Every concept was forced out by a concrete problem
2026-08-31 · Methods and Limits
Not one core concept in this field came out of "studying complex systems". Every one of them came out of some specific job that had to be done — how a gun hits a plane that is deliberately zigzagging, how a pot of chemicals grows stripes without breaking the second law. If you want to know where a concept applies, look up the job, not the definition.
Intellectual history is easy to write as a parade of names and dates, and it leaves you with nothing but "they were all very clever". That reading is useless. This issue does one thing: it reduces each concept back to the problem that forced it out — because how far a concept reaches is more or less the reach of that original problem.
The twist is that the trick runs in reverse too. Take any complexity-science term you own. If you cannot find the specific job it was built for — not "studying complex systems", an actual job — then it was probably never forced out of anything. It was just written down.
The last three issues asked how complexity is measured, how a model goes wrong, and what the field stands accused of. This is the last piece: where these ideas came from — who got cornered by what, how several disciplines ended up in one room, where the money came from and what it filtered out. The final section is reserved for something unflattering: how badly the method used to tell this story holds up.
In 1940 German bombers flew zigzag courses over England specifically to defeat anti-aircraft fire. A shell takes several seconds to reach altitude, so the gunner has to aim at where the plane will be, not where it is. That is a pure engineering problem: given a jittery trajectory so far, estimate the next position.
The mathematician Norbert Wiener and his assistant Julian Bigelow took the job. They rewrote it as a statistics problem: treat the pilot's manoeuvres as a noisy random signal and find the linear estimate that minimises prediction error. The report went to the US National Defense Research Committee (NDRC) in February 1942 as a classified document. The device itself was judged too complicated to build in wartime and the project was shut down in 1943.
Two things survived. One was the report, declassified and published in 1949, which founded the theory of signal prediction and filtering. The other was the thought Wiener had while chewing on the problem: the gun, the gunner, the plane and the pilot should be computed as one loop — the pilot sees the flak and changes course, the gunner sees the course and changes aim. Which parts are human and which are machine does not matter; what matters is how information goes round. In 1948 that thought became Cybernetics. → ref · Cybernetics and Its Lineage
This is not an exception. Turn the field's signature concepts over one by one and there is a concrete job pressed underneath each of them:
Take the last row. What Geoffrey West and colleagues set out to answer in 1997 was narrow: why does an animal's metabolic rate scale as body mass to the 3/4 power rather than to the first power? Their answer was that nutrients must be delivered to every cell through a fractal branching network, and how the efficiency of such a network changes with size pins the exponent. → ref · Allometric scaling The explanation holds only where there is a physical branching supply network. So it travels naturally to a city's pipes and cables — and when you push it onto something with no physical network (the activity level of an online community, say), the premise snaps. Not forbidden, but you owe yourself an account of which assumption you just threw away.
For any complexity term you are about to use, spend ten minutes on two questions: which paper it first appeared in, and what specific job that group was trying to do. If you find the job, copy down its assumptions — those are your limits of use. If all you find is "used to describe complex systems", keep the term out of your conclusions; at most let it into your questions.
From 1946 to 1953, ten conferences were held in New York under the title "Circular Causal and Feedback Mechanisms in Biological and Social Systems", paid for by the Josiah Macy Jr. Foundation — hence the Macy Conferences. The convenor was the neurophysiologist Warren McCulloch. In the room: the mathematicians Wiener, von Neumann and Shannon; the anthropologists Gregory Bateson and Margaret Mead; psychiatrists, psychologists, ecologists.
There is a clean description of this in the language of networks. Each group had been its own dense cluster: mathematicians see mathematicians, anthropologists see anthropologists, and almost no edges run between. What the Macy Conferences did was add a few shortcuts — a tiny number of them (ten meetings, a few dozen people) — and the distance between the clusters collapsed from "no path" to "three steps". That is the small-world mechanism. → ref · Small-world model
The trouble is that the same mechanism predicts something else, and it matters more: a shortcut has limited bandwidth, so it carries detachable units far better than structures that have to be swapped in whole. Words are detachable: "feedback", "information", "homeostasis" — copy them and go. Constraints are not: the stationarity assumption inside the filtering derivation, the definition of error, the observation that would falsify the model. Those have to travel with an entire body of mathematics, and they don't.
What followed fits exactly. "Feedback", "information" and "homeostasis" were inside anthropology, psychiatry, management and education within a few years; Wiener's filtering mathematics stayed roughly where it was. By the 1960s cybernetics' standing in the Anglo-American academy was sliding, and one reason cited again and again is precisely this: its words were everywhere and its testable content had not travelled with them. The Macy transcripts, read today, are serious work. This is what the shortcut mechanism does on its own.
To judge whether a cross-disciplinary exchange actually happened, don't look for shared vocabulary — look for one shared constraint: a criterion both sides accept that can rule a specific proposal out. If a cross-team meeting yields nouns and no prohibitions, don't hold the second one. Replace it with half an hour in which the other side's only job is to try to kill one of your concrete proposals. One round tells you whether the channel is open.
Intellectual history usually puts the money in a footnote. But if you believe in selection at all — variation is produced constantly, the environment decides who stays — then funding is not a footnote. It is the selection function itself.
To be precise: money does not make a conclusion true or false. Whether Wiener's filtering theory is right has nothing to do with who paid. What money changes is which questions get asked seriously, and for how many decades. And what a field ends up looking like is decided by exactly that.
From the 1940s through the 1960s, a large share of American basic research was funded through defence channels. That channel has a very specific taste: it prefers problems that improve control and prediction. So filtering and prediction, coding and channel capacity, operations research and optimisation grew fast — their acceptance criteria were clear, results came quickly, and the results were legible to the funder. The historian of physics Peter Galison put it harder in his 1994 essay "The Ontology of the Enemy": cybernetics' whole worldview — treat the adversary as a modellable, noisy servo system — is the shape of the fire-control job.
Now look at the other end. The Santa Fe Institute (SFI) was founded in 1984. Apart from David Pines and Murray Gell-Mann, its founders — George Cowan, Nicholas Metropolis, Peter Carruthers and others — were Los Alamos scientists. It was first called the Rio Grande Institute, and its founding workshop was held on 5–6 October 1984. → ref · The Santa Fe Institute and its lineage
SFI's founding motive is recorded again and again in the same phrase: a place to do theoretical research outside traditional departmental boundaries and outside government agency science budgets. That sentence matters, because it concedes that the earlier filter existed and states plainly that this is an attempt to swap it. What grew afterwards fits: at the September 1987 workshop, Kenneth Arrow brought ten theoretical economists and ten physicists and biologists, the volume The Economy as an Evolving Complex System came out of it, and in 1988 it became SFI's first resident research programme. Under the defence filter, that sort of topic could hardly have won long-term support — not because it was wrong, but because it could not deliver that kind of deliverable.
Before reading any field's review article, spend twenty minutes on where its money has come from for the last twenty years and what that source accepts as a deliverable. You come out with a list of questions nobody worked on — not refuted, just off-criterion. Almost everything cheap that is left to be picked up is on that list.
If this field really has caught hold of some cross-domain mechanisms, that should leave a testable trace: the same rule will be invented over and over by people who never met, each giving it a new name. It happened, and the clearest case is "the more you already have, the more you get".
The rule is one sentence: the more of something a thing already holds, the higher its chance of getting more. Rich gets richer, poor gets poorer, and the resulting distribution is the brutally unequal shape called a power law. → ref · Preferential attachment
How many times it has been invented:
Merton's case is worth pausing on. In 1968 Robert K. Merton published a piece in Science about how scientific credit is allocated: for the same work, the already-famous name on the paper takes nearly all of the recognition. He called it the Matthew effect, after the line in Matthew's gospel — "unto every one that hath shall be given". That is the "more gets more" mechanism turned on knowledge production itself.
Which yields a neat self-reference: the mechanism this field uses to explain wealth, cities and networks is simultaneously deciding who gets remembered as the inventor. The statistician Stephen Stigler formalised it in 1980 as Stigler's law of eponymy: no scientific discovery is named after its original discoverer. He added the twist himself — the law should be credited to Merton.
This has a practical consequence for reading. When a concept has appeared in exactly one field but is being sold as a "universal mechanism", treat it as a local fact about that field. Conversely, when the same generating rule has been written down independently in four or five unrelated fields, the rule is worth your time — but spend it on the rule, not on any of the names. → ref · Identifying (and misidentifying) power laws
Run a "reinvention search" on any concept: look for the same generating rule in three or four fields' jargon (search the rule, not the name). Three or more independent sources → put your weight on the rule itself and read the earliest source that states it cleanly. Exactly one source, but marketed as a universal mechanism → demote it to a local fact and stop using it for cross-domain inference.
The previous four sections did something rather arrogant: they used the field's own mechanisms to explain the field's own history. That is tidy. Tidiness is exactly where to get suspicious.
First, and worst: a selection story can barely be falsified. For "defence funding filtered out certain questions" to be a claim rather than a mood, you need to compare two populations — the questions that were funded and left records, and the questions that were not and therefore left none. The second group has almost no data: never applied for, rejected, thought of but never written, written but never cited and gone. None of it enters the denominator.
Second, "concepts are forced out by problems" misses an entire category. Some concepts were not forced by any external job; someone just worried at them. Benoit Mandelbrot chewing on "pathological" curves had no client waiting for the length of a coastline; John Conway designing the Game of Life was building something close to a mathematical toy. → ref · Cellular automata Section 1's search comes back empty on those — and when it does, the correct response is "this concept is of another kind", not "this concept is suspect".
Third, section 4's reinvention argument may itself be stepping into a hole this site dug two issues ago. Pareto, Zipf, Yule and Simon were not strictly talking about one thing: more than one mechanism generates power laws (multiplicative random processes, extreme-value statistics, critical phenomena all do), and rolling them all into "six names for one mechanism" is precisely the move of treating "looks alike" as "is the same". The described/generating split in the figure exists for that reason — but even among the four green ones, equivalence is conditional and has to be checked case by case, not stamped in a sentence.
Fourth, sample size. This field's history has a few dozen key episodes, no control group and no repeatable experiment. Using the tools of a science of large-sample statistical regularities on a case with n that small is not obviously self-consistent. So read this issue as follows: sections 1, 2 and 4 give you moves you can actually run (find the original problem, check whether constraints crossed the bridge, search for reinvention); section 3 gives you a hypothesis worth suspecting and short of proof. Filing them at the same grade is the easiest mistake this issue offers.
Not illegitimate — priced. The cost of generalising is stating explicitly which assumption of the original problem you are discarding and what replaces it. West's 3/4 law extends to cities because cities really do have physical supply networks; extended to something without one, you first owe an answer to "what limits transport here?". No answer means it isn't a generalisation, it's a loan of vocabulary.
Because the banner carried too much, and the testable parts were long ago broken up, renamed and absorbed elsewhere: feedback and stability into control engineering, information and channels into communications, state estimation into filtering and navigation, self-organisation into physics and biology. A shared banner lets anyone hang anything on it, so nobody owns its bad debt — while each piece that moved out acquired its own acceptance criteria, where bad debt cannot be parked. The lesson applies to any large word today.
Yes, with one word changed: read the earliest source that states the rule clearly, which is not necessarily the earliest to notice the phenomenon. Pareto saw the distribution in 1896; Yule wrote down the generating rule in 1925. What separates them is not time but the difference between description and mechanism — a distinction this site keeps insisting on.
Yes, at a lower grade. An unfalsifiable hypothesis cannot go in your conclusions, but it can go on your list of things to look up: it hands you the action "find out where this field's money comes from", and that action produces checkable facts — funders, terms of award, which projects were declined. An untestable hypothesis can still generate a productive search. The difference shows up in what you finally write: "because of funding, therefore X", or "nobody worked on these questions at the time; here is the list".
There is no model answer, but there is a concrete first question: where does this field's money come from now, and what does that source accept as a deliverable? If the answer is "a high-impact-factor paper with a model that looks universal", then what gets filtered out is work that explains only one system but explains it to death — and several earlier issues here have argued that the second kind is worth more.