More parts isn't complexity — entanglement is
2026-07-21 · What Counts as Complex
A Boeing has six million parts and it is not complex. A flock of starlings runs on three rules and it is. We get these two words backwards almost every time — and the price of getting them backwards is that you keep taking apart the one kind of thing that should never be taken apart.
An aircraft maintenance manual is as thick as a brick, but read it and every part is in there: what it is, what it connects to, how to swap it when it fails. The millions of connections between parts are all connections written down in the manual. However vast the machine, it is the kind of thing you can trace, one line at a time.
Why a city clogs into gridlock out of nowhere on an unremarkable Wednesday afternoon is in no manual anywhere. Nobody designed that jam; every driver in it just wants to get home; the rules could not be simpler — and yet you can know every car and every driver's intent and still not predict where the next jam forms.
This is the dividing line the whole site turns on: "many parts, thick manual" is one thing; "simple pieces entangling into behaviour you cannot compute" is another. The first is complicated, the second is complex. Telling them apart is the foundation for everything that follows — because the two call for opposite methods.
Start by putting "complex" on a single axis. The left end is perfectly ordered, the right end is perfectly random, and complexity lives in the middle — only in the middle.
The left end: a salt crystal, atoms laid out in a perfect grid. See one corner and you know the whole thing, because it is a single motif repeated forever. One sentence describes it completely: "tile this cell out endlessly." A pendulum, a marching column, all live at this end. They look orderly, but they carry almost no information — no surprises, the next step always guessable.
The right end: television static, gas molecules colliding at random, a run of coin flips. Here every spot is unrelated to every other; stare at ten thousand points and you can't guess the ten-thousand-and-first. It looks information-rich, but the information is mutually unrelated and adds up to no structure — so describing it is also easy: one statistical parameter ("half heads, half tails") is enough, and the rest of the detail doesn't matter.
Here's the crux: both extremes are actually "simple". Order is simple because it compresses (one sentence); randomness is simple because it has no structure to speak of (one statistic and you're done). The genuinely hard thing is neither pure repetition nor pure noise, but the layer that has structure and surprise at once — a living cell, a language, a city, an ecosystem. It can't be squeezed into a sentence, and it can't be statistically averaged away; you have to keep almost all of it to describe it. That layer is complexity.
The trichotomy is not new. In 1948 the mathematician Warren Weaver (one of the founders of information theory) cut scientific problems into three: problems of simplicity (a few variables, the Newtonian kind), disorganized complexity (vast numbers of random variables, mopped up wholesale by statistical mechanics — a tank of gas), and, wedged in between and hardest of all, organized complexity — a moderate number of variables that are also entangled with each other, where life, economies and ecosystems all sit. He already saw that science was a triumph at both ends and nearly blank in the middle. Complexity science is the field that later went straight at that blank.
The middle layer has a louder name too: the edge of chaos — too ordered and a system freezes, too random and it falls apart; only pinned to the narrow band between the two can it hold structure while still throwing up new patterns. It's a useful phrase, and also the most abused one — its trouble waits for the last section.
English keeps the two words apart, which is lucky, because they name two entirely different things. The difference has four counts, and every one is practical.
First, can you take it apart. A complicated thing decomposes: strip a jet engine down to pieces, study each on its own, bolt it back together, and your understanding of the whole is intact. A complex thing dies on the operating table: put an ecosystem's species one by one into separate jars and the relations that made it an ecosystem — who eats whom, who pollinates whom — were severed the moment you started separating.
Second, is there a chief designer. Complicated things are usually designed — there are blueprints, someone accountable for the whole. Complex things typically have no central designer: nobody designed English grammar, and nobody designed a city's commute flow; they grew out of countless local actions on their own.
Third, is the behaviour predictable. A complicated-but-not-complex system, as long as it isn't broken, behaves repeatably — same input, same output, which is precisely why we trust machines. A complex system is full of loops and mutual influence, so the same starting point can run to wildly different outcomes (Topic 8, "chaos", takes that to its conclusion).
Fourth, and most consequential — where the extra comes from. Double the parts of a complicated system and you get a bigger machine. Double the units of a complex system and you often get a different kind of thing: ten people are a team, ten thousand are a market, and the laws of their behaviour aren't even on the same level.
Before you set out to solve a problem, spend thirty seconds deciding whether it is complicated or complex. The fastest of the four tests: can you split it, solve the pieces, and reassemble without losing anything? If yes — decompose as usual, bring in experts, add process, the finer the better. If no — stop; don't pour more detail into the model. For a complex system, "look more closely" does not bring you nearer the answer; you need to move up to a higher level of description (next section spells out that move). The cost of the wrong method isn't wasted effort, it's backfire: take the machine-disassembly approach to an ecosystem and you sever, with your own hands, the relations that made it work.
In 1972 the physicist Philip Anderson (a future Nobel laureate) wrote a four-page essay that rewrote how an entire field saw itself. The title was four words: More Is Different.
What he set out to oppose was a belief then almost unquestioned: reductionism. Reductionism says everything is built from more fundamental things, so once you understand the deepest particles and laws you have, in principle, understood everything; chemistry, biology, psychology above are just "applied physics".
Anderson granted the first half: yes, you can reduce anything down to fundamental particles. Then he cut the second half clean off: being able to take it apart is not the same as being able to put it back. Whenever a great many units gather together, new laws appear that don't exist in a single unit and can't be derived directly from the units' laws. One water molecule has no "temperature", and no "liquid" or "solid"; temperature and phase are concepts that only appear once water molecules are in bulk. One neuron has no "memory"; one person has no "inflation". These aren't applied-physics exercises — they are a new level, with its own vocabulary and its own laws.
Why does this matter? Because it turns "the whole is greater than the sum of its parts" from a platitude into an operational judgement: when you're facing a complex system, the move "dig one level deeper" has a ceiling — not because you didn't try hard enough, but because the law you're after simply isn't on the level below; it appears only on the level you're already looking at. To understand gridlock, taking apart each engine is useless; to understand inflation, brain-scanning every person is useless. You have to stay on the right level to find its laws.
When a problem gets "more detailed yet less clear the more you look", treat that as a signal: you may be on the wrong level. Concrete move — ask yourself, "the law I'm after, is it on the level I'm looking at, or have I drilled down to a level where it doesn't exist?" If several rounds of added detail haven't sharpened the judgement, stop adding; jump up a level and find that level's own macro-variable (temperature for molecules, price for trades, tail latency for requests). "One more level down" is sometimes not diligence — it's diligence in the wrong direction.
Now the other side. "Complex ≠ complicated" is a useful cut, but it too gets abused, and you should know three soft spots before you use it.
First, that axis has no sharp fence posts. Ordered, complex and random aren't three drawers but a continuous transition with blurry, even gradual, boundaries. The same system, viewed at a different scale or asked a different question, can slide from "this category" into "that one". So "this is a complex system" is rarely a black-and-white fact; it's mostly a judgement relative to the question you're asking right now. Treat it as a permanent label pinned to the system and you've gone wrong.
Second, "the edge of chaos" is the most romanticised of all. "Poised between order and chaos is where creativity lives" is such an attractive sentence that it's been carried into countless management and innovation talks. But as a testable scientific claim it stands on shakier ground. Around 1990 it was proposed that computational power peaks near a certain critical parameter (later written as λ) — beautiful to hear. Yet in 1993 Mitchell, Crutchfield and Hraber re-examined it with evolutionary experiments and found the "optimum" did not appear as reliably as advertised. The lesson: the moment someone reasons at you from "the edge of chaos", ask — this edge you speak of, what are its units, and how do you measure it? If there's no answer, it's just a pretty metaphor.
Third, and most abused: "complex" as a shield. "The system is just too complex" is often used to explain why something didn't get done or why responsibility can't be pinned. But complexity is not a disclaimer. Admitting a system is complex and not fully predictable means precisely that you should switch to a different toolkit (find a macro-variable up a level, run small reversible experiments, watch the coupling) — not throw up your hands. And people make the opposite error too: taking a problem that is really complicated but not complex, solvable by decomposition and expertise, and romantically calling it "complex" to duck the dull analysis it actually needs. Both misuses are the same laziness wearing the word.
Before you say "this is a complex system", run one self-check, the same ruler as the last section: can you name the macro-variable it should be watched on right now, and how you'd measure it? If yes — the judgement is operational, carry on. If not — you're probably just using "complex" to dodge the problem: either it's actually a complicated problem you should honestly take apart, or you haven't yet found the right level. "Complex" with no units, like "edge of chaos" with no units, is just a word.
Yes, but the unit of experiment has to change. You can't isolate one species and expect the relations to hold, but you can perturb at the level of the whole system: change a parameter, add a connection, watch how the whole responds. Topic 27, "agent-based modeling", is exactly this route — instead of dissecting the real system, regrow it from local rules inside a computer and then abuse the copy freely. The honesty is in this: what you study is always the whole with its relations intact, never the shards.
No. The new emergent laws never violate the base physics — water molecules always obey quantum mechanics, and temperature never breaks it. The new laws are what the base permits but does not specify: the underlying physics bounds the space of the possible, but doesn't dictate which macroscopic pattern a mass of units will settle into. So it's "added", not "opposed" — which is why anti-reductionism need not be anti-science.
It can, and that's the abuse to watch hardest this issue. The right consequence of judging something complex isn't surrender, it's shifting gears: from "look closer" to "find a macro-variable up a level, run small reversible experiments, watch the coupling and loops". If someone says "too complex" and then nothing changes and it's business as usual, the judgement wasn't made properly — it was supposed to bring a different set of actions.
"Simple" here means simple to describe, not low in information. A purely random string can't be compressed (in that sense its information is indeed maximal), but you also don't need to memorise it bit by bit — one statistical description ("each bit independent, fifty-fifty") captures everything useful about it, and the specific values don't matter. The complex thing loses on both counts: it won't compress to a sentence, and its specific detail can't be thrown away. Topic 41, on quantifying complexity, settles this tension formally.