The short version: The modern economy’s growth in output per person — around 1.5% a year — is, in effect, the fraction of the population that automation frees up each year to go do something else. The “System of the World” — governance + free markets + engineering — is precisely tuned to reabsorb that rate. AI now automates the automation layer, which means the freed-up rate itself starts to climb. When it climbs past what the system can reabsorb, the system breaks. The only open questions are how fast and whether we build the next system before the current one fails violently.

Who this is about: the rich-world core

A boundary first. This thesis is about the rich-world core — roughly 1.2 billion people, about 15% of humanity^{SourceThe World Bank classifies ~87 economies as “high income,” with a combined population of about 1.42 billion (2024). The “rich-world core” used here is a deliberate subset of that — high-income economies that also allocate labor through the free market and permit internal mobility — so the ~1.2B figure is my own narrower estimate, not the World Bank aggregate. World Bank, income classifications 2024–25Blue numbers are sources — hover any of them for the citation and a link. The orange dashed terms are short readable explainers.}: the high-income market economies that allocate labor through the free market, feed it with mass education, and let people relocate freely to where the work is within their own borders — within a country, or within the EU’s internal free-movement zone. The US and Western Europe are the template; Japan, South Korea, Taiwan, Canada, Australia, Israel, and a handful of others run the same operating system.

It deliberately excludes two large groups. Economies where the state directs employment or restricts internal movement are out — China’s hukou household-registration system^{Background · ChinaThe hukou system (1958) ties a person’s access to welfare, schooling, and public services to their registered home locality — historically blocking rural residents from settling for work in cities. It’s a state restriction on internal mobility, the opposite of the free-market labor allocation this thesis runs on. Partial reforms began in 2014. Overview} disqualifies it on both counts. And the lower-income world is set aside not because it’s untouched, but because displacement there lands in a vast informal and subsistence sector — a different, and harsher, shock absorber than the institutional one described below.

That institutional absorber is the whole point. The machine below is the one that’s been tuned, over 250 years, to a specific rate. Everything that follows is about what happens when that tuning fails.

The economy is a machine, and everyone in it gets a role

Within that core, those ~1.2 billion people get sorted into functions, which come in three broad kinds:

• Frontier work — expanding what we know. Scientists, mathematicians, and artists pushing the edge of understanding and representation. Each new piece of fundamental knowledge grows the toolset available to everyone else.
• Automation work — applying what we know to how we make things. Engineering, definitionally, is the practice of applying knowledge to enable capital and energy to reduce the human labor required to make something.^{Background · VocabularyWhat I’m calling Engineering here is close to what endogenous growth theory (Romer, 1988) calls Innovation: knowledge as both input and output of production. The framings overlap, but the one here specifically picks out the application of frontier ideas to production — the application slice inside Romer’s broader category. Same animal, different label. Endogenous growth theory · overview}
• Running the machine — keeping the day-to-day system working. Government, education, healthcare, agriculture, manufacturing, logistics, news, finance. Cells in the body, cogs in the machine.

All three kinds of role have existed since human society existed — in earlier, smaller societies as part-time functions rather than full-time specializations. But the specialization isn’t the genuinely new thing. What was new with the modern system was the combination: Enlightenment ideas assembled into a machine designed to change over time — government that enables the self-optimizing market rather than imposing a rigid hierarchy on it. Not a fixed order to be maintained, but a system built to adapt.

The free market is a self-optimizing allocator

Take a news beat nobody’s covering. Inside any single organization, an editor might deliberately assign reporters to it — that part is a real decision, not market magic. The market’s actual work happens between organizations. If one outlet ignores the beat, a competitor can pick it up and win market share as its reward; or it turns out nobody cared, and the outlet that stayed lean is the one that thrives. Either way the allocation gets made — not by a central planner deciding what’s worth covering, but by competition rewarding whoever guessed right. That selection pressure, running across thousands of niches at once, is the source of the market’s efficiency.

This needs supporting technology — currency chief among it. Currency is a universal value token: one thing exchangeable for anything anyone wants. It collapses the impossible problem of barter into a simple two-step. Without a universal token, the allocation machine can’t run.

Borrowing Neal Stephenson’s framing, the modern System of the World bundles three pillars: governance (rule of law, property, trustworthy contracts), the free-market economy (the allocator above), and engineering (the steam engine and everything downstream of it). The governance pillar is doing subtler work than it looks: the framers of a constitution couldn’t have known that software engineering would one day matter — the point was never to anticipate specific innovations, but to build something flexible enough to nurture whatever the frontier produced next. This bundle replaced what came before because it was better at distributing people to the machine. The result was a phase change: pre-modern systems grew at fractions of a percent per year; the new one took output per person to one to two percent a year, compounding, sustained for 250 years.^{SourceFor most of history, output per person barely moved — long-run growth ran in fractions of a percent. The Maddison Project’s reconstruction of the world economy shows sustained growth only begins with industrialization, then compounds for two centuries. Our World in Data · Maddison ProjectBackground · TheoryOded Galor’s Unified Growth Theory formalizes exactly this phase change — a single framework spanning the Malthusian epoch, the escape from it, fertility decline, and the modern era of sustained per-capita growth. It was developed specifically because earlier growth models (Solow, endogenous) could explain growth within the modern regime but not the transition into it. The transition is the thesis of this essay. Overview}

What that rate actually is

Here’s the reinterpretation everything hangs on.

That per-person growth rate — call it ~1.5% a year — is, in effect, the fraction of the population that productivity gains free up every year to go do something else.^{SourceRobert Solow’s foundational growth-accounting study concluded that technological progress accounts for roughly 80% of the long-term rise in U.S. per capita income, with capital investment explaining only the remaining 20%. John Kendrick’s later work reached the same conclusion: productivity, not factor accumulation, is what compounds. Solow residual · overview} Automation reduces the human labor content of existing goods; the freed-up capacity gets reallocated — most of it to making things we didn’t used to have, some to the frontier, some back into automation itself.^{SourceRobert Gordon, The Rise and Fall of American Growth (2016): roughly 60% of U.S. consumer spending in 2013 went on goods and services that did not exist in 1869. The freed-up capacity didn’t just sit there — it built entire categories that hadn’t existed a century earlier. Princeton UP overview}

That last loop is the engine: by freeing up labor, the system reinvests a slice of it into freeing up more labor next year. That’s why it compounds instead of running down.

SAME NUMBER — TWO FACES OF ONE 1.5% +1.5% richer / person the level — Chart ① = 1.5% of labor freed the same number — this picture

OLD WAY AUTOMATION NEW WAY

5 people make it AUTOMATE + capital · energy ③ inside the gear ⌕ 1 person · same widget 4 freed hands something NEW (didn't exist in 1990) Net unemployment ≈ 0 ⑥ → the gap opens rehire = release · joblessness is only the GAP between the two rates, and for 250 years that gap was essentially zero.

It is worth being precise about what’s doing the work, because it matters later. Engineering, strictly in our terms, is applying knowledge to capital to automate — and not every engineer is doing that; many carry the title but are working inside the machine, keeping the system running rather than advancing it. And conversely, not all of the engineering gets done by engineers. And what gets automated is dictated by the market. Engineering effort is scarce and expensive — it carries a high opportunity cost — so a business spends it where the payback is best, or a competitor that allocated better takes its place (and the misallocating firm’s engineers get reallocated when it folds). That payback filter is why automation advances along the highest-leverage front first.

The rare, compounding case is automation that lowers the cost of automation itself — a classic example being CADExample · AutomationBefore computer-aided design, technical drawings were produced by rooms of draftsmen — skilled cogs who hand-drafted engineers’ designs onto paper. Software engineers built CAD, and that entire role collapsed into a tool. But the job it eliminated isn’t the point: CAD lowered the cost of the next automation, the compounding kind that makes automating cheaper across the board.And CAD itself was enabled by the frontier — it wasn’t possible until computers had the graphics power and the user interfaces to drive it. Every innovation builds on the ones beneath it; that stacking is a large part of why the system compounds at all..

The mechanism is unemployment — and that’s not a bug. The “freeing up” is the labor market continuously moving people out of jobs automation displaced and into jobs that didn’t exist last year. Economists already have all the pieces — Schumpeter’s creative destruction^{Background · ConceptJoseph Schumpeter popularized creative destruction in Capitalism, Socialism and Democracy (1942) — growth working by continuously dismantling old economic structures to build new ones. He also coined technological unemployment for the displacement this produces. Overview}, the natural-rate decomposition into frictional + structural unemployment, the job-reallocation literature showing that churn from less- to more-productive firms is where long-term productivity gains come from.^{SourceDavis & Haltiwanger, Gross Job Creation, Gross Job Destruction, and Employment Reallocation (NBER, 1992) — the foundational measurement of these flows. Their finding: reallocation from less- to more-productive plants drives a major share of long-term productivity gains. nber.org/papers/w3728} Most people don’t even know the pieces exist. Assemble them and you get one picture: the “healthy” rate of unemployment is the throughput of reallocation, not a defect to be eliminated. The System of the World has had 250 years to tune its schools, safety nets, and markets to digest exactly this rate.

But the freeing up often does not look like layoffs at all. The plant-level data^{SourceCounterintuitively, plants that automate tend to add workers: a 1% rise in automation is associated with roughly +0.2% employment in the near term and +0.4% after ten years — freed capacity reabsorbed as growth, not layoffs. Aghion, Antonin, Bunel & Jaravel, The Direct and Indirect Effects of Automation on Employment. Paper (PDF)} shows most of the churn behind productivity gains happens within growing firms — the freed-up capacity gets reabsorbed as expansion, so the business does more per person rather than the same with fewer. Displacement is the cleanest case to reason about, but it’s only one channel; the more common one is the same headcount producing more, which frees labor in aggregate without anyone visibly losing a job. Either way the capacity comes loose — and either way it has to find a new role.

The new gear: AI automates the automation

Now add a new component to the machine — one that operates on the automation layer itself.

Set aside the strong, still-debatable claim that AI will soon do frontier science. Here’s the claim that holds regardless: AI has already reached the engineering level. It can build automation. And building automation is precisely the activity that produces that freed-up rate in the first place.

So we are now automating the automation. If engineering effort is what generates that rate, and AI is now multiplying engineering effort, then the rate itself starts to grow. At minimum this changes the exponent on the curve. More likely it applies a compounding rate to the compounding rate — a rate on the rate.

My back-of-envelope shape (and I want to be clear these are my estimates, not established figures): maybe the freed-up rate itself grows ~25% a year, probably along an S-curve. Year one adds less than a single percent — barely noticeable. But each year’s increment is bigger than the last.

Why this breaks the system: the absorption ceiling

The disruption isn’t the technology. It’s the mismatch between the new displacement rate and the rate the system is tuned to absorb.

It’s not a light switch — it’s a rising ceiling. The honest question isn’t “when does it break” but “at what reallocation rate does it break, and how fast do we get there?”

• 4% / year? Maybe still absorbable — inside the band past adjustments have handled.
• 8% / year? This is the bet. Sustained 8% almost certainly breaks the system. Markets can’t clear labor that fast; schools can’t retrain people that fast; safety nets weren’t built for that throughput.
• 15%+? Past the point where the existing institutional answers mean anything.

Underneath the institutional bottlenecks sits a sharper limit — the Schumpeterian one. The System works because creative destruction generates new profitable categories at roughly the rate productivity dismantles old ones. Today’s ~1.5% net displacement is met by roughly equal new-niche creation, and the market stays stable. If AI pushes destruction to 4–6% a year while the rate at which genuinely new profitable activity emerges stays near today’s, the difference is permanent labor-market exit, accumulating every year — even with perfect retraining and zero search friction. The institutional bottlenecks (markets clearing, schools retraining, safety nets absorbing) are the symptoms. The binding constraint is the new-new emergence rate: profitable new categories aren’t slow to find, they don’t yet exist.

This is already partly running, hidden in the wrong statistic. US Labor Force Participation has slid from ~67% (2000) toward ~62%; people who give up looking stop counting as “unemployed.” The destruction-creation gap shows up in participation and quality declines, not the headline unemployment number.

The Limit Case

Take it to the logical limit — which is what this project is for.

Is there any fundamental reason capital + energy cannot do everything currently classed as traditionally productive? No. Food, goods, shelter, logistics, manufacturing, most administration. By the energy-as-currency derivation, at the limit, production is capital + energy. So the long-term share of human labor required for productive work approaches zero.

That leaves at most two categories of human role:

• Frontier work — genuine discovery at the boundary of what’s been thought. Be honest about the count: this is a small slice today — low single digits of the workforce — and even counting all non-automatable work generously, you don’t credibly clear ~20% of working-age people. And if the strong AI claim lands, AI eventually does the frontier work too.
• Preference-protected roles — care, performance, mentorship, intimacy. These survive not because they can’t be automated but because we prefer humans for them. They’re real and meaningful — and typically low-paid or unpaid, and they don’t scale with productivity the way machinery does.

The Limit Case, fully stated: In the limit, capital + energy do everything traditionally productive. Even optimistically, no more than ~20% of work resists automation — and that is the ceiling on all non-automatable work, not the frontier alone, which is a small slice today and which AI may eventually take too. Preference-protected roles exist but don’t scale economically. The labor → income → consumption mechanism has no long-term equilibrium that includes everyone. A different allocation mechanism is required — and the current System of the World does not contain one.

This is the thesis the whole project points at. Everything else on this site is either a piece of the derivation or a sketch of what the successor mechanism might be.

The stake: build the next system before the current one breaks

If the absorption ceiling is somewhere below 8% — and the bet is that it is — the modern System of the World will break. Not philosophically. Materially. The institutions that allocate people, distribute value, and absorb slack will stop functioning at the rate they’re being asked to. The only question is what comes next, and how it arrives.

Two paths:

• By design. We notice the curve early, name the absorption problem honestly, and start building the next system before the current one fails — new mechanisms for allocating people to meaningful activity, and for distributing surplus when productivity decouples from wages.
• By revolution. We don’t, and failure forces a successor into existence the hard way — through the structural violence that has historically accompanied transitions of this magnitude. The Industrial Revolution’s early phases weren’t bloodless; the French Revolution is in that lineage.

It would be naïve to pretend the designed path is guaranteed. The violent path may be unavoidable. But “maybe unavoidable” is not “not worth attempting.” Naming the problem clearly, mapping the trajectory, and building the conceptual tools the next allocation machine will need — that is what this project is for.

Questions for the Reader

• What rate of annual job displacement do you think your own industry could absorb before it stopped functioning? 2%? 8%?
• If most “productive” work no longer needs humans, what should a person’s claim on the surplus be based on?
• Which of the two paths — design or revolution — do you actually expect? And does your answer change what you think we should be doing now?

---

This is the spine. The other essays are limbs hanging off it: energy as fundamental currency is where production grounds out; the baseball analogy is what happens to labor; the desire/demand distinction is the engagement mechanism; regulation as friction is governance as a deliberate distortion on the allocator; the exponential blind spot is why the early increments get ignored until it’s too late.