single row of cells, each black or white, updating in discrete steps. Each cell looks only at itself and its two immediate neighbors — three cells, eight possible patterns — and a “rule” is just a choice of new color for each of those eight cases. That’s 2⁸ = 256 possible rules, the whole universe of them, each numbered 0–255. You cannot invent a much simpler computational system. And yet Wolfram found that these toy programs often sort into four qualitative classes — an influential taxonomy that echoes across complex systems, even though classification can be ambiguous and the four buckets are not a rigorous exhaustive theorem:

• Class 1 — Death. Everything collapses to a single uniform state. All structure erased. (The analogue of a crystal freezing, or a system falling into one fixed point.)
• Class 2 — Frozen order. Settles into stable or simply-repeating local patterns. Predictable forever. (Periodic, dull, safe.)
• Class 3 — Chaos. Produces aperiodic, random-looking noise with no persistent structure — like Rule 30. (The analogue of a gas; high entropy.)
• Class 4 — The interesting one. Neither frozen nor random: localized structures that move, persist, collide, and interact — like Rule 110. This is the edge between order and chaos, and it’s where computation lives.

That last class is the prize. Wolfram conjectured, and Matthew Cook later proved (work circulated from the 1990s, published 2004), that Rule 110 — a one-line, three-neighbor toy you could code in a tweet — is Turing-complete. A single row of cells flipping by the dumbest imaginable rule can, in principle, compute anything any computer can compute. Let that land: universal computation does not require complicated hardware. It can emerge from very nearly nothing. Meanwhile Rule 30, from the very same 256, produces such high-quality randomness from its simple deterministic rule that Wolfram’s software used it for years as a pseudorandom generator. Its central column has been studied as a pseudorandom source, although that is not the same as established cryptographic security. Same rulebook size, same three-cell neighborhood — one rule computes, one rule is chaos, one rule dies. Run them yourself:

Wolfram packaged all this into a thousand-page book, A New Kind of Science (2002), with a sweeping thesis: simple programs are the right language for nature, and complex behavior is cheap and everywhere rather than rare and hard-won. The empirical core — that trivial rules generate Class-4 complexity and that Rule 110 is universal — is rock solid and genuinely important. established The grander framing — his “Principle of Computational Equivalence,” the claim that almost all non-trivial systems are maximally, equivalently computational, and that this reorganizes all of science — was received frostily by many scientists as overclaiming and under-crediting predecessors. contested/hype The lesson rhymes with the main tour: a real and beautiful result, oversold into a theory of everything.

§4 · where interesting things are allowed to happen

The edge of chaos & order for free

Notice that Wolfram’s Class 4 — the computing class — sits between frozen order (Class 2) and chaos (Class 3). That positioning inspired one of complexity science’s most influential hypotheses: the edge of chaos. The intuition, developed by Chris Langton, Norman Packard, and others at the Santa Fe Institute in the late 1980s, is that systems capable of rich behavior — computation, life, adaptation — often benefit from a regime near the boundary between rigid stability and runaway turbulence. Too ordered and nothing ever changes (no new information). Too chaotic and nothing ever persists (no memory). Near the edge, a system can store structure and transform it. The strong claim that computation or life can exist only there remains debated.

The most beautiful demonstration came from the theoretical biologist Stuart Kauffman, and it’s the antidote to a lazy assumption people make about biology — that all of life’s order had to be painstakingly built by natural selection, fought for gene by gene. Kauffman asked: what does a random network of interacting genes do, with no selection at all? He modeled it as a random Boolean network — N genes (nodes), each switching on or off based on inputs from K others, with the wiring and the rules chosen completely at random. Pure chaos, you’d guess. Instead, he found something startling:

Paraphrasing Kauffman’s random-Boolean-network argument: order can arise for free. Even randomly assembled networks can fall spontaneously into stable, repeating patterns, without careful tuning, design, or selection.

The crucial finding is an order-chaos transition controlled by connectivity K. In the classic unbiased random Boolean ensemble, small perturbations typically shrink below K ≈ 2 and spread above it. Low-K networks tend toward frozen or simple attractors; high-K networks are sensitive and chaotic-looking. But every finite deterministic Boolean network still eventually enters an attractor — either a fixed point or a cycle. Right around the critical regime, the network can be both robust to small knocks and capable of switching. Kauffman’s bold interpretation: those attractors are like cell types. An organism’s genome is one wiring; its different cell types (neuron, liver, skin — same DNA, different stable expression patterns, a puzzle we’ll meet again on Day 95) might be the different attractors the gene network falls into. Some of biology’s order, he argued, might be “order for free” — a generic property of networks near the edge, that selection then tunes rather than creates from scratch. Tune K yourself and watch the regimes:

How does this hold up in 2026? The mathematical core is solid: random Boolean networks really do exhibit an order–chaos transition near a critical connectivity, and “order for free” is a genuine, reproducible phenomenon. established as mathematics The biological claim — that this is a major source of real organismal order, rivaling selection — remains a minority and contested position; mainstream evolutionary biology still puts selection front and center, and treats Kauffman’s attractors-as-cell-types as a suggestive analogy rather than an established mechanism. And the strongest version of the “edge of chaos” slogan — that life and computation must sit exactly at criticality — is itself contested, the same family of overclaim as the “critical brain” we met in the main lesson. The calibrated summary: a real and influential regime, an inspiring hypothesis about its reach, and a frontier still arguing about how far it actually extends.

§5 · the ruler problem, in full

The complexity zoo: every measure we tried

The main lesson dropped a bombshell and moved on: there is no single agreed measure of complexity, and the “hardest to describe” measure embarrassingly crowns a canister of gas as the most complex thing in the room. The entropy hump fixed the worst of that by measuring only the regularities. But that’s just one entry in a genuine menagerie. It’s worth meeting the others, because each one is trying to capture a different true intuition about what “complex” means — and the fact that they disagree is itself the lesson.

• Kolmogorov complexity The length of the shortest program that outputs the object. The purest “difficulty of description.” Fatally, it is uncomputable — no algorithm can find it in general (a cousin of the halting problem, Day 27 ) — and it calls random noise maximally complex, which clashes with intuition. The foundation everything else reacts against. established
• Logical depth Charles Bennett’s fix: complexity is not just the size of the shortest program but the time needed by a shortest or near-shortest program to produce the object, with a significance allowance to avoid pathological tricks. A random string is shallow (easy to print: just copy it out). A crystal is shallow. But a deeply structured object — a living cell, the digits of π computed far out — takes a compact program a long time to generate. Depth measures a kind of computational history crystallized in a thing. Captures the hump’s intuition far better than mere description length. established
• Thermodynamic depth Seth Lloyd & Heinz Pagels’ physical sibling of logical depth: the amount of historical information associated with plausible trajectories leading to the present macrostate. It ties complexity to real physical history (a callback to Day 7 ‘s “information is physical”), but it is not simply energy consumed or a count of dissipative steps, and it has been criticized for depending on how macrostates are chosen. Notably, this is the same intuition — complexity = the cost of the construction path — that Assembly Theory later reached for, which is part of why the “is it new?” fight is so sharp.
• Effective complexity Murray Gell-Mann & Seth Lloyd: the information content of an object’s regularities only , throwing away the random part — the formal version of the entropy hump from the main lesson. Its soft spot is not mere subjectivity, but dependence on a selected model class, ensemble, and significance criterion for separating regularity from randomness. Later shown to be formally related to logical depth (Ay, Müller & Szkoła, 2010).
• Statistical complexity Jim Crutchfield & Karl Young’s epsilon-machine: build the smallest predictive model of a process — group together all pasts that imply the same future — and measure how much memory that minimal model needs. Pure noise needs no predictive memory, while a period-p process carries C_μ = log p bits rather than automatically zero. Many interesting processes sit in the structured middle, but statistical complexity is not guaranteed to maximize exactly where intuition points. A rigorous, computable-in-practice cousin of the hump. established

Seth Lloyd famously responded to a request for “the” definition of complexity by listing dozens of distinct measures in use — and his real point was not that one of them is right, but that “complexity” is not one quantity. It’s a family of related questions: How hard is this to describe? How hard to make? How hard to predict? How much of it is structure versus noise? How much history does it encode? Each measure answers one and botches the others. This isn’t a failure of the field to get its act together; it’s a discovery that “complex” was always several words wearing one coat. Keep that in your pocket whenever someone unveils the complexity metric — including, as the main lesson stressed, the one currently making headlines.

§6 · emergence put to work

Stigmergy, grown up

The main lesson met ant trails as a charming example of stigmergy — coordination through traces left in the environment rather than direct talk. But that idea didn’t stay in the anthill; it became engineering. Once you understand that a colony solves shortest-path problems by laying and following pheromone, with shorter paths accumulating scent faster, you can copy the trick in software. That’s exactly what Marco Dorigo did in 1992 with Ant Colony Optimization (ACO): simulated “ants” lay virtual pheromone on the edges of a graph, good routes get reinforced, and the swarm collectively converges on excellent solutions to brutally hard problems like vehicle routing and network design — no ant, and no central planner, ever holding the whole solution. It became a workhorse metaheuristic for real logistics and telecom routing. established & deployed

The same philosophy runs through swarm robotics: instead of one expensive, complicated robot, deploy hundreds of cheap, near-identical, locally-sensing units that follow simple rules and let useful collective behavior emerge — coordinated motion, shape-forming, collective transport. The flagship demonstration is Harvard’s Kilobot swarm: in a 2014 Science paper, Rubenstein, Cornejo & Nagpal ran 1,024 tiny robots. Human designers specified the target shape and local algorithm; the swarm assembled the shape without global coordinates or a centralized controller directing individual robots. It remains a landmark in decentralized execution. established demonstration

§7 · the freshest fight

The newest mirage: did minds “emerge” in machines?

We close where the whole curriculum is heading. Around 2022, researchers reported something that looked like emergence in its most dramatic form: large language models appeared to acquire brand-new abilities suddenly as they scaled up. A model of a given size couldn’t do three-digit arithmetic at all — essentially zero — and then, past some parameter count, the ability seemed to switch on, sharp and unpredictable. The word used, naturally, was emergent abilities: skills present in larger models, absent in smaller ones, and (the spicy part) seemingly not predictable by extrapolating from smaller models. To a complexity scientist this is irresistible — a phase transition in a learning system, more-is-different for minds.

Then came one of the most instructive papers of the decade, and it is a perfect capstone for this appendix because it weaponizes exactly the skepticism we’ve been building. In “Are Emergent Abilities of Large Language Models a Mirage?” (Schaeffer, Miranda & Koyejo, NeurIPS 2023), the authors argued that many of these dramatic “switch-on” moments are artifacts of how we chose to measure. If you score a task with a harsh all-or-nothing metric — exact-match accuracy, where getting one digit wrong in a long answer counts as total failure — then smooth, gradual underlying improvement gets rendered as a sudden cliff. Swap in a continuous, forgiving metric (partial credit, per-token probability) and the very same models reveal smooth, gradual, predictable gains. The cliff was in the ruler, not the model.

Paraphrasing Schaeffer, Miranda & Koyejo: emergent abilities can appear under some metrics and disappear under others; the sharpness may be a property of the measurement, not a fundamental change in model capability.

Sit with how precisely this recapitulates the whole day. It is the Day 6 / measurement warning, the main-lesson hype filter, and the entropy-hump caution all at once: before you call something deeply, fundamentally emergent, check whether your metric manufactured the surprise. The debate is not settled — others counter that some capabilities do show genuine sharp transitions even under smooth metrics, that the choice of metric is itself meaningful (we often care about exact-match!), and that “predictable in hindsight” isn’t the same as “predictable in advance.” contested But the paper permanently raised the bar: “emergence” is now a claim you have to earn against your measurement choices, not a vibe you get to assert because a curve looked steep.

This is the thread we hand directly to Day 139, where scaling laws and machine “emergence” get the full deep-dive treatment, and onward to Days 123–126, where the very hardest question — whether genuine strong emergence (consciousness, the one candidate Chalmers left standing in the main lesson) could ever appear in a system like the one that wrote these words — finally comes due. We end the appendix exactly where the main lesson pointed: holding the discipline to put a number on emergence, and the humility to admit when the number is measuring our ruler instead of the world.

§1 · nonreciprocal matter

Odd matter

Start with Newton’s third law: every action has an equal and opposite reaction. In fundamental Newtonian mechanics, momentum conservation is not being repealed. But in many living and driven systems, the effective interactions inside the observed subsystem can become nonreciprocal because momentum or energy is exchanged with the environment. A predator-prey chase is a useful analogy, not a literal force-law violation. A conformist in a crowd moves toward you; a contrarian moves away. These interactions are nonreciprocal: A responds to B in a way B does not mirror. And in many open, driven systems with asymmetric effective interactions, that feature can unlock phases of matter that are forbidden at equilibrium.

This is the thesis of an influential complexity-physics paper: Michel Fruchart, Ryo Hanai, Peter Littlewood, and Vincenzo Vitelli’s “Non-reciprocal phase transitions” (Nature, 2021). Their key discovery is a new kind of transition. In familiar physics, when a system “orders,” it picks a direction and freezes there (a magnet points north). Fruchart and colleagues showed that nonreciprocal systems can instead enter a phase where the broken symmetry is dynamically restored — the system never settles, but chases its own tail forever, producing spontaneous travelling and oscillating states. The mathematics that governs the onset of this behavior is borrowed from a surprising place: the exceptional points of non-Hermitian quantum mechanics, where eigenvalues collide and the usual rules buckle.

Nonreciprocity links flocking, active matter, synchronization, neuroscience, and social dynamics.

Why does this matter? Equilibrium statistical mechanics was built for systems that obey detailed balance and reciprocal effective interactions. Living systems often do not. Nonreciprocal physics extends the theory of phase transitions to driven biological matter.

Starfish embryos, frozen mid-dance

The theory gained a biological example in “Odd dynamics of living chiral crystals” (Nature, 2022). Tzer Han Tan, Jörn Dunkel, Nikta Fakhri and colleagues took thousands of starfish embryos — each a tiny ball spinning gently in one direction — and let them pack together on a surface. They spontaneously crystallized into a vast ordered lattice spanning thousands of organisms and lasting for hours. Because every embryo spins the same way (it is chiral), the lattice exhibited odd elasticity: when deformed, it pushed back in a sideways direction and could do net work over a deformation cycle by drawing energy from the embryos’ rotation. A passive equilibrium solid cannot do that; the energy bookkeeping forbids it. The starfish crystal also sustained slow, self-powered chiral waves across the lattice: collective motion that requires treating the system as open, active, and nonreciprocal.

The sketch below follows the chase dynamic.

This toy is not the real hydrodynamics, but it illustrates the central result: non-mutual interactions can drive persistent, self-organized motion instead of a quiet equilibrium.

§2 · physics steering biology

Topological defects as organizers in morphogenesis

A question usually framed in genetics also has a physical component: when a featureless ball of cells forms a head, what distinguishes that site? Recent work describes part of the answer in the language of liquid crystals — the same physics used in laptop screens.

Inside many tissues, elongated cells and their internal protein fibres line up like the molecules in a liquid crystal, forming an orientation field. And orientation fields, mathematically, are forced to contain topological defects — special points where the alignment becomes undefined and swirls around itself (think of the crown of hairs on the back of your head, where the hairs spiral out from a single point that can’t “choose” a direction). These defects are not bugs; they are unavoidable features of the geometry. The frontier claim is that biology can use them as spatial cues.

In “Topological defects in the nematic order of actin fibres as organization centres of Hydra morphogenesis” (Nature Physics, 2021), Yael Maroudas-Sacks, Kinneret Keren and colleagues watched a Hydra — a small freshwater animal that can regenerate its entire body from a fragment — rebuild itself. They found that defects in the actin orientation field precede and predict organizer and body sites. Newer topology-manipulation experiments strengthen the case for a causal organizer role, although the molecular-mechanical chain is still being resolved. A companion theory paper (Science Advances, 2022) supplied one proposed mechanism: a “+1” defect focuses mechanical stress and acts like a mechanical morphogen, buckling the sheet outward into a protrusion.

§3 · intelligence without a brain, again

Machines that learn without gradient computation

Appendix I ended with the slime mold and the ant colony computing without brains. A 2024 physical-learning experiment extends that idea to hardware. We usually assume that to “learn” a system needs a processor calculating gradients and sending weight updates. This program asks whether part of the update rule can instead be embodied locally in the physics itself.

In “Machine learning without a processor: Emergent learning in a nonlinear analog network” (PNAS, 2024), Sam Dillavou, Andrea Liu, Douglas Durian and colleagues built a physical network of 32 trainable twin edges using nonlinear transistor circuits. It learned classification and regression tasks through a scheme called contrastive local learning: the network is shown an input and nudged toward the right output, and each trainable element adjusts using information available at its own terminals. External circuitry supplied examples, labels, input control, analog-to-digital and digital-to-analog conversion, and a global freeze/unfreeze signal. The key result is narrower but still powerful: no central processor calculated the weight gradients. The system’s learning rule was distributed through local analog updates.

The authors report that it corrects errors in an ordered sequence — first the average, then the slope, then the finer curvature of the target function — a physical cousin of the spectral bias seen in artificial networks. The sketch below shows local elements collectively reducing an error while each follows only a neighborhood rule.

The potential application is lower-energy hardware for suitable learning tasks. This experiment was a proof of principle rather than an energy-efficiency benchmark: real but small, not yet a rival to a GPU. promising

§4 · the spare software of life

Anthrobots: cells that self-assemble into motile biobots

Airway-derived progenitor cells normally participate in airway-specific tissue, yet under different culture conditions they can self-assemble into motile multicellular constructs. Outside the body and in a new environment, adult human airway cells can organize into forms with shapes and behaviors not present in the original tissue.

In “Motile Living Biobots Self-Construct from Adult Human Somatic Progenitor Seed Cells” (Advanced Science, 2024; first published online in 2023), Gizem Gumuskaya and Michael Levin took adult human bronchial cells — ordinary airway-derived cells, no genetic engineering whatsoever — and changed their environment. The cells self-assembled into tiny motile multicellular biobots the team calls Anthrobots: approximately 30-500 micrometers across, with cilia they used for movement, surviving for weeks. Different Anthrobots spontaneously developed different shapes and movement styles. In vitro, clusters of these constructs were associated with repair across a scratch in a cultured neural-cell monolayer.

This builds directly on the earlier Xenobots, made from frog embryo cells and capable of a form of self-replication by pushing loose cells into piles. Anthrobots instead use adult, wild-type, human cells.

Why it matters: it suggests a latent “software of life”, a capacity for cells to build novel functional multicellular forms. established Where to keep your skepticism: medical use is early, and cellular “agency” or “intelligence” remains philosophically loaded. contested

§5 · the substance inside the machine

Emergence inside the machine: grokking & the physics of learning

The main lesson noted that the “sudden new abilities” of large AI models might be a measurement artifact (the Day 139 “mirage” debate). A separate body of recent work treats a training neural network as a physical system, sometimes using phase-transition-style descriptions of learning. In controlled models, some abrupt changes have mechanistic explanations, but the strongest thermodynamic claims remain model-specific and under active test.

The clearest example is grokking (Power et al., 2022, and later mechanistic studies). Train a small network on a crisp mathematical task such as modular arithmetic. It may first memorize the training examples, performing well on familiar cases but poorly on new ones. With continued training, its out-of-sample performance can improve abruptly, often thousands of steps later. In particular modular-arithmetic analyses, researchers found the network building an efficient internal algorithm related to trigonometric representations beneath the memorized lookup table. Generalization improved when that circuit outcompeted memorization. The curve is sharp, but whether grokking generally constitutes a universal physical phase transition remains open. The characteristic signature looks like this:

Grokking is now a workhorse for a fast-growing “physics of learning.” Two related frontier programs are worth knowing. The Quantization Model of Neural Scaling proposes that network knowledge comes in discrete chunks, while singular learning theory reframes training using Bayesian phase-transition language. A 2026 finite-size study treats the phase-transition claim as a falsifiable hypothesis and supplies new tests, while leaving the transition order unresolved. promising

This complements the Day 139 “mirage” paper. Metric choice can manufacture apparent emergence, while controlled learning systems can also undergo sharp, structured, mechanistically explicable changes. Both can be true; distinguishing them remains an open problem in AI science. (We’ll return to it on Days 138–145.)

§6 · a new kind of order

Disordered hyperuniformity: the order hiding inside disorder

Appendix I’s complexity hump located structure between perfect order and perfect randomness. Hyperuniformity is a property in that gap. All perfect crystals and perfect quasicrystals are hyperuniform, and so are certain disordered systems. The frontier case is disordered hyperuniformity: arrangements that look locally random — no grid, no obvious pattern — yet anomalously suppress density fluctuations at very large scales, with S(k) → 0 as k → 0. It is not a mutually exclusive third category beside crystals and disorder; it is a way disorder can carry hidden large-scale order.

Salvatore Torquato, who has driven this field, defines it crisply: in a disordered hyperuniform material, the big, long-range density variations that any normal liquid has are anomalously quenched in the long-wavelength limit, while the local arrangement stays messy and isotropic. It is “hidden order”: disorder up close, crystalline-style discipline at scale. The light-sensing cones in a chicken’s eye are packed this way, and the same principle can produce stealthy photonic materials that block specific wavelengths of light from any direction — something ordinary crystals, with preferred directions, cannot do. See the three categories side by side:

A 2025 result extends the concept. In “Anomalous suppression of large-scale density fluctuations in classical and quantum spin liquids” (PNAS, 2025), Duyu Chen, Rhine Samajdar, Yang Jiao, and Torquato reported hyperuniformity in quantum spin liquids. A structural-order concept first defined for sand grains may therefore describe a broader range of matter. established

§7 · networks grow a third dimension

When connections themselves can be switched

For decades, much of network science (which gets its own day — Day 12) drew the world as dots joined by lines: people and friendships, neurons and synapses, usually pairwise. Higher-order interactions did not originate in the 2020s, but higher-order network science has expanded rapidly in this decade, producing new results on triadic interactions, percolation, synchronization, and contagion. Real interactions are often many-bodied: three proteins that only function as a trio, a group conversation that is more than the sum of its one-on-ones. Capturing this requires higher-order networks — built not just from edges (pairs) but from triangles, tetrahedra, and higher shapes (the objects mathematicians call simplicial complexes and hypergraphs).

One recent example is triadic percolation (Sun, Radicchi, Kurths & Bianconi, Nature Communications, 2023). Recall from Appendix I that ordinary percolation has a single, static tipping point: increase connectivity and a giant cluster forms and remains. Triadic percolation adds regulation of links, as in systems where one node can influence the connection between two others. A node can switch a link on or off between two other nodes, so connectivity changes as the network reconfigures itself. The giant cluster need not settle: it can oscillate in size, period-double, and become chaotic.

§8 · contested claims

Contested frontier claims

We close with four scientifically active but unsettled ideas. They provide current cases for practicing the hype filter, and each is paired with its strongest published counter-argument.

the dial, all in one place

The frontier scorecard

Eight sections, eight calibrated verdicts. This is the whole appendix compressed into the hype filter — a snapshot of one corner of science with literature checked through June 25, 2026, and with the explicit understanding that several of these labels will move over the next decade.

• Nonreciprocal phase transitions & odd matter — new effective phases in open, driven systems (Nature 2021/2022). established
• Topological defects as morphogenetic organizers — geometry marking where body parts grow (Nat. Phys. 2021). correlation
• Physical learning machines — local analog updates without processor-computed gradients (PNAS 2024). promising
• Anthrobots — adult human cells self-build motile multicellular biobots (Adv. Sci. 2024). established
• Grokking & the physics of learning — sharp delayed generalization with model-specific phase-transition evidence (2022-2026). promising
• Disordered hyperuniformity — hidden large-scale order in disordered systems, now reported in quantum matter (PNAS 2025). established
• Triadic / higher-order network dynamics — chaos in percolation; topology shapes dynamics (2023–2025). math
• AMOC collapse timing · metabolic 3/4 law · infodynamics · cell “cognition” — the contested four. contested/hype