Why High-Output Systems Are Often the First to Stop Growing为什么最高产的系统往往最先停止成长

On two kinds of novelty: the kind that compounds and the kind that only piles up.

“The limits of my language mean the limits of my world.” – Wittgenstein, Tractatus 5.6

For a while it looked like progress.

For a week, an AI agent pipeline I ran kept shipping. A commit landed roughly every hour: small fixes, minor improvements, an activity graph that looked alive. But the product itself did not get any bigger ¹. Nothing was broken and nothing was idle; the pipeline was busy, but only within the vocabulary it already had.

This is a common confusion. Conway’s Game of Life has kept producing new patterns for half a century, and when a long-standing open question about its oscillators was settled in 2023, the community did not wind down; it celebrated for a week and turned its new tools on the next problems. Speedrunners deliberately ban whole classes of glitches and still find faster routes within the smaller rule book. Super Smash Bros. Melee, a fighting game whose code has not been officially patched since 2001, still has a thriving tournament scene. Most viral internet ideas, by contrast, burn out quickly and are dropped.

From the outside, all of these look like activity. The easy verdict is “the well has run dry,” but that is usually too crude: the Game of Life community could be plateauing, the agent pipeline could still be growing, and the viral fad might come back in another form. What separates these cases does not show up in the usual activity metrics, but it can be named, and in principle measured.

So the question is: what is the difference between a system that is producing and a system that is growing? What separates motion from progress? A gut feeling is not a measure, so the difference deserves a name.

The distinction itself is not subtle; after the next two sections, most readers will find it obvious. Yet most of the systems we run, build, and reward are exactly the ones that blur it. So this essay is less about the distinction than about the structural reason we keep falling into it.

More Is Not New

There are two kinds of novelty, and we routinely confuse them.

The first is instance novelty: an output not seen before, such as a new solution, a new feature, or a new generated artifact. The second is primitive novelty: a new reusable building block that enters the vocabulary and changes what becomes cheap to say or do next.

The trap is that a finite grammar can generate infinitely many sentences; linguists, after Humboldt, call it making “infinite use of finite means” ². A system can produce an unbounded stream of genuinely new instances without adding a single new primitive. It looks open-ended from the outside while its vocabulary has stopped growing: there is always something new to say, but nothing new to say it with.

Once you see the distinction, examples are everywhere. A pop-song generator can produce ten thousand chord progressions without introducing a new chord. A research subfield can publish for a decade, every paper different, while still running on the three theorems it started with. An organization can ship features for years on an architecture nobody is allowed to question. A solo creative practice can move between one-off experiments, each piece new and none of them making the next one easier. In each case the outputs are new but the vocabulary is frozen: the system is productive without becoming more generative.

Fig 1: Instance novelty vs primitive novelty. Filling a level with more instances (horizontal) is not the same as promoting a new reusable primitive that opens a level (vertical). A finite grammar yields unboundedly many instances at fixed primitive height. The ladder is the Game of Life’s classic one: gliders, guns that emit them, logic gates built from glider streams, and computers built from the gates.

Cellular automata give the cleanest example, and here the record is public. In 2023 the last two missing periods were found, and Conway’s Game of Life was shown to have an oscillator of every period: it is “omniperiodic” ³. The question, is there a pattern that blinks with period N?, had been open for as long as there are records of hobbyists discussing it, and the way it closed is the distinction in miniature.

Most periods were never settled one pattern at a time. In 1996 David Buckingham built a set of “Herschel conduits”, pieces of track that carry a signal and can be joined into a closed loop of any length; with them, an oscillator of every period from 61 up can be assembled. Later Mike Playle’s Snark, a small and fast glider reflector, brought that bound down to 43. Two constructions answered infinitely many questions. What they could not reach had to be found one period at a time, through new mechanisms and ever better computer searches. Of one common kind of search, the paper says each result “is a roll of the dice: there is no way of specifying in advance which period you would like to find”. At the turn of the millennium twelve periods were still missing. The last two, 19 and 41, fell in 2023.

So the finish line was crossed by instances, but most of the distance was covered by primitives. And the primitives are what lasted. The paper ends by noting that the problem “has served as a rallying point for the Life community, leading to the creation of the many tools and programs” it describes, and then lists the questions still open, in one of which even period 7 is unresolved. One of the authors described the mood to Quanta: “Everyone’s celebrating for a week — and then moving on to other things. There are so many other problems to solve” ⁴. The parts were reused at once: the first true period-41 glider gun was built the day after the first period-41 oscillator, and guns of other missing periods followed through 2024 and 2025, many of them from the same kinds of hasslers and reflectors ⁵. The oscillators were instances; the conduits, reflectors and search programs were primitives. The two run on different clocks, and confusing them is an easy way to misjudge whether a field, a community, or any other system is still alive.

None of this is a new distinction, and it is worth saying where it comes from. In 1969 the biologist C. H. Waddington wrote that it is “not sufficient to produce new mutations which merely insert new parameters into existing programmes; they must actually be able to rewrite the programmes” (quoted in ⁶). Artificial-life research has since made the idea formal. Banzhaf and colleagues separate variation, novelty inside a model, which “does not change the set of entities” and so leaves the space of possibilities as it was, from innovation, which adds a new type to the model, and emergence, which changes the terms the model itself is written in ⁷. They are stricter than this essay: variation, for them, is not open-ended at all, only “the ‘normal’ regime of any dynamical system”. Taylor calls the same three tiers exploratory, expansive and transformational, in terms close to Margaret Boden’s account of creativity ⁶. What follows adds something narrower: a test for when a new element counts as a primitive, a way to watch the two kinds of novelty separately over time, and an account of why the systems we build keep losing sight of the second.

The Two Curves

The rest of the argument needs only one picture. Run an open-ended system over time and track two numbers.

$g_{\text{inst}}$ (instance yield): how surprising each new output is, given everything before it; in other words, how much genuinely new stuff is arriving. Surprise has to be of the kind that can be learned from: noise is maximally surprising and adds nothing, which is why open-endedness researchers ask for novelty that stays learnable ⁸.

$g_{\text{prim}}$ (primitive yield): how often a genuinely new reusable building block is promoted into the vocabulary.

Fig 2: The two-curve criterion. Instance-level marginal complexity $g{\text{inst}}$ stays bounded away from zero while primitive-level marginal yield $g_{\text{prim}}$ decays to zero: unboundedly many novel instances, yet generative closure._

What matters is not the level of either curve but their divergence. The failure that looks healthy from a distance goes like this: $g_{\text{inst}}$ stays high and new instances keep coming, while $g_{\text{prim}}$ decays toward zero and no new building blocks appear. The result is infinitely many instances with generative closure, and the activity metrics stay green exactly when the curve that matters has gone flat.

Schmidhuber’s theory of curiosity draws a similar line from the observer’s side. What makes data interesting, on his account, is not how compressible it is but how fast it is becoming more compressible: “compression progress”, the first derivative. Pure noise and fully predictable data both score zero, because neither lets the observer find anything new to compress ⁹. $g_{\text{prim}}$ is close to that progress term. Once a system stops finding regularities in its own history, its output can stay varied forever and still teach it nothing.

The same shape appears at very different scales. A research group that keeps publishing variations of the same trick has high $g_{\text{inst}}$ and near-zero $g_{\text{prim}}$: every paper is new, but the toolkit is frozen. A craft practice that moves from one one-off experiment to the next has the same signature: a new piece each week, none of them building much on the last. A late paradigm in science can look similar: work continues and anomalies are absorbed, but no new explanatory primitives enter the canon. In each case the first useful question is where new primitives could form, rather than how to produce more instances of the old ones.

The shape has also been measured on real data. Every US patent is tagged with technology codes, the patent office’s vocabulary of capabilities. Following the record across the nineteenth and twentieth centuries, Youn and colleagues found that patents, codes and combinations of codes grew together until about 1870. After that, new codes came much more slowly, while patents and new combinations kept growing in step, at about six new combinations for every ten patents ¹⁰. Invention did not stop; it changed kind, toward what the authors call “tinkering, gradual modification and refinements” on a vocabulary that grew only now and then. This is not a story of failure. Patents are still useful, and the authors argue that the occasional new code is “more than enough to sustain a combinatorial inventive search”. But it is the two curves, drawn by an institution for well over a century. One detail anticipates the next section: when the office introduces a new code, it goes back and reclassifies older patents that turn out to have embodied the capability all along. A new primitive is recognized partly by what it lets you say about the past.

Open-endedness research has run into the same wall in its own systems. POET invents obstacle courses for simulated walkers and learns to cross them, and it tracks progress by counting new courses that are both created and solved. In the original version that count stopped rising after about 20,000 iterations, and the authors of the follow-up gave the reason plainly: the generator “can only sustain a finite number of obstacle types with predefined regular shapes and limited variations” ¹¹. More courses, same obstacles. The fix, in Enhanced POET, was a richer way of encoding terrain, in effect a new vocabulary, and the researchers designed it; the system did not grow it.

One caution about reading the second curve. Even a vocabulary that grows forever grows more slowly as it gets bigger. In books, music listening, Wikipedia and social tagging, Tria and colleagues found the number of distinct words, songs, pages or tags rising more slowly than the volume of activity, so that the rate of new ones falls over time as a power of $t$, a pattern known as Heaps' law ¹². A falling $g_{\text{prim}}$ is therefore normal and, by itself, not a diagnosis. The question is what the total does. Under a power law it keeps growing without bound; under an exponential decay, the shape in Fig 2, it levels off at a ceiling. A slowing vocabulary and a closed one differ in the tail, and that is also what separates a mature system like the patent office from a finished one.

So the first practical step is simple: watch both curves, not one. Output volume and instance novelty are vanity metrics that can look great during a plateau. The number to watch is primitive yield, which behaves very differently. The artificial-life literature has suggested as much: Banzhaf and colleagues propose applying its counting measures to each kind of novelty separately, “to determine if the activity of the respective type is continuing” ⁷.

What Counts as a New Primitive

“Reusable building block” is still vague. Program synthesis offers a precise test, and it is the one formula in this essay worth remembering.

A candidate counts as a primitive only when adding it to the library makes the total description of past work shorter. Write the sum of two costs:

$$\text{cost}(\text{library}) ;+; \text{cost}(\text{work, expressed in terms of the library})$$

An abstraction has a cost, because the library grows; it pays off only if the work written with it gets cheaper. The demanding part of this test is that it is retrospective. You cannot judge a new “abstraction” by its design, or by how clean it looks in code review. You have to apply it to old work and see the description get shorter. If it does not compress past work, it probably will not compress future work either, at least while future work resembles past work; I come back to that condition below. This is the working form of the distinction the rest of the essay builds on.

Library-learning systems do exactly this. Stitch states its objective in this form: an abstraction is worth the size of the corpus before, minus its size after being rewritten with the abstraction, minus the size of the abstraction itself ¹³. DreamCoder frames the same step as Bayesian inference, and its authors gloss the result as a preference for “libraries that best compress programs found during waking” ¹⁴; because it refactors programs before comparing them, it can find an abstraction such as map that appears verbatim in none of them. Both build the library greedily: find the abstraction that compresses most, rewrite everything with it, and repeat until nothing more compresses. DreamCoder’s authors also ran the obvious control: a variant that simply added whole past solutions to its library, keeping them without compressing them, solved fewer held-out tasks in every domain ¹⁴. Keeping everything is not the same as learning. Nothing here depends on intuition. A reusable skill is one that lets past work be re-described more compactly, because it factors out a pattern that kept being derived again. That gives a concrete test of whether a system is accumulating capability or only output: is its library compressing its own history?

The test has known failure modes, and they are instructive. Compression can prefer the less general abstraction: in a case reported by DreamCoder’s own authors, the most compressive way to draw polygons beats a slightly longer one that also takes the side length as a parameter, the one that intuitively generalizes better ¹⁵. And compressive is not the same as usable. When the authors of LILO gave a language model the abstractions Stitch had found, which come out as anonymous functions with numbered names, the model solved fewer problems, by 31 percentage points in one domain. Once another step gave each abstraction a readable name and a short description, results improved in two of the three domains ¹⁶. A primitive nobody can call by name is not yet part of anyone’s vocabulary. The test also has a convention built into it. “Shorter” depends on how each piece is priced, and systems price them differently: Stitch uses hand-set costs, DreamCoder a learned distribution. Theorists of minimum description length warn that the crude two-part form of the idea “is in danger of becoming arbitrary” for exactly this reason ¹⁷.

The deeper limit is in the word past. The test assumes that future work will look like past work. When the tasks themselves drift, people do something these algorithms do not: they choose abstractions for the work they expect next. Hernandez Cano and colleagues found this in a pattern-building experiment; people’s choices were consistent with what the authors call prospective compression, and “cannot be captured by existing retrospective compression-based algorithms” ¹⁸. When the sequence of tasks was steady, the retrospective algorithm did about as well as an oracle. So the rule holds with a condition attached: past compression is good evidence while the world stays put, and weak evidence when it is changing, which is exactly when a new primitive matters most.

This also changes what a good open-ended structure is for, whether it is a research group, an artistic movement, a software organization, or an autonomous system. Completing tasks is not enough for generativity ⁸. The system needs a persistent, shared library and an incentive to compress into it: some reason for someone, or something, to notice that the same pattern has appeared five times and should become a primitive. Without that, every cycle starts from the same vocabulary, and the primitive curve never rises.

Two very different literatures describe what such a library does. Primatologists distinguish chimpanzee traditions, which vary “within the species' existing cognitive repertoire”, from human culture, which “accumulates modifications over time”: improvements stay in the population “with relatively little loss or backward slippage until further changes ratchet things up again” ¹⁹. The ratchet is the library. And when Arthur and Polak simulated invention by wiring logic gates together at random, a circuit that met some need was kept “as a new component that can serve as a building block”. Complex circuits appeared only by way of simpler ones: when one intermediate goal, the full adder, was left out, “not even a 2-bit adder was found in one million steps” ²⁰. A library grows through stepping stones, and something has to decide which ones are worth keeping.

Why We Keep Forgetting This

So far this is mostly a definition, and an obvious one. Anyone with experience running systems already knows, in some form, that output is not capability. So why does the same pattern keep appearing, in product after product, agent system after agent system, lab after lab?

The reason is structural. Every running system has an accounting layer: a changelog, a velocity dashboard, an OKR sheet, an evaluation harness, a paper count. These track output; none of them tracks library compression. The changelog records what the system did; the library records what it became. The two never share a dashboard, because the second is hard to define and harder to reward, and people optimize what they can see being paid for.

None of this is news to economists. In 1975 Steven Kerr collected cases of what he called the folly of rewarding A while hoping for B, and named this one directly: “some parts of the task are highly visible while other parts are not”, so that “publications are easier to demonstrate than teaching” ²¹. Holmström and Milgrom later worked out the mechanism. When an agent has several tasks and only some can be measured, pay for the measured ones draws attention away from the rest: “if volume of output is easy to measure but the quality is not, then a system of piece rates for output may lead agents to increase the volume of output at the expense of quality”, or to neglect shared equipment. Their sharpest case is an asset whose value changes over time and cannot be measured. There the best contract gives “only muted incentives” for output, so that effort is not drawn away from maintaining the asset ²². A library is that kind of asset. Its value is how cheap it makes the next piece of work, and no quarterly count sees it change.

In practice, what gets rewarded is what gets measured. A pull request closes a ticket, and the closing is visible. An engineer who instead spends a week retiring three modules in favor of one cleaner primitive does about as much work and produces nothing the system can count as “shipped.” Over enough quarters, the incentives wear the library down, and the system drifts, predictably, toward higher $g_{\text{inst}}$ and lower $g_{\text{prim}}$.

Engineers say so when asked. In a survey on refactoring at Microsoft, one respondent put it this way: “How do you measure the value of a bug that never existed, or the time saved on a later undetermined feature? … it makes it difficult to justify to management.” More than half of the responses said refactoring was driven by “immediate concrete, visible needs” rather than by long-term maintainability ²³. Lehman’s second law of software evolution describes the result: as a program is changed, its complexity grows “unless work is done to maintain or reduce it” ²⁴. A well-known paper on technical debt in machine-learning systems proposes a cultural remedy: teams that “reward deletion of features, reduction of complexity” as much as they reward gains in accuracy ²⁵.

Consider two systems. The first ships a feature every day, its dashboard is green, and its team gets praised. The second ships nothing for a month while its team rebuilds an internal representation that nobody outside will ever see. Six months later, the second is producing twice as many primitives per quarter as the first, and the first is heading toward exhaustion, yet most accounting layers show neither fact. Read honestly, the picture is the opposite of the obvious one.

The same pattern appears outside commerce. Research labs that count publications converge on instance work, and the fields they make up show it. Across 90 million papers in 241 subjects, Chu and Evans found that as a field publishes more each year, citations concentrate on papers that are already well cited, the list of most-cited papers stops changing, and new papers become unlikely ever to join it: “a deluge of papers does not lead to turnover of central ideas in a field, but rather to ossification of canon” ²⁶. Milojević, measuring how much conceptual ground physics covers by the variety of phrases in its paper titles, found that this ground grows far more slowly than the number of papers: physics is still expanding, “but at a rate that is slower than at any time in the last 50 years” ²⁷. A better-known result, that papers and patents are becoming less “disruptive” ²⁸, rests on a citation measure whose decline is still being argued over ²⁹.

AI agent benchmarks that score pass rates select for agents that solve more problems with the same vocabulary, not for agents that grow a vocabulary at all. Chollet made the general point years ago: “skill is merely the output of the process of intelligence”, and with enough priors or training data an experimenter can “buy” arbitrary levels of it ³⁰. Even systems built to grow a library can fool their own scoreboard. Two language-model systems that reported gains from learned libraries of functions turned out, on closer study, almost never to reuse those functions; the gains came mostly from self-correction and repeated sampling ³¹. Generative-AI products priced by the token are economically indifferent to whether the next token came from a compressed concept or a re-derivation. In each case the accounting layer cannot see the variable that matters, so the system optimizes a proxy.

So the problem is not that “more is not new” is a subtle insight that organizations fail to grasp. The insight is obvious, and almost everyone running these systems would agree with it at a whiteboard. The systems still produce the wrong answer because nobody has written the right one down where it affects what gets paid.

The Second Trap: Lock-In

Now suppose the primitives are available: the current framing of the problem is exhausted, but a better one exists, such as a new representation, a new theoretical toolkit, or a new architecture that would reopen the primitive curve. Does the system move to it?

Usually not, or not for a long time. This failure differs from “nothing left to find”: it is lock-in, where the system should move and cannot. Kuhn’s account of scientific revolutions is largely a study of it, and he put it in the language of cost: “As in manufacture so in science—retooling is an extravagance to be reserved for the occasion that demands it” ³². A field drops a paradigm only when another is ready to take its place; “to reject one paradigm without simultaneously substituting another is to reject science itself.” Even then the move is slow. For one scientist, Kuhn compares it to a gestalt switch, all at once or not at all; for the community it is “an increasing shift in the distribution of professional allegiances”, and Copernicanism “made few converts for almost a century after Copernicus' death”. He quotes Planck on the bleakest version: a new truth wins “because its opponents eventually die, and a new generation grows up that is familiar with it.” Economists have since tested Planck. When an eminent life scientist dies unexpectedly, papers in the field by outsiders rise markedly, draw on different literature, and are disproportionately likely to be highly cited; while the star was alive, outsiders appear reluctant to challenge the field’s leadership ³³.

The model below is my own reconstruction, not Kuhn’s; he denied that paradigm choice is a comparison of problem-solving ability, calling it a decision that “can only be made on faith”. But the reconstruction is useful. A member of a system switches frames when its gain from switching, $\Delta g$, exceeds its cost of switching, $c$, and stays whenever $c > \Delta g$. The catch is that both numbers depend on what everyone else does: moving alone costs the most and gains the least. Two kinds of lock-in follow, and economists have names for both.

Pure convention. The alternatives are about equally good, and the system is stuck with one for historical reasons because changing is not worth the cost. Which side of the road a country drives on is the clean case. Liebowitz and Margolis call this first-degree path dependence: a path “that cannot be left without some cost, but that path happens to be optimal (although not necessarily uniquely optimal)” ³⁴. QWERTY is the famous case, and it has been told both ways. For years it was the textbook example of lock-in to an inferior layout, “standardization on the wrong system” ³⁵. A closer look at the studies behind the Dvorak keyboard’s reputation found “no scientifically acceptable evidence that Dvorak offers any real advantage over Qwerty” ³⁶; the argument has not entirely ended. This kind of lock-in is often harmless.

Coordination lock-in. The new frame is genuinely better: summed over everyone, the gain exceeds the cost of relearning the shared vocabulary, rebuilding the tooling and agreeing on new standards. But no one gains by moving alone, because what the new frame is worth depends on others moving too. Farrell and Saloner call this “excess inertia”. In its purest form everyone prefers the new technology “and yet they do not make the change”, which “is purely a problem of coordination”; in another, everyone is a fence-sitter, “happy to jump on the bandwagon if it gets rolling but insufficiently keen to set it rolling themselves” ³⁷. Arthur showed how such states arise: when returns increase with adoption, small early events give one option a lead that the dynamics do not forget, and the lead locks in whether or not the option was the better one ³⁸. This is the expensive kind.

Astronomy is the classic case of lock-in in science, though not in the way it is usually told. The popular story has Ptolemaic astronomers stacking epicycles on epicycles until the system collapsed under its own weight. When Gingerich recomputed the thirteenth-century Alfonsine Tables, he found them “based on a pure Ptolemaic theory”, with parameters “almost all identical to those originally adopted by Ptolemy”, and called the embroidered system “a latter-day myth” ³⁹. The real history is more instructive. Copernicus’s system used about as many circles as Ptolemy’s and gave results “as accurate as Ptolemy’s, but it did not give more accurate results” ⁴⁰. On the measure everyone used, predicting where the planets would be, the better frame was not yet better. Its gain lay in coherence and in what it would make possible, and that is the hardest kind of gain to count when deciding whether to move. Laudan later turned this into a rule of method: judge a research tradition for acceptance by how much it has solved, but for pursuit by how fast it is progressing, since “it is always rational to pursue any research tradition which has a higher rate of progress than its rivals (even if the former has a lower problem-solving effectiveness)” ⁴¹. That is the primitive curve, used as a reason to move.

Mature engineering organizations can end up the same way when their process has no room for the leap, so that “pause delivery for three weeks and rethink the topology” is not a proposal the structure can accept ⁴². Hannan and Freeman argue that this is less a failure of management than a product of selection: organizations survive by being reliable and accountable, which requires structures that reproduce themselves faithfully, and such structures resist change ⁴³. A frame does not question itself.

Fig 3: Lock-in as a potential landscape. Even when a better framework $V'$ exists ($\Delta g > 0$), the collective stays in $V$ if the escape resistance $[c - \Delta g]+$ exceeds the available perturbations: a metastable basin. Pure-convention lock-in is the special case $\Delta g \approx 0$._

Putting the two traps on perpendicular axes gives a fuller picture.

Fig 4: Two axes generate four corners. Generativity (does discovery keep producing new primitives?) is the horizontal axis. Epistemic lock-in (when yield decays, does the collective stay?) is the vertical. Low-generativity, low-lock-in is the flash-in-the-pan fad: once people see through it, they leave. Low-generativity, high-lock-in is the long-lived culture on closed rules: a game whose code stopped changing decades ago and whose competitive scene is still going. High-generativity, high-lock-in is the open community deliberately bounding itself: speedrunning a still-rich game under “glitchless” rules. High-generativity, low-lock-in is sustained deepening: Game of Life fifty years on.

The corners are tendencies, not boxes, and the real cases are more interesting than the labels. Super Mario Bros. came out in 1985; its “glitchless” category bans five named glitches, and the record still fell in 2022 and again in 2024 ⁴⁴. Melee’s players have not only stayed with their game but taken over its upkeep: a fan-made fix to how the game reads controllers was adopted at major tournaments within weeks of its release in 2017, and it is now a required part of the community’s own online build ⁴⁵. That is lock-in so complete that the community maintains the frame itself.

Most writing about open-endedness looks at one axis at a time. With a stalled system, it helps to ask about both before settling on an answer. Even a genuinely generative system will plateau if it is tied to a representation whose primitive curve has flattened, and no amount of local cleverness will get it out. Migration is a different operation from optimization: it has a real cost, and someone, usually still a human, has to decide to pay it.

One more point applies to both traps. What we can measure is never a system’s raw capacity for producing primitives, only how many primitives it actually produced, given where its attention went. Two systems with access to the same material can end up with very different libraries depending on how they allocate attention ⁴⁶. March’s own summary is bleak: adaptive processes, “by refining exploitation more rapidly than exploration, are likely to become effective in the short run but self-destructive in the long run.” Levinthal and March later named the mechanism the competency trap: as an organization grows more competent at an activity, it does more of it, “thus further increasing competence and the opportunity cost of exploration” ⁴⁷. Lock-in is the special case where attention is fixed on a flat curve and the structure will not let it move.

The Hard Part Is Not the Idea

The argument consists of a definition, a test, an explanation of why systems keep failing the test, and a separate trap that catches those that pass it. None of these pieces is difficult. What is difficult is that knowing the distinction does not get a system out of the trap, because the trap lives in the accounting layer, not in anyone’s understanding.

The counting can be crude and still be useful, as long as it counts reuse rather than size, since a library can grow while almost nothing in it is used again ³¹. For a codebase: of the abstractions added this quarter, how many are now called from code their authors did not write? For a research group: which of this year’s methods did anyone else, or the group itself a year later, pick up as a tool rather than cite as a result? For an agent: are new skills called by later skills, or only once, by the task that produced them? Voyager, an agent that writes its own library of skills for Minecraft, tended to plateau in its later stages when that library was taken away ⁴⁸. Each of these is a reuse count. None of them is on a standard dashboard.

The working version of the test is short. A system’s real state is not in its output but in its library: the changelog says what it did, and the library says what it became. Most measurement counts only the first. The few systems that grow have found some way to count the second as well; the rest stay busy.