APoV 006Meta-work, and Production Relations in the AI Era

On a flight home to the San Francisco Bay Area after three weeks on the East Coast, I read a new study from Anthropic, “How Employment Is Changing at Anthropic,” examining the company’s own workforce. Surveyed in August, Anthropic’s employees expected on average that by 2030, AI would be able to handle 75 percent of their current work.

The year 2030 brought to mind John Maynard Keynes. In his 1930 essay “Economic Possibilities for our Grandchildren,” the British economist predicted that within his own generation’s lifetime, agriculture, mining, and manufacturing would require just a quarter of the human effort to which the world had grown accustomed. Within a century, by 2030, humanity’s economic problem might be solved, “or be at least within sight of solution,” leaving his grandchildren to work fifteen-hour weeks. Now, looking toward that very same year, Anthropic’s staff expect AI to absorb three-quarters of their work, leaving a quarter in human hands. A century apart, two visions of technological abundance converge on 2030, and on remarkably similar arithmetic.

Yet no one at Anthropic is easing up. The report gives what keeps employees occupied a name, meta-work. Put simply, meta-work is the labor of establishing the conditions under which a task gets done, including writing documentation, configuring environments, refining workflows, and setting up oversight, while handing direct execution to AI. At Anthropic, that means building reusable skills for Claude, drafting plans and reviews for other agents, assembling dashboards to monitor agent pipelines, and codifying working methods into standing documents. Between fall 2025 and spring 2026, meta-work’s share of the average employee’s interactive Claude sessions more than doubled, rising from 10 percent to 22 percent.

Reading the term, I immediately recognized my own story. On a typical workday, the time I spend writing code by hand has fallen to nearly zero. Most of my day goes to articulating my reasoning and plans clearly enough to serve as means of production for my agents. Alongside that, I build dedicated environments around recurring bottlenecks on the job so that AI can resolve problems more reliably and efficiently. That shift reflects both the practical demands of the work and the way my company’s incentives are structured. Outside the office, the pattern holds. For this column, APoV, the editorial frameworks, style guides, and rendering scripts I have written for AI now run longer than the essays themselves.

First Sharpen the Tools

Is meta-work new? It is not, nor was Anthropic the first to use it. In workplace research, an influential early formulation appeared in a 2004 study by Victor González and Gloria Mark at the University of California, Irvine. Shadowing office information workers as they switched among multiple projects through the day, the researchers noticed how much time went to activities tied to no single project, effort spent purely to keep all the others organized. They called these activities “metawork.” Two decades earlier, in 1985, the sociologist Anselm Strauss had analyzed a closely related phenomenon under the name “articulation work.” Once the division of labor fractures a job into parts, someone must stitch the parts back together, sequence tasks, track progress, and absorb the unexpected. Because such effort rarely leaves behind a tangible artifact, it is easily mistaken for idleness. In online productivity circles, “meta-work” has even acquired a faintly derisive edge as shorthand for tinkering with tools and workflows to procrastinate on the task itself.

The practice is far older than any modern label. In 1776, in the opening chapter of The Wealth of Nations, Adam Smith recounted a story while discussing the division of labor: on the earliest steam engines, a boy had been hired solely to open and shut the valve between the boiler and the cylinder with every stroke of the piston. Eager to play with his friends, one boy tied a string from the valve’s handle to another moving part of the engine, allowing the valve to open and close on its own. One of the greatest improvements to the machine since its invention, Smith wrote, was “the discovery of a boy who wanted to save his own labour.” Historians treat the tale as more legend than fact, yet its logic endures. The boy stopped operating the valve and redesigned how the valve was operated. That is meta-work in embryonic form.

In the very next paragraph, Smith noted that while many mechanical improvements came from machine-makers, others came from those “called philosophers or men of speculation, whose trade it is not to do any thing, but to observe every thing,” who, for that reason, are often able to combine the powers of the most distant and dissimilar objects. As society advances, he observed, such speculation becomes, like any other trade, the principal occupation of a distinct class of citizens.

In the late nineteenth century, the Austrian economist Eugen von Böhm-Bawerk illustrated a parallel logic through fishing. Catching fish bare-handed in tidal pools is no match for taking the time to weave a net and build a boat before putting out to sea. “That roundabout methods lead to greater results than direct methods,” he wrote in The Positive Theory of Capital, “is one of the most important and fundamental propositions in the whole theory of production.” On the days spent weaving, no fish are landed; once the net is finished, there is no going back to fishing by hand.

By 1903, Frederick Winslow Taylor, the father of scientific management, had codified that impulse into an industrial system. In Shop Management, he argued that “all possible brain work should be removed from the shop” and concentrated in a planning department, which would specify each worker’s task in advance and supply the “standardized conditions” and tools needed to execute it reliably. Seven decades later, in Labor and Monopoly Capital, the American political economist Harry Braverman mounted a searching critique of Taylor’s system, arguing that management had monopolized conception and reduced workers to pure execution.

More than two millennia before Smith or Taylor, Chinese classical thought had already distilled the principle into an aphorism: “A craftsman who wishes to do his work well must first sharpen his tools.” Recorded in The Analects, the line was Confucius’s metaphor when his disciple Zigong asked how to cultivate virtue. From Confucius’s whetstone to Smith’s piece of string, Böhm-Bawerk’s fishing net, and Taylor’s planning room, meta-work has worn many names and guises. Its function has remained constant, getting the conditions right first and then setting the work in motion.

Meta-Work in the Age of AI

Why, then, is so much human labor in the age of AI converging on meta-work? The environments I configure for AI are Taylor’s “standardized conditions”; the specifications and prompts I draft for agents are the written instruction cards his planning department once handed to machinists. Meta-work is roundabout production in cognitive labor; specifications, skills, and playbooks are the nets we weave and the string tied to the valve.

The value of meta-work depends largely on the capability of the underlying AI. The more problems AI can solve on its own, the greater the return on time spent on meta-work. Whether a net is worth the days spent weaving depends on the size of the catch it can haul in; a small, loose mesh scarcely beats wading into the shallows, whereas a wide net dragged through deeper waters repays the upfront effort many times over. An agent’s capability determines the size of that catch. According to Anthropic’s report, combined hours worked by autonomous and interactive agents at the company surpassed human working hours for the first time in March 2026. By August, even counting only interactive sessions directed by individual employees, agents were logging six times as many working hours as their human handlers, assuming a forty-hour week per employee. A single hour of meta-work now mobilizes an order of magnitude more output than it did a year ago.

The rapid turnover of AI engineering jargon over the past three years tells the same story from another angle. First came prompt engineering, the craft of phrasing a single query just right. In June 2025, Tobi Lütke, the chief executive of Shopify, and Andrej Karpathy, a founding member of OpenAI, championed “context engineering,” shifting attention from an isolated prompt to the full slate of information supplied to a model. In February 2026, OpenAI published “Harness Engineering,” chronicling a team that shipped a large application with virtually no hand-written code by focusing instead on building environments and verification gates for agents. By June 2026, “loop engineering” had gained currency, advising developers to stop prompting coding agents turn by turn and instead design automated loops that prompt, verify, and re-run them. Beneath the shifting labels, every stage is a form of meta-work, each moving one rung higher up the ladder of abstraction. Today, standalone prompt engineering has largely faded because models infer intent well enough and a robust harness cushions an imperfect instruction. Plain language addressed to a more capable model now outperforms a meticulously tuned prompt fed to a weaker one.

Inside frontier AI labs, meta-work evolves in lockstep with model capability. Across the wider economy, however, most people do not operate at the research frontier. While reporting this piece, I spoke with friends working in technology, finance, and brokerage, as well as startup founders, to examine how their daily workflows actually run. Placed alongside my own process for producing APoV, their experiences pointed to a different driver. For most people, changes in meta-work have relatively little to do with breakthroughs in frontier models. Meta-work grows organically out of each person’s business problems and deepening command of existing AI tools.

APoV evolved the same way, step by step. At first, I simply passed finished drafts to an AI model to polish the prose. Next, I began codifying my stylistic preferences and editorial rules. That grew into an end-to-end playbook governing everything from topic selection to publication, and eventually into custom rendering scripts and a structured fact-checking protocol. Every layer was added because the column’s workflow demanded it or because I understood the tools better, and none was directly tied to any new model release.

The pattern recalls the electrification of American manufacturing a century ago. In a classic 1990 paper, the economic historian Paul David examined a famous puzzle: electric motors were available by the late nineteenth century, yet factory productivity barely budged for decades. Early adopters, it turned out, merely bolted electric motors onto existing overhead drive shafts, leaving multi-story factory buildings, belt transmissions, and production routines unchanged. Only in the 1920s, roughly forty years after the first central power stations opened, did manufacturers redesign factories around the logic of electricity, mounting an individual motor on every machine inside single-story plants laid out along the natural flow of materials. Only then did productivity surge. Replacing the power source without re-architecting the factory yielded almost nothing. The Stanford economist Erik Brynjolfsson and his colleagues later formalized this dynamic as the “productivity J-curve,” showing that general-purpose technologies pay off only after organizations make substantial, largely invisible investments in new processes, structures, and skills. In the age of AI, meta-work is that intangible investment, made one workflow at a time.

New Relations of Production

Thus far, meta-work looks like a straightforward story about productivity, with the same people, working the same hours, using AI to multiply their output. Yet whenever the forces of production shift, the ownership of productive assets and the relations of production eventually shift with them.

In a 1999 essay, the management theorist Peter Drucker observed that knowledge workers, unlike manual laborers, own their means of production, namely the knowledge between their ears, “a totally portable and enormous capital asset.” When a factory hand quits, the machinery stays on the shop floor; when a knowledge worker leaves, the scarcest asset walks out the door. That portability has long been the deepest divide between knowledge work and industrial labor, and the foundation of white-collar bargaining power.

Meta-work is profoundly changing that. At Anthropic, written standing instructions for agents grew from a few hundred in mid-2025 to 8,500 a year later. Once committed to writing, standing instruction after standing instruction, skill after skill, and playbook after playbook all remain inside corporate repositories; when an employee departs, the codified know-how stays behind, still directing fleets of agents. In the “Fragment on Machines” in his Grundrisse, Karl Marx used the English phrase “general intellect” to describe accumulated social knowledge once it has been absorbed into machinery as fixed capital. Industrialization embedded the artisan’s physical craft into iron and steam; meta-work embeds the professional’s judgment and institutional knowledge into software. “We can know more than we can tell,” the philosopher Michael Polanyi famously wrote in 1966. Piece by piece, meta-work records tacit knowledge as explicit text. The moment that judgment settles into digital form, intellectual capital migrates out of the worker’s head and into the company’s systems.

The only difference lies in who performs the transfer. In the industrial era, it was the capitalists who dissected the machinist’s craft and moved it into the planning room and the machines, leaving workers largely passive. In the age of AI, knowledge workers codify their own expertise by hand, and mostly do so willingly, because externalizing what they know immediately amplifies their personal output.

Once the means of production have been handed over, where does labor’s edge in the market lie? The MIT economist David Autor and his colleagues found that as of 2018, roughly 60 percent of U.S. employment was in job specialties whose titles did not exist in 1940. Innovation that augments human labor generates new work, whereas automation that merely replaces workers does not. Anthropic’s report offers a vivid microcosm: writers on a product-design team who once crafted user-facing copy automated that workflow and rotated wholesale into writing model evaluations and system prompts. As the report puts it, “Writers who used language to drive human behavior are now using it to drive Claude behavior.”

Historically, many of the new occupations born of industrial revolutions have grown out of meta-work. Analyzing the U.S. Department of Labor’s landmark 1899 Hand and Machine Labor study, the economic historians Jeremy Atack, Robert Margo, and Paul Rhode found that while mechanization displaced vast numbers of manual steps, it simultaneously created tasks with no handicraft equivalent, such as firing and maintaining steam engines or supervising factory workflows. None of those jobs fashioned the product directly; they kept the machinery and the workforce running. Emerging forms of AI meta-work deserve to be taken seriously, because they may well point the way to the next generation of roles. Consider the forward-deployed engineer, or FDE, who embeds directly with enterprise clients to configure models, connect legacy systems, and weave AI into day-to-day business processes. According to the labor-market analytics firm Lightcast, as reported by Fortune, U.S. job postings for forward-deployed engineers between January and August 2026 jumped more than tenfold from a year earlier, standing more than forty times above their 2023 level.

What distinguishes AI from earlier technological revolutions, however, is the velocity of its iteration. Because model capabilities advance so rapidly, new meta-work roles are likely to have far shorter shelf lives than their industrial predecessors. From prompt to context, harness, and loop, the conceptual stack of AI engineering has turned over four times in barely three years. In 2023, the prompt engineer was hailed as tech’s hottest new career; by April 2025, a Wall Street Journal headline had declared the job obsolete, noting that models had grown far better at inferring user intent while companies had begun training ordinary employees to write their own prompts.

Fleeting job titles are the lesser concern; headcount is the greater one. In the past, new jobs could always fill the gaps left by old ones, but in the face of the revolution brought by AI, I deeply doubt whether new roles can emerge fast enough to match the pace at which old ones disappear. The Nobel laureate economist Daron Acemoglu and his colleague Pascual Restrepo model technological change as a contest between two forces, automation that displaces workers from existing tasks and new-task creation that reinstates them. For four decades after World War II, from 1947 to 1987, the two forces roughly balanced out. After 1987, displacement accelerated while reinstatement slowed. Britain’s First Industrial Revolution endured a comparable stretch, which the economic historian Robert Allen termed “Engels’ pause”: from roughly 1780 to 1840, output per worker climbed substantially while real wages barely moved. Subsequent scholars, re-estimating the period with updated data, have found a narrower gap and little change in labor’s overall share of national income, yet most series still show workers’ real consumption lagging behind productivity growth for decades. We may be entering a similar interregnum today, with a crucial difference. Earlier machines supplanted muscle and manual dexterity, leaving humans a refuge in cognitive work. For the first time, humanity confronts systems that match and increasingly surpass human performance across cognitive tasks themselves, leaving no obvious higher ground.

Even where new work emerges, who captures the gains that meta-work creates? In the Grundrisse, Marx noted capital’s dual tendency, creating vast amounts of disposable time only to convert that freed time back into surplus labor. In his 1932 essay “In Praise of Idleness,” the British philosopher Bertrand Russell illustrated the paradox with pin manufacturing. Suppose an invention allows the same workforce to make twice as many pins in an eight-hour day, though the world needs no additional pins. In a sensible society, everyone would work four hours instead of eight. In the actual world, half the workforce continues working eight-hour days while the other half is thrown out of work. As Russell put it, “The morality of work is the morality of slaves.”

Prisoners of Surplus Value

Whenever I caught up with friends during my trip, the conversation eventually circled back to a single word, busy. AI can write software, draft documents, and build spreadsheets, yet everyone’s calendar seems more crowded than ever. A familiar refrain now echoes across the internet, claiming that every industrial revolution has left humanity more harried, never freer.

The historical record tells a more nuanced story. Annual working hours did surge during Britain’s First Industrial Revolution, but over the longer arc they fell dramatically. In 1870, workers across industrializing economies averaged more than 3,000 hours a year; by 2000, annual hours across those same countries had dropped to roughly 1,350 to 1,900. Average work time has never simply marched upward. What changed is who works the longest hours. As the economist Dora Costa showed, in the 1890s the longest workdays fell on the lowest-paid laborers, who worked long hours simply to subsist; as wages rose, they steadily bought back their time. By the 1990s, that relationship had inverted, and the longest hours were being logged by the highest earners. Technological progress does not distribute busyness evenly. Rather, the closer workers stand to the newest means of production, the busier they become, for two reasons.

The first is how direct economic returns are distributed. In 1821, in the third edition of On the Principles of Political Economy and Taxation, David Ricardo added a celebrated chapter titled “On Machinery.” Having once believed that mechanization benefited every social class, Ricardo recanted, concluding that “the substitution of machinery for human labour, is often very injurious to the interests of the class of labourers.” If even Ricardo, once an optimist about machinery, conceded as much, technological gains clearly do not flow to labor on their own. In contemporary terms, they accrue disproportionately to those closest to meta-work and to the newest means of production, the very people who consequently work the most relentlessly.

The second is how scalable technology amplifies differences in skill. Studying U.S. labor data from the 1980s through the early 2000s, the economists Peter Kuhn and Fernando Lozano found that among college-educated, salaried men, the earnings payoff for working beyond forty hours a week more than doubled, drawing a larger share of top earners into long workweeks. The more an extra hour pays, the harder it is to leave on the table. In his classic 1981 paper “The Economics of Superstars,” the University of Chicago economist Sherwin Rosen identified a closely related mechanism: once technology allows one person’s output to be replicated at vast scale, small differences in talent translate into large differences in income. Meta-work obeys the same logic. Autonomous agents will execute a lucid playbook thousands of times, and a muddled one just as dutifully, so the gap between first-rate and passable meta-work compounds with every hour the agents run. The better someone is at meta-work, the more each human hour is worth, and the harder it becomes to stop.

And even for those who want to stop, stopping alone is hard. This is the classic prisoner’s dilemma. Consider how Western Europe acquired its modern culture of leisure. Until the late 1960s, Europeans worked at least as many hours as Americans. The divergence began in the 1970s and 1980s, when unions in declining European industries rallied behind the slogan “work less, work all,” pushing through laws that shortened the workweek and mandated paid vacations. Those laws failed to create more jobs. What they did instead, almost by accident, was reset the rhythm of an entire society. Because leisure carries a social multiplier, with vacations becoming more rewarding the more friends and colleagues take them at the same time, collective mandates shifted social norms toward time off. A campaign meant to ration employment ended up unlocking leisure, demonstrating how a prisoner’s dilemma is broken: no individual can afford to slow down unilaterally, but shared rules allow everyone to slow down together.

Today’s AI arena has no such rules. As I see it, the AI giants are locked in an unforgiving contest in which they have no choice but to keep pushing forward, and whoever stops first is bound to be eliminated. Neither state machinery nor the leading frontier companies have the power to alter that trajectory single-handedly. Even the capitalists themselves can no longer rein in the colossus of capital as it drags the wheels of history relentlessly onward. The high-paid knowledge workers closest to the means of production, along with the capitalists themselves, in seeking to extract and magnify the outsized returns of meta-work, have trapped themselves in an inescapable prisoner’s dilemma and the limitless circulation of capital.

As individuals caught in the current of AI, we must be careful not to let meta-work pull us into the trap of a technological system with no purpose of its own, in which the pursuit of enabling conditions displaces the pursuit of the thing itself. In Capital, Marx observed that money circulated as capital becomes an end in itself, and that “the circulation of capital has therefore no limits.” In his 1930 essay, Keynes mocked the “purposive” man, for whom “jam is not jam unless it is a case of jam to-morrow and never jam to-day.” By nature, meta-work is purely instrumental, preparing the ground for something else. Yet a system unmoored from final purposes excels at making instruments multiply without end. Configuring the environment becomes an absorbing pastime in its own right, building an agent to build the agents, a dashboard to watch the dashboards, and ever finer rules for an undertaking that never quite begins. Efficiency was never meant to be more than an instrument. Once it becomes the objective, means and ends have traded places.

Then again, isn’t our pursuit of intelligence itself a way of mistaking the means for the end?

References

  1. Wilmers, N., Catipon, R., Massenkoff, M. & McCrory, P. (2026-10-16), “How employment is changing at Anthropic,” Anthropic Research (with companion survey report by Sacher, S. et al., “The AI expectations gap,” 2026-10-16).
  2. Keynes, J. M. (1930), “Economic Possibilities for our Grandchildren,” The Nation and Athenaeum (reprinted in Essays in Persuasion, Macmillan, 1931).
  3. González, V. M. & Mark, G. (2004), “‘Constant, Constant, Multi-tasking Craziness’: Managing Multiple Working Spheres,” Proceedings of CHI 2004, pp. 113–120.
  4. Strauss, A. (1985), “Work and the Division of Labor,” The Sociological Quarterly, 26(1), pp. 1–19.
  5. Smith, A. (1776), An Inquiry into the Nature and Causes of the Wealth of Nations, Book I, Chapter 1.
  6. Böhm-Bawerk, E. von (1889; English trans. W. Smart, 1891), The Positive Theory of Capital, Macmillan.
  7. Taylor, F. W. (1903), “Shop Management,” Transactions of the American Society of Mechanical Engineers, Vol. 24 (republished as Shop Management, Harper & Brothers, 1911).
  8. Braverman, H. (1974), Labor and Monopoly Capital: The Degradation of Work in the Twentieth Century, Monthly Review Press.
  9. Confucius, The Analects, Book 15 (Wei Ling Gong), 15.10.
  10. Lütke, T. (2025-06-19), post on context engineering, X.
  11. Karpathy, A. (2025-06-25), post on context engineering, X.
  12. Lopopolo, R. (2026-02-11), “Harness engineering: leveraging Codex in an agent-first world,” OpenAI.
  13. Osmani, A. (2026-06-07), “Loop Engineering,” AddyOsmani.com (quoting Boris Cherny, 2026-06-06, and Peter Steinberger, 2026-06-07).
  14. David, P. A. (1990), “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox,” American Economic Review, 80(2), pp. 355–361.
  15. Brynjolfsson, E., Rock, D. & Syverson, C. (2021), “The Productivity J-Curve: How Intangibles Complement General Purpose Technologies,” American Economic Journal: Macroeconomics, 13(1), pp. 333–372.
  16. Drucker, P. F. (1999), “Knowledge-Worker Productivity: The Biggest Challenge,” California Management Review, 41(2), pp. 79–94.
  17. Marx, K. (1857–58; English trans. M. Nicolaus, 1973), Grundrisse, “Fragment on Machines,” Penguin.
  18. Polanyi, M. (1966), The Tacit Dimension, Doubleday.
  19. Autor, D., Chin, C., Salomons, A. & Seegmiller, B. (2024), “New Frontiers: The Origins and Content of New Work, 1940–2018,” Quarterly Journal of Economics, 139(3), pp. 1399–1465.
  20. Atack, J., Margo, R. A. & Rhode, P. W. (2019), “‘Automation’ of Manufacturing in the Late Nineteenth Century: The Hand and Machine Labor Study,” Journal of Economic Perspectives, 33(2), pp. 51–70.
  21. Fore, P. (2026-09-03), “One of the fastest-growing jobs in Silicon Valley sends engineers straight to customers’ offices to get AI up and running, and pays more than $188,000,” Fortune (citing Lightcast data).
  22. Bousquette, I. (2025-04-25), “The Hottest AI Job of 2023 Is Already Obsolete,” The Wall Street Journal.
  23. Acemoglu, D. & Restrepo, P. (2019), “Automation and New Tasks: How Technology Displaces and Reinstates Labor,” Journal of Economic Perspectives, 33(2), pp. 3–30.
  24. Allen, R. C. (2009), “Engels’ Pause: Technical Change, Capital Accumulation, and Inequality in the British Industrial Revolution,” Explorations in Economic History, 46(4), pp. 418–435.
  25. Crafts, N. (2020), “Slow Real Wage Growth during the Industrial Revolution: Productivity Paradox or Pro-Rich Growth?,” CAGE Working Paper No. 474, University of Warwick (published in Oxford Economic Papers, 74(1), 2022, pp. 1–13).
  26. Russell, B. (1932), “In Praise of Idleness,” Harper’s Magazine (October 1932), pp. 552–559.
  27. TIANYU2FM (对谈未知领域), Episode 157 with Ma Zhaoyuan, 2026-09-08 (episode title and show notes).
  28. Voth, H.-J. (2000), Time and Work in England 1750–1830, Clarendon Press, Oxford University Press.
  29. Huberman, M. & Minns, C. (2007), “The Times They Are Not Changin’: Days and Hours of Work in Old and New Worlds, 1870–2000,” Explorations in Economic History, 44(4), pp. 538–567.
  30. Costa, D. L. (2000), “The Wage and the Length of the Work Day: From the 1890s to 1991,” Journal of Labor Economics, 18(1), pp. 156–181.
  31. Ricardo, D. (1821), On the Principles of Political Economy and Taxation, 3rd ed., Chapter 31, “On Machinery.”
  32. Kuhn, P. & Lozano, F. (2008), “The Expanding Workweek? Understanding Trends in Long Work Hours among U.S. Men, 1979–2006,” Journal of Labor Economics, 26(2), pp. 311–343.
  33. Rosen, S. (1981), “The Economics of Superstars,” American Economic Review, 71(5), pp. 845–858.
  34. Alesina, A., Glaeser, E. & Sacerdote, B. (2006), “Work and Leisure in the United States and Europe: Why So Different?,” NBER Macroeconomics Annual 2005, Vol. 20, pp. 1–64.
  35. TIANYU2FM (对谈未知领域), Episode 159 with Yu Mingfeng, “God Is Dead, Yet Science Cannot Cure Our Spiritual Poverty,” 2026-10-08 (show notes and chapter titles, on the purposelessness of technological systems).
  36. Marx, K. (1867), Capital, Vol. 1, Chapter 4, “The General Formula for Capital.”