The Automation of Automation: A Ten-Thousand-Year Mountain of Humans and Machines

msn — Fri, 25 Sep 2026 21:36:58 +0000

TL;DR

Human civilization keeps handing work off, from individuals to other people, to organizations, to tools, to machines, and now to computational systems. Read as a simplified narrative, that's a slow drift from direct execution toward organization, mechanization, and automation.

This piece tells that story as a mountain climb: one compressed image standing in for centuries of technological and social change. The climb isn't a claim about a single, predetermined historical path. It's a device for holding onto one question: as technology and social organization grew more complex, how much human labor did it actually take to keep civilization running?

World population has multiplied roughly two-thousandfold since 10,000 BCE. Annual working hours in industrial economies have fallen by nearly half since 1870. Modern job classifications catalog thousands of distinct occupations and tens of thousands of tasks. None of that adds up to a single law of declining need for humans, but together, the numbers are worth taking seriously.

The model separates five things that are usually lumped together: jobs, occupations, tasks, skills, and required labor volume. The question isn't how many jobs exist per person. It's how much functional diversity a civilization can sustain per unit of human labor it actually needs.

AI sits at the far end of the story. Systems that write code, design, analyze, and decide represent a different kind of technological wave, the open question is whether they've entered a genuine, independent loop of designing their own successors, or whether that's still mostly ahead of us.

So the "summit" here is a scenario built for asking a question, not a place history has confirmed we've reached. The real question isn't whether the future is fixed, it's what happens to the relationship between production, labor, income, ownership, and demand once design itself becomes something a machine can do.

Preamble: Nobody Drew This Mountain in Advance

Nobody sketched this mountain ten thousand years ago. Early societies weren't working toward the factory, the computer, or AI.

Looking back with enough distance, though, a pattern shows up in the wreckage of a thousand unrelated decisions: people keep splitting work away from direct, individual execution and handing it to someone else, another person, an organization, a tool, an energy source, a machine, a computer.

This piece follows that pattern as a mountain.

Real technological history is a tangle, not a line, branches that stall out, reappear decades later, or race ahead of everything around them while unrelated domains barely move. The mountain compresses that tangle into one legible shape, along a single axis worth tracking:

As technology, organization, and specialization increase, how much of civilization's function can a given amount of human labor sustain?

Part One: The Plain

Wind moves across a grassland split by a river. A few dozen people live along it, a sketch, not a census.

Roles here aren't sharply defined. The same person hunts one day, gathers the next, shapes a tool after that, watches children, helps raise a shelter. None of this maps onto a modern "job", we can't break Paleolithic life down into occupations, tasks, and skills the way ISCO or O*NET break down a modern economy, because the data simply doesn't exist at that resolution. The honest move is to call this a period of low role differentiation, not to assign it a job count it can't support.

One member of the group turns out to be sharper at tracking animals. Another has a knack for knapping stone. The group starts leaning on the difference, sending people toward what they're good at, an early, informal division of labor, though certainly not the first one in human history; division of labor is older and messier than any single scene can capture.

What the scene actually buys the model is one idea:

A person can live off someone else's labor without doing every task their own survival requires.

Analytical Branch 1, Automation, or Something Else?

This piece uses a working definition: an outcome that once required someone's direct involvement becomes more automatic for that person once it can be reached with less of their direct involvement.

Under that definition, division of labor splits in two depending on whose side you're standing on. For the person who wants meat but doesn't hunt, part of the food supply has been handed off. For the hunter, nothing has automated at all, they're still doing the hunting, every time.

Division of labor, then, is a transfer of execution from one person to another, not automation in the technical sense. It functions as an organizational layer that sets the stage for automation later, without being automation itself.

Picture a river splitting into branches: same water, several specialized channels.

Part Two: The Market

Generations pass. Small groups grow, unevenly and on no shared timeline, into settlements, then towns.

As social networks widen, people trade goods and services they didn't make themselves. A hunter needs a tool; a toolmaker needs food. Custom, obligation, and informal contract settle the question of whose is whose long before anything like formal property law exists.

Ownership functions here as one mechanism among several that make more complex trade possible, not the sole cause of markets, and not something that arrived on a straight line from village to town to nation. Each of those transitions has its own separate history.

Analytical Branch 2, The Market as Coordinator

Friedrich Hayek's "The Use of Knowledge in Society" argues that price systems coordinate scattered, local information without needing a central authority to hold all of it at once.

That's a specific claim about coordination, it doesn't make the market "history's first automation." Automation, in this piece's terms, is the automatic execution of a defined process; a market is a coordination system. Hayek himself split orders into two kinds: those someone designs (a factory, an army) and those that simply grow (language, custom, the market itself). The market sits in the second category, a network of scattered decisions aligning on their own, not a machine one designer built for one task.

Call the market social coordination technology. The question worth tracking from here is when technology stops coordinating people and starts replacing their decisions or their execution outright.

Part Three: The First Machine

Centuries pass. A windmill turns wind into the motion that grinds grain; a watermill does the same with a river's current. Someone still has to feed the mill, adjust it, repair it, move the grain, the mill automates the grinding, not the surrounding work.

In 1804, Joseph Marie Jacquard introduced a loom controlled by punched cards encoding the weave pattern, an early split between an instruction and the hand that carries it out. The loom still needed a weaver to load it, thread it, and keep it running.

Analytical Branch 3, Older Than the Industrial Revolution

Automation didn't start with steam power. Watermills, windmills, mechanical clocks, and Al-Jazari's thirteenth-century automata all ran sequences of operation without moment-to-moment human input, centuries before anyone built a factory.

Jacquard's loom is a genuine bridge between encoded information and mechanical execution, calling it "the first stored program" stretches the term further than its own history supports; call it an important antecedent instead, and leave the "first" to the people who invented stored-program computing a century and a half later.

The Luddites weren't simply afraid of machines, their grievances ran through wages, working conditions, and control over a trade, not blind hostility to mechanization. What actually distinguished the Industrial Revolution from everything before it wasn't the existence of automation. It was scale: a steam engine, unlike a windmill, could run a train, a factory, or a ship, the same source of power, dozens of unrelated applications. That's the shift from local, single-purpose automation to general-purpose infrastructure.

Part Four: The Brain That Stayed Human

Factories grow. Muscle work shifts onto machines and new energy sources, but designing, maintaining, and running those machines stays a human job. People design the machines, define the processes, decide where to point the technology.

Plenty of calculation and control had already been mechanized before computers existed, so this isn't a claim that human judgment was uniquely, permanently untouched by automation. What's true is narrower: designing and defining new systems stayed heavily dependent on human capability for a very long time. And employment history isn't "machines arrived, then more jobs appeared", technology shrinks some jobs, reshapes others, and creates demand for work that didn't exist before, often all three at once.

Analytical Branch 4, Comparative Advantage

In 1930, Keynes coined "technological unemployment" for a specific worry: that labor-saving discoveries might outrun the discovery of new uses for labor. He framed it as a long-run transition problem, not a flat prediction that jobs would run out for good.

The Depression's unemployment came from a financial crisis and a demand collapse layered on top of a production collapse, technology wasn't the driver, though the full historical accounting of causes is more contested than the popular version admits.

Nineteenth-century British bootmaking is better documented: mechanization eliminated roughly 152,000 traditional shoemaking jobs and created about 144,000 new ones requiring different skills. Much of that shift ran through what researchers call erosion of entry, younger workers simply stopped entering a declining trade rather than being laid out of it. Whether that's genuinely less painful than a direct layoff, or the same pain spread thinner and made less visible, isn't something this piece can settle; it's an open question, not a comfort.

One case doesn't make a law. Comparative advantage is the actual mechanism at work here: trade and specialization stay worthwhile even when one side is better at nearly everything, because what matters is relative cost, not absolute skill. It doesn't guarantee a human niche survives forever, if a future technology can cover an extremely wide range of tasks at very low cost, the underlying structure of comparative advantage shifts too. Treat it as the mechanism that has explained employment's persistence so far, not a promise about what comes next.

Part Five: When the Books Stopped Needing Hands

Computers move into offices mid-century. Calculations that took a room full of clerks or a mechanical calculator now run in seconds.

Accounting and finance didn't disappear, cheaper, faster computation opened up data analysis, more sophisticated auditing, financial management, and information systems that hadn't existed before. Work that once ate a person's whole day moved to a machine; human effort moved up a level.

Analytical Branch 5, The Automation Paradox

Automation cuts both ways on employment: it substitutes for labor in some tasks and complements it in others, and cheaper goods tend to sell more of themselves, which can raise demand for the labor that remains.

"Automation increases employment," stated with no period, industry, or technology attached, is too broad to mean much, the actual effect runs through demand elasticity, wages, investment, policy, and trade. The Acemoglu-Restrepo task-based framework gives this a cleaner shape: automation shifts some tasks from labor to capital while technology simultaneously creates new tasks that may need labor to fill them.

So the mechanism isn't:

Technology → job destroyed

It's:

Technology → the task set changes → the human/machine split shifts → productivity and demand shift → employment's structure shifts.

This is where the mountain starts doing real work.

Part Six: Halfway Up

Work has moved, over time, from individuals to other people, to organizations, to tools, energy sources, machines, and computers, a rough sequence, not a strict one; real history skipped steps and doubled back constantly.

Population has grown roughly two-thousandfold since 10,000 BCE. Annual working hours in several industrial economies have nearly halved since 1870. Put those two facts side by side and you get something worth modeling: more people, and less time spent working, at the same time. Neither fact alone proves a causal link between automation and the population a civilization needs, together, they're the actual puzzle this piece is built to examine.

Each step up the mountain is a jump in organizational or technological capacity, not automatically more humans needed, not automatically fewer. Sometimes technology cuts demand for labor; sometimes higher productivity expands both production and demand; sometimes new tasks appear; sometimes old ones just change shape. Often several of these happen from the same technology at once.

Every technological wave redraws the line between human and non-human work.

Part Seven: The Summit

Call this point the summit, not because it's a confirmed historical fact, but because it's where technology starts reaching past task execution into design, analysis, planning, and the creation of new tools.

Current AI systems generate code, produce text and designs, analyze complex problems, and use external tools with limited supervision, real capabilities that have expanded fast in the last few years. That's different from "the machine now independently builds its own successor," which depends on exactly how much independence, what kind of tool, how much human oversight, and what "designing the next generation" even means. Call this the automation-of-automation scenario, not a settled fact:

Part of what used to depend on humans for designing and building new automation may itself be becoming automatable.

Analytical Branch 6, Why This Point Might Actually Differ

Past waves of automation took over physical or computational execution while leaving design, goal-setting, and error correction to people, even though feedback systems and algorithmic control have their own long history of automating parts of that decision layer.

Frame the difference through three layers: energy (supplying and converting power), structure (executing tasks, physical or computational), and control (planning, deciding, designing, defining what the tasks even are). Most historical innovation lived in the energy and structural layers; computers automated a large share of computation itself. What AI potentially does is push into the control layer directly, reasonable grounds for taking this wave seriously as different in kind, though it doesn't by itself prove a full self-improvement loop is running.

Anton Korinek and others working on transformative-AI economics ask a sharper version of the same question: if the range of machine-doable tasks gets wide enough, what happens to labor's comparative advantage? In the extreme case, a machine that can do every human task at very low cost, wage structure and income allocation could shift fundamentally. That's a theoretical scenario, not a description of today's economy, which is exactly why the summit is a place for examining the scenario, not a claim that we've arrived at it.

Part Eight: The Descent Isn't Symmetric

Nothing requires the way down to mirror the way up. Software ships in months; a physical robot needs hardware, supply chains, energy infrastructure, safety testing, and capital, years, not months. Cognitive and physical tasks have no reason to automate at the same speed.

Analytical Branch 7, Moravec's Paradox

High-level reasoning has turned out to be cheap for machines; picking up a coffee cup or folding laundry has turned out to be hard. That asymmetry, Moravec's Paradox, has shaped robotics and AI for decades.

A newer economic model builds the asymmetry directly into automation forecasts. But treat any specific number attached to it, "a decade for cognition, a century for physical work", as a scenario parameter, not a law: it rides entirely on assumptions about compute growth, robotics costs, data, and algorithmic progress, any of which could move the estimate by decades in either direction.

The digital layer may move fast while the physical layer runs into a different, harder set of constraints.

That's a hypothesis to test, not a conclusion about the next hundred years. The "boiling frog" image sometimes attached to this kind of gradual, underreacted-to change is worth dropping, the biology behind it is fiction, and social change doesn't need a fake frog story to make the point that gradual shifts go unnoticed longer than sudden ones do.

Part Nine: Silent Factories

Picture the extreme case: factories running near-empty of people, software writing software, robots moving their own raw materials, AI managing the supply chain. Production costs fall hard.

One question doesn't go away:

If income comes mainly from economic participation, and the need for labor drops sharply, what keeps households able to buy anything at all?

Analytical Branch 8, The Effective-Demand Paradox

Effective demand is a standard macroeconomic concept: falling income and purchasing power drag demand down with them. In the US specifically, household consumption runs around 70% of GDP, a figure that belongs to the American economy, not a universal constant.

None of that proves automation collapses demand on its own. It does sketch a mechanism worth taking seriously: automation cuts production costs and labor demand together, labor income falls, purchasing power can follow, and falling demand cancels out part of the gain from higher output. Whether that mechanism actually closes into a feedback loop in a real economy depends on capital ownership, income distribution, prices, fiscal policy, and trade, none of which are fixed.

Martin Ford makes a forceful version of this case in Rise of the Robots, an influential, widely read argument, and a non-academic one; where it actually stands in the professional economics literature is a separate question this piece hasn't settled. More formal models run the same scenario several different ways, with automation raising output while reshaping distribution and demand in directions that aren't predetermined.

In a highly automated economy, what actually keeps production connected to purchasing power?

Part Ten: Three Valleys

Three paths lead down from the summit, scenarios for examining consequences, not predictions.

Valley One: The Abandoned Market. Capital ownership concentrates; redistribution stays weak. If household income still runs mostly through labor while labor demand falls, the gap between what gets produced and what people can buy widens. How badly depends on ownership structure, capital prices, and fiscal policy, a scenario of high concentration and low redistribution, not a forecast of collapse. Picture a lake full of water with every outlet channel blocked.

Valley Two: Institutionalized Redistribution. Society redistributes part of automation's returns through taxation, cash transfers, public services, or public ownership, without necessarily collectivizing capital itself. The policy menu includes Universal Basic Income, Negative Income Tax, Universal Basic Capital, wage insurance, the EITC, and retraining. None of them is a solution independent of cost, institutional feasibility, and the economic conditions they'd have to work in.

Valley Three: A Repeat of the Historical Pause. AI produces large effects, but institutions take decades to catch up, rising productivity, wage pressure on some groups, rising inequality along the way. Engels' Pause names the actual precedent: several decades of early British industrialization where productivity growth ran well ahead of real wages before the gap closed. A useful comparison, not a guarantee of a repeat. Picture a long winter that has to run its course before anything like a new equilibrium shows up.

Analytical Branch 9, Not Necessarily Socialism

Redistributing income and collectivizing the means of production are different things by definition, and no redistributive policy becomes socialist just by redistributing. Valley Two assumes public institutions take on a bigger role in distributing automated capital's returns, that's an expanded redistributive role, not a change in who owns the capital.

Analytical Branch 10, A Real Policy Menu

Research from Jacobs and Imas (2026), Economic Policy for AGI, scores eleven policy instruments, unemployment insurance, retraining, wage insurance, EITC, job guarantees, NIT, UBI, UBC, sovereign AI dividends, basic services, industrial policy, against different disruption scenarios. No single instrument wins across the board. Under broad displacement, NIT scores well. Where capital ownership itself becomes the central issue, UBC does better, since it expands ownership directly rather than just transferring income.

Valley Two, then, isn't one policy, it's a family of mechanisms aimed at the same handful of goals: keeping purchasing power intact, spreading risk, and broadening who actually holds a claim on capital's returns.

Part Eleven: The Base of the Mountain

Nobody knows what the base of this mountain looks like. Population, technology, working hours, and job structures have all changed continuously, that alone doesn't prove history is heading toward the total elimination of human labor.

The model asks a narrower, testable question:

How much human labor does a given level of civilizational function actually require, and how does that number move as technology changes?

More task coverage by technology can lower the labor a given output requires; new tasks and new needs can offset part of that; and if technology reaches into domains that used to depend on distinctly human ability, the historical link between productivity and employment can shift along with it. These are open questions, not foregone conclusions.

The mountain is a model, not a prophecy. The climb stands for growing organizational and technological capacity; the height, formally, is the human labor required to sustain a given set of functions; the descent is a scenario for testing whether technology can shrink that number. The model doesn't hand you the future, it gives you a way to measure and compare the futures on offer.

Appendix: A Mathematical Definition of the Mountain

If the mountain isn't just a metaphor, something has to stand in for its height. That means separating the data we actually have from the variables the model needs.

The Solid Data

World population. Roughly 4 million people around 10,000 BCE; about 595 million in 1700; 983 million in 1800; 1.6 billion around 1900; past 8 billion in 2022. The UN's World Population Prospects 2024 central scenario projects a peak near 10.3 billion in the mid-2080s, settling to about 10.2 billion by 2100. This is the most solid series in the whole model, ancient and prehistoric estimates carry far more uncertainty than the modern numbers do.

Occupational diversity, and why the count depends on who's counting. ISCO-08 organizes the global labor market into 10 major groups, 43 sub-major groups, 130 minor groups, and 436 unit groups. O*NET, covering the US alone, runs far finer, cataloging over 19,000 distinct task statements. Neither number is "every job in the world", each is the resolution of one particular classification. Split one occupation into ten tasks and the count jumps without the economy's real diversity moving at all, which is exactly why this model tracks weighted diversity instead of raw category counts.

Working hours. German annual working hours fell from roughly 3,284 in 1870 to about 1,371 in 2017; British hours fell from about 2,984 to 1,670 over the same stretch. Solid data, but "work" in a modern labor-hours dataset and "work" in a foraging or subsistence economy aren't the same unit, so pre-industrial hours belong in the model as an uncertain variable, not a fixed number pulled from the industrial series.

The Model

Earlier drafts collapsed job count, task diversity, and required labor into one variable. This version keeps them apart:

$$
\text{Job} \rightarrow \text{Occupation} \rightarrow \text{Task} \rightarrow \text{Skill}
$$

Required human labor. Let $N_T(t)$ be the number of distinct tasks at time $t$, $L_i(t)$ the reference labor each requires, and $a_i(t)$ the share of task $i$ a machine performs:

$$
L_H(t) = \sum_{i=1}^{N_T(t)} L_i(t)\,[1-a_i(t)]
$$

$L_H$ is a model variable, not an observed historical series.

Minimum human population. If each person supplies $W(t)$ effective work-hours a year:

$$
H_{\text{labor}}(t) = \frac{L_H(t)}{W(t)}
$$

A civilization can need more people than the raw labor-hour total implies, to keep critical skills, coordination, and resilience against shocks intact:

$$
H_{\min}(t) = \max\left[H_{\text{labor}}(t),\, H_{\text{skill}}(t),\, H_{\text{coordination}}(t),\, H_{\text{resilience}}(t)\right]
$$

This is the mountain's height. It isn't directly observable in the historical record, estimable in part for recent decades, a latent variable with wide uncertainty for everything older.

Functional diversity. Two economies with 100 tasks each look identical by raw count even if one concentrates 90% of its labor on three of them and the other spreads evenly across all 100, which is why the model uses the Shannon index instead:

$$
D_{\text{eff}}(t) = \exp\left(-\sum_i p_i \ln p_i\right), \qquad p_i = \frac{L_i}{\sum_j L_j}
$$

Tasks with a tiny labor share carry proportionally little weight. This is well-defined mathematically; its values for ancient periods still aren't observed data.

Civilizational specialization density.

$$
S(t) = \frac{D_{\text{eff}}(t)}{H_{\min}(t)}
$$

This term belongs to this model specifically, it isn't a standard index from the economics literature. Rising $S$ means more functional diversity per unit of required labor; falling $S$ means the opposite. Right now it's a proposed measurement framework, not an established historical indicator.

On Bettencourt, and its limits. Bettencourt et al. (2014) found professional diversity scaling sublinearly with city population across American metro areas: $D \propto H^{0.84}$. That's a real result, scoped tightly to contemporary US cities and one specific measurement method. It does not support:

$$
D_{\text{human history}} \propto H_{\text{world population}}^{0.84}
$$

Use the 0.84 exponent as one data point suggesting sublinear size-diversity relationships can exist, not as a universal historical law.

What still isn't measured. We have archaeological and historical evidence that division of labor, specialization, and craft existed in ancient societies. We don't have a defensible way to compress that evidence into "civilization Y had exactly 37 kinds of jobs in year X." Future versions of this model should carry prehistoric $D_{\text{eff}}$ as a latent variable with an explicit uncertainty range, not a plotted point that looks more precise than the evidence behind it.

The defensible conclusion. World population has grown enormously, especially since industrialization. Working hours have fallen substantially in the economies with long time series. Modern classifications record enormous occupational and task diversity. Professional diversity scales sublinearly with city size in the one dataset that's actually measured it, and that relationship doesn't generalize to all of human history without more evidence. Job count, task diversity, and skill diversity are three different quantities, and none of them alone determines how much human labor a civilization actually needs. For the modern era, the mountain is a testable model. For most of pre-industrial history, it's a hypothesis waiting on evidence that may never arrive with the same precision as the population data.

Closing

This piece doesn't claim history ran a fixed course from the cave to artificial intelligence, or that division of labor, markets, machines, computers, and AI were scripted stages in some larger plan. It doesn't claim the future ends in the total elimination of human labor either.

What it claims is narrower: technological history, from one angle, is the history of shifting task execution among people, organizations, and non-human systems, through division of labor, then new energy sources, then mechanization, then computation, and now, increasingly, through systems that can take on cognitive work, planning, and design. That last shift is still in progress. If it continues, the relationship between human labor, production, ownership, income, and demand moves with it, how far, is still an open empirical question.

The mountain isn't a prophecy. It's a question:

If civilization can sustain more function with fewer humans, where does the next equilibrium sit?

And once machines start designing a growing share of the next tools themselves, "which job replaces the old one" may not be the only question worth asking anymore. The deeper one might be:

In a civilization whose output no longer depends heavily on the volume of human labor, what do humans actually do in running it?

Today's data can't answer that. It can be measured, modeled, and tested against whatever comes next, which is the point where a mountain metaphor stops being a story and starts being a research program.

References

Keynes, J. M. (1930). Economic Possibilities for our Grandchildren.
Anslow, L. "Robots have been about to take all the jobs for more than 200 years."
MIT Technology Review (2024). "People are worried that AI will take everyone's jobs. We've been here before."
Smith, A. (1776). An Inquiry into the Nature and Causes of the Wealth of Nations.
Hayek, F. A. (1945). "The Use of Knowledge in Society." American Economic Review.
"The Paradox of Automation: How Labor-Saving Technology Can Create More Jobs." EconomicsOnline.
Korinek, A. Economic Policy Challenges in the Age of AI. NBER Working Paper No. 32980.
Korinek, A. & Juelfs, M. Preparing for the (Non-Existent?) Future of Work.
"Moravec's Paradox and Restrepo's Model: Limits of AGI Automation in Growth." arXiv:2509.24466.
"Moravec's Paradox: Why Robotics Lags Behind AI."
"The AI Paradox: How Automation Eats Its Own Tail."
"When Robots Take Your Job, Will the Government Pay You Instead?"
Ford, M. Rise of the Robots: How Artificial Intelligence Will Transform Everything.
Allen, R. C. Work on the concept of Engels' Pause.
Bettencourt, L. et al. (2014). "Professional diversity and the productivity of cities." Scientific Reports, 4, 5393.
Acemoglu, D. & Restrepo, P. (2019). "Automation and New Tasks: How Technology Displaces and Reinstates Labor." Journal of Economic Perspectives.
Our World in Data. "How has world population growth changed over time?"
United Nations. World Population Prospects 2024: Summary of Results.
International Labour Organization. ISCO-08.
O*NET OnLine. Occupation and task databases.
Our World in Data. Annual working hours per worker.
Huberman, M. & Minns, C. (2007). Historical working-hours research.
"Technological unemployment in Victorian Britain: a tasks based approach." LSE, 2025.
Jacobs, J. & Imas, A. Economic Policy for AGI, 2026.

DEV Community: msn