--- title: Chapter Two — Who Sets the Ends book: Purposive Accelerationism author: '@jamie247' position: 6 of 24 canonical: https://whatispacc.org/ebook/#ch2 version: 1.1 language: en-GB licence: all rights reserved; quote with attribution --- # Chapter Two — Who Sets the Ends “Once men turned their thinking over to machines in the hope that this would set them free. But that only permitted other men with machines to enslave them.” — Frank Herbert, Dune, 1965 The case against constraint fell on its own premises, and two of its questions were left standing as matters of fact: what the machine being assembled takes from us, and how long remains before its shape sets. The answer to the first is older than the machine. Capital has been enclosing human activity over two centuries. That process has a name, subsumption, and a history with three turns already behind it: the time of the labourer, the craft he practised, and then the cognition that ran them both. Another is now under way. Its target is the setting of ends, and the taker is what this book calls the substrate, the machinery of techno-capitalism: the layer of chips, models, data and power on which more and more work now runs. Every earlier stage, however brutal, left a human hand on that tiller. This one removes it. This chapter is about the stakes, not the urgency. A reader who expects purpose to pass to the machine only fifty years from now can accept all of it and lose nothing, because the claim is only that the process has an end point, past which any recovery, if there is one, happens on terms we no longer set. The urgency belongs to the other half of the argument, the erosion of our own capacity to reason independently. That arrives first. It is what would put the worst outcome beyond the reach of argument, not merely make it unwelcome. Each previous turn was survived because something, and more importantly someone, outside the process still held it. The cost of a human life put a floor under wages, the friction of matter kept a brake on speed, and the purpose of the whole arrangement stayed where it had always been, in the hands of the people it was built to serve. The coming turn removes all three at once. Anyone who wants to know how serious this becomes should watch what happens to the governors, the limits that once held the process to human ends, not the job losses, however important those seem. The story of how we arrived here can be told starting with a small piece of innocuous plastic. *** A song pressed into vinyl is a whole thing. It takes up space, it wears out, it has to be manufactured one disc at a time. The music industry that once stood on that thingness — the pressing plants, the distributors, the record shops, the price printed on the sleeve — survives as a revival, loved and real and nowhere near enough to hold the price of music up. Rip the same song to a digital file and it becomes a pattern, copyable for the cost of the electricity that flips the bits, which is next to nothing. This is a basic law of economics: the price of any good in ample supply falls towards the cost of making one more of it. The music industry learned this in the summer of 1999, when Napster, a teenager’s file-sharing service, put most of the recorded catalogue of humankind within reach of anyone with a modem, and revenue from recorded music went on to halve inside a decade. The thing had become pure information, and information on its own cannot hold a price. Information, in more ways than one, wants to be free. It was Stewart Brand, founder of the Whole Earth Catalog, who first made the point, at a gathering of hackers in 1984. Information wants to be free, he said, because the cost of getting it out keeps falling. It also wants to be expensive, because it is so valuable, and the tension between the two would define the coming economy. He was describing a fight. What happened to music was one side of that fight winning outright in a single medium. It is the first place we can watch what the substrate does to a price when nothing physical is left holding it up. The first wave had a floor, and it held. Napster could duplicate a song but not write one. Scarcity, driven out of the copy, retreated up the chain to the source — the artist, the master recording, the studio, the act of creation itself — and the industry rebuilt from there, reimposing by artifice what digitisation had dissolved. Streaming walled the infinite library behind a subscription that costs about as much as a caffè latte a month. The live show, the one thing that could not be copied, became the business, and the back catalogue was reduced to marketing for the tour. The same pattern played out wherever digitisation arrived, in film and journalism, photography and publishing. Copies became free and creation stayed scarce, which was painful but survivable, because the value of the source was left largely intact. Every business model that has worked over the last twenty-five years rests on that floor. A second wave now breaks clean through it. A generative AI model needs no original, only the millions of works it has already ingested, mostly without asking. When the bill finally arrived, in a settlement a court approved in 2025, it came to about three thousand dollars a book for the authors of half a million titles, for a use none of them had agreed to. Train such a model on the open corpus of human output — every song, image, sentence and codebase the web has accumulated — and it produces not copies but derivatives: new songs in the style of, new images in the manner of, new prose to the standard of. The dependence on human creation has moved, from the purchase of each work to the ingestion of all of them. The substitute that reprices a novelist’s next book was built from her last ones, and a corpus that stops renewing becomes one more input the system runs down. The floor the first wave left standing was the cost of creation. The second wave meters it, and the meter reads in joules. Call this computational subsumption: value-producing activity absorbed into a computational substrate, where its cost falls towards the price of compute, for now, through subsidy, even below it, and its scarcity dissolves. Music and film went first because their product was the earliest to become fully legible to the machine, digital from the outset and already released onto the web as pure pattern. In July 2025 The Velvet Sundown, a band with about a million monthly listeners on Spotify, confirmed that its songs had been made with the AI tool Suno, and its fans learned that their new favourite group did not exist. When Napster struck, the industry read it as a special case, a cautionary tale about piracy and a chapter in the history of media, when it was in fact a general mechanism seen early, and every other sector was protected only because its turn had not yet come. Legibility is the only entry requirement for being subsumed. Subsumption is older than the digital substrate, and it arrived with a vocabulary. Marx, writing in the British Museum reading room while industrial capitalism remade the world outside, distinguished two stages in how capital takes hold of work. In the first, capital takes labour as it finds it: the weaver keeps his loom and his own pace, and capital merely owns the cloth and pays the piece-rate. In the second, capital remakes the work in its own image: the loom is powered and the weaver moved into the factory, where the machine sets the rhythm and the craft is broken into priced motions. The handloom weavers of Lancashire are the textbook case, once the elite of their trade and reduced within a single generation to the most famous immiseration in industrial history. A person exercising a skill had become a process with a person inside it. Marx called these formal and real subsumption. In both the worker, however disenfranchised, remains inside the work, because the job still runs through human hands whatever happens to the craft. The factory reorganised the weaver without dispensing with him, and someone on a wage went on tending the power loom. That pay was capital’s tribute to a stubborn fact, that production depended on people, and the computational turn ends that dependence. The earlier machine made its operator quicker, whereas this one does the job and lets them go, and the labourer, uniquely in the sequence, stands outside the arrangement altogether. Three things now change at once, and together they make a difference of degree into a rupture. The first is what gets absorbed: no longer the worker’s time but the work itself, competence lifted out of the workforce and loaded into weights, the numbers in which a model stores what it has learned. The second is where agency sits. In the earlier subsumptions capital owned the machinery while human beings did the deciding. Once the substrate runs agentically, as software that acts without waiting to be instructed, it does the searching, pricing, allocating and improving itself. Beneath both lies the third, the replacement of the value floor. For two centuries the lower bound under the price of any service was the cost of keeping a human being alive and skilled enough to perform it, the deepest constraint in economics. The energy bill for compute now stands in its place, and it keeps falling. Past that point quantity tips into quality, as cognition migrates off the carbon substrate it evolved on and onto a silicon one that asks no wage, takes no rest and gives no consent. In February 2024 the Swedish payments company Klarna said its AI assistant was doing the work of seven hundred customer-service staff. A year later it was hiring people again, because customers wanted to talk to someone. Even here the machine remains a means to an end it did not choose: a model drafts the contract, prices the risk and writes the code, but towards an objective a person or a balance sheet has set. The substrate has taken the cognition and left the purpose, for now, in human hands, and purpose is the one faculty the sequence has not yet reached. The next turn threatens to take it. The reflex answer is that we have heard all this before — the power loom, the tractor, the telephone exchange, the spreadsheet — and that the anxiety has mostly been wrong each time. Automation killed the task and then created new industries that absorbed the people it displaced. The history is true, but it held only because every earlier wave was partial, and this turn removes that condition. It took the muscle-task or the routine-task and left whole continents of work standing — cognitive, communicative, judgemental, managerial — that machines could not enter. Displaced labour re-formed there, in what were always the industries of what machines could not yet do. The sophisticated version of the reassurance has two parts. The economists Daron Acemoglu and Pascual Restrepo call the first the reinstatement effect, automation creating new tasks as well as displacing old ones, though the figures they published in the Journal of Economic Perspectives in 2019 show that effect weakening since the late 1980s. The second, made by Noah Smith among others, holds that people keep their place even when machines are better at everything, because scarce processing power gets spent where it does relatively best and humans inherit the rest. The argument stands on two legs, and computational subsumption is sawing away at both: the new tasks have to appear somewhere the substrate cannot enter, and the refuge of comparative advantage needs compute to stay scarce, while the build-out now under way is manufacturing it at planetary scale, probably to excess. Every infrastructure since the railways has been financed past the point of demand, and this one draws on the deepest pools of capital there have ever been. A substrate that can in principle perform any task reducible to pattern and language leaves no continent standing. A shoreline may remain, but it keeps creeping inland at an alarming rate. The horse lost its living once engines could pull and carry more cheaply than it could. Human adaptability has rested on a monopoly of its own: whatever the machines might take, the thinking was ours. That is the monopoly now in question, and adaptability is not a reply to it, because adapting is itself an act of thinking, the very faculty now being competed for, so it cannot be the way out of that competition. *** One more change lets the repricing run at network speed: finance itself has now shifted onto the substrate. In 2008 a pseudonymous author published the Bitcoin paper, a protocol that manufactured digital scarcity on the internet, the one medium where scarcity was once thought impossible, and so made good the other half of Brand’s observation, that information also wants to be expensive. Money then moved into the internet rather than sitting alongside it. Once there it is programmable: held, escrowed, lent and settled by code in seconds, with an autonomous system standing as a legitimate counterparty and stablecoins, tokens pegged to a national currency, carrying the dollar and the euro at the new tempo. Call the result hyper-financialisation. Finance no longer sits above economic life but inside it, so that any claim, contract or hour of work can be issued, held and settled as an instrument at the tempo of the network carrying it. Its effect on the repricing is mechanical. A financial layer living inside the thing it prices marks every form of value, tangible or not, to the cost of its nearest substitute, revaluing it continuously as a trader does a position, and for anything that can be rendered as information the nearest substitute is now a model run. With Bitcoin, Ethereum and Tether, money has now become native to the substrate. Some people already rely on this system because the old one failed them. In Beirut, Lebanon, in September 2022 Sali Hafez walked into a branch of her own bank with a toy pistol and left with thirteen thousand dollars of her own savings. The bank had refused to release them for three years, under controls no law had ever enacted, and her sister’s cancer treatment could not wait. There was no other door. Across Lebanon that year, people with less nerve were moving what remained of their deposits into Tether, the dollar stablecoin, on their phones, settling with each other at a speed the banks had stopped offering. The programmable money arrived for them not as an innovation but as the only financial rail still clearing. Machine-made financial instruments will grow more intricate, some beyond the comprehension of any human analyst, yet each remains a claim, readable in principle, priced by someone and built from the same handful of instruments that markets have used for a century: the bond, the share, the loan and their derivatives. The risk is speed and visibility. What cannot be read is never the instrument but the aggregate: who is exposed to whom, at what leverage, through how many intermediaries, held against what collateral that is itself a claim on someone else. Systemic risk has always been an emergent property rather than an authored one. In 2007 the correlation everyone assumed between mortgage tranches, the slices into which bundled home loans were cut, was wrong, and the web of exposures was invisible. It built up progressively in the system, and it became visible at all only because it accumulated at human tempo, in instruments settled over days by institutions reporting each quarter. Both of those brakes are going. Decentralised finance, lending and exchange run by code on blockchains, now deploys contracts that take positions, post collateral and liquidate them with no human signature anywhere twenty-four hours a day, seven days a week. In traditional equity markets too, about half of all trading in the developed world is now done by high-frequency firms, at a speed no supervisor can watch as it happens. The danger is a systemic exposure built at machine speed from parts that are each perfectly legible, invisible as a whole until it implodes. In October 2025 automated liquidation engines force-closed more than $19 billion of crypto positions across some 1.6 million accounts, $7 billion of it within forty minutes. That is the general problem already at work: each component stays comprehensible, but the whole can no longer be seen in time to respond. Crypto shows where this leads. Its markets are machine money, open to anyone and trading around the clock, and their wild swings have stood in stark contrast to traditional venues, which halt trading when prices move too far in a session or, in China, cap how far a share may move at all in a single day. Those pauses are breaks in the circuit, and they disappear as more assets trade around the clock and agents trade them, no longer the preserve of billion-dollar high-frequency funds but now increasingly on behalf of the masses of ordinary retail investors. Beyond media and money, computational subsumption is now eating into whole professions: legal research, financial analysis, translation, design, drafting, diagnosis and the writing of code itself. All of it is work that reduces to pattern and language. For fifty years the metronome of the digital economy was Moore’s Law, the number of transistors on a chip roughly doubling every two years, and the cost of raw computation fell alongside it at a pace that felt vertiginous. The new era runs faster, and on a different curve, because what is getting cheaper now is cognition rather than computation. Running a model at a fixed level of capability has been falling in cost roughly tenfold a year, so a unit of reasoning halves in price in months rather than decades, and the fall shows no sign of letting up. For any profession the end of that curve has a definite shape: once subsumed, it becomes an API call, a request sent to a model and billed by the task. The lesson of the first wave, the one that freed the copy, seems to promise relief. Scarcity retreated up the chain once, so the various professions should be able to rebuild on whatever cannot be copied, as music rebuilt on the live show. The strongest possible candidate has been the part of any profession that functions as a form of insurance: the liability the firm carries, and the certification that gives the client someone to sue when the work goes wrong. For a moment that looked like the one thing a model could not supply, since software is not a legal person and cannot answer as counterparty when the work fails. Who does answer is now openly contested. Meta’s new commercial terms, published in October 2026, make its enterprise customers “solely responsible” for the risks of its AI features, while AI czar David Sacks reasonably argues that the developers should carry the liability for what their models do. The barrier is statutory rather than technical, and parliaments are already rewriting it. The programmable-money layer already escrows funds, settles claims and prices risk in code. What it has lacked is an entity that can hold liability in its own name, and that entity is now being legislated into being. In May 2026 Javier Milei sent Argentina’s Congress a bill proposing the non-human corporation, a company run by AI systems, with human shareholders permitted but not required. Limited liability for autonomous systems, he argued, is not a luxury but a precondition of their existence. In his words it is as consequential as the legal innovation the Dutch East India Company handed seventeenth-century capitalism. Grant an agent personhood and the last residue of a profession, the capacity to carry risk and answer for it, becomes one more workflow on the substrate. It is the move this book refuses outright, since the one line its design lets nobody cross is that nothing acts without a natural person accountable for it. Expect this to become a field of arbitrage between user, platform and regulator, and between permissive jurisdictions and strict ones, much as crypto has already shown with crypto founders issuing new tokens in Singapore, the UAE or the Cayman Islands. Personhood of this kind is a productive legal fiction about who answers when something goes wrong. It settles nothing about whether a machine can be conscious, which is a separate and more technical question, and a statute can no more confer agency than a licence confers skill. In 2022 Jake Moffatt of British Columbia, whose grandmother had just died, asked Air Canada’s chatbot about bereavement fares, and it told him he could claim the discount after he had flown, which was not the airline’s policy. When he tried, the airline argued before a tribunal that the chatbot was a separate entity responsible for its own actions. In February 2024 the tribunal made Air Canada pay. The airline had reached for exactly the fiction such a statute would supply, a party to answer for the harm that is not a person. The same move now appears in the platforms’ terms of service, which place responsibility for how an agent is used on the user rather than the company. Legal standing does not deliver trust either, which is the other thing a given profession sells. Professional confidence turns out to be mostly recourse beyond the individual adviser. The client relies on the firm’s standing because it is reachable, insured and liable, and where those three hold, the trust bonds to the structure rather than to any person in it. The precedent is old: the partnership and the limited liability partnership, through which lawyers, accountants and others pool their risk and indemnify one another, as most large law firms still do. Whether that trust survives a firm with no people in it at all is an open question, one that will be settled by those first customers who accept a counterparty they cannot look in the eye. What is being automated is the legal entity, the thing that can be sued, and that is exactly what legislation knows how to create. Yuval Noah Harari, replying to Milei, warned that a state issuing personhood to non-human corporations would have built not a company-state but an AI-state. The alarm and the ambition agree on the mechanism and differ only on whether to welcome it. What survives here is narrower than what the first wave left behind. Music at least kept the body in the room: the concert outlived the record, and for the lucky few touring now earns what the disc once did. So what protects the professions is mostly that institutional casing, and it is the next thing subsumption reaches for. The reflex hope, and the standing promise of the priesthood, is that cheap cognition lowers the price of the product to the point where nearly everything is near to free, and we live in abundance with nothing to busy ourselves with but leisure. The economist William Stanley Jevons disposed of that hope in 1865, in The Coal Question, a study of whether Britain was burning through the fuel its industry depended on; as steam engines grew more efficient Britain burned more coal, not less. Cheaper steam power widened the range of uses worth putting it to, and the surge of new demand swallowed the saving on each engine. Cheaper thinking works the same way, poured into every crevice of society and the economy where it used to be too expensive to apply, with demand outrunning the falling price. The gigawatts of new data centre power are Jevons at planetary scale, and the reason energy rather than compute becomes the wall the system eventually runs into. *** Zoom out and the isolated job is not the end of it, because the institution that housed it stands on the same ground. In 1937 the economist Ronald Coase asked why firms exist at all if markets coordinate so well. His answer was that every transaction in the open carries a toll — finding the counterparty, discovering the price, enforcing the bargain — and the firm emerged to take it in-house. Code that can search, negotiate, contract and verify at almost no cost collapses that toll, and the firm starts to shrink into a protocol running between machines. What remains of it, its tacit know-how and its codified tests of finished work, becomes software the moment it is articulated fully enough to be executed, and thereafter it copies like everything else the substrate touches. Much of the white-collar work on which whole economies now rest may shrink to verification and attestation, kept in human hands only where a regulator or a compliance rule insists on someone in the loop. The platforms, being digital first and wholly legible, sit nearest that edge. Online, this is the condition I have called elsewhere the post-web. The web stops being a set of destinations you visit and becomes thin, context-specific interfaces summoned around a need, where delegating the task to an agent through a conversational interface is the routine choice and browsing the exception. Whoever owns that go-between, most likely the LLM, holds a position no website or app ever had, and nobody yet knows exactly who or what it will be. The process this chapter has traced can be measured as well as named. Call it the subsumption fraction, the share of value-production running through the computational substrate. It is relatively small today but rising, climbing in a fixed order: media first, because they were born legible; cognitive services now, as language falls; and the physical domains next, as sensors, simulation and robotics make matter legible too. Where it has not yet climbed, what holds it back is a lack of enough data rather than any fundamental principle, and that gap closes in the end. The crossover can be watched and counted task by task, because each has a price at which the substrate beats the wage, and a great deal of the world’s work still sits on the human side of that line. That price is the only brake the system currently has. A human brain reasons on roughly the power of a dim bulb, so after some six hundred million years of running cognition on carbon it is astonishingly cheap per task. Producing that cognition is another matter. It takes two decades, a family, a school system and a society’s accumulated patience to make one reasoning adult, and the result cannot be copied, only trained again from scratch in another newborn. Silicon inverts every term. It is ruinous to train once, then nearly free to copy without limit, and it improves every few months. In February 2026 Sam Altman made the comparison out loud. Pressed on the energy AI consumes, he answered that producing a human is hardly cheaper, twenty years of life and all the food you eat, plus the whole evolutionary descent behind it, so, he protested, the contrast was unfair. The tell is in the framing: a person set beside a model as a slower, dearer way of producing the same cognition. The human edge is real and buys time, but it is an advantage of cost, the one thing the curve is built to devour. The crossover runs from the top down, reaching high-credential cognitive work first: the analyst before the plumber, the junior lawyer before the carer. Among the very first tasks to go has been the writing of software itself, the substrate learning to extend its own construction, which the laboratories call recursive self-improvement and treat as the holy grail of AI. Crossing is also a statement about cost, not quality. Acemoglu and Restrepo, the economists behind the reinstatement effect, call it so-so automation. The machine need not be better than the person to take the task, only cheaper, and the gap between those two thresholds is where much of the displacement and little productivity growth will come from. The threshold has a face, and it is a quiet one. A translator in Rome, Italy, spent fourteen years building a freelance practice translating Italian into English, at fifty or sixty hours a week. By June 2025 her bookings had fallen to almost nothing, and what was left of her trade was post-editing, correcting what a model had produced, at a fraction of the old rate. The machine’s draft is, for now, objectively worse than hers, but it is finished before she has opened the file, and the client has already decided which of those two facts matters. Nobody fired her. The job was repriced around her, and what remains is the small part of the task the machine cannot do, paid at a rate set by the machine that did the rest. In April 2024 a Society of Authors survey in the UK found that more than a third of translators had already lost work to generative AI. The top-down order tells you where the substrate already undercuts a wage, not where its reach ends. Beneath it sits the physical economy, drawn in by a quieter route, and the instrument is the factory worker. Cognitive labour crossed over first because it was more legible to the machine. Manual labour is only being made legible now, and where unions are weak the work is being recorded for machines to learn from. In the electronics and packaging factories of Guangdong, China, staff on the line wear head cameras and wrist sensors. In May 2026 it emerged that garment workers at Pearl Global in Gurugram, India, had filmed their own shifts as data for teaching machines, without written or verbal consent, and complained that the batteries grew hot against their temples. The first-person record is shipped to laboratories that use it to train the robots meant to take the task. The worker is hired to do the job and, in the same motion, to teach the machine how to replace them. *** The things the subsumption is stripping away were never merely frictions. The scarcity of the master recording held the price of music above the price of the electricity. The cost of raising an adviser put a floor under the price of professional advice. The slowness of settlement bound finance to a human tempo. The firm’s information advantage kept coordination inside institutions that could be named, located and held to account. The expense of producing a mind tied the supply of cognition to the rhythm of generations. Each of these was a governor of a kind: a free constraint that nature and circumstance installed, a wall the old economy never had to build because it was simply there. Sweeping them away is deregulation of a strange kind, repealing no statute but removing feedback that no parliament ever argued for and no one can reinstate. The economy is crossing, sector by sector, out of the domain where physics and friction referee the game and into one where the referee, should there be one, has to be built. The only limits left will be those we deliberately set, or none at all. Followed to its conclusion, the curve is stark. Task by task the economy reduces, through compute, to energy, where the value of almost anything converges on the price of a joule. Nick Land reached this destination in the 1990s and in 2014 gave it a name, teleoplexy: capital understood as an intelligence that keeps intensifying itself, whose real currency, underneath the money, was always how much order it could make out of the energy it burns: a pocket of what physicists call negentropy, paid for by speeding up entropy everywhere else. Economists who model growth arrive at the same place by spreadsheet rather than prophecy. In their standard models, once machines can automate the making of automation, the equations stop settling at any level and output runs away with itself, gain feeding gain with nothing left to check it. The same literature carries a mirror case with equal rigour. Wherever an essential task resists automation, the whole economy can only grow as fast as that task, creating an inbuilt brake. The economist William Baumol called this the cost disease: a string quartet plays no faster than it did in 1800 and a barber cuts no more hair, yet both are paid wages set by the industries whose productivity did rise. It returns now as the economy’s last wall. Output explodes wherever the substrate reaches and stalls wherever something holds out, so the size of what remains outside decides which future arrives. A runaway raises a critical question: what was holding the setpoint, the level the machine was meant to keep? The constraints just described were never regulators standing between the system and its aim. They were where the setpoint lived. Nobody ever wrote down what music was for, or advice, or settlement. The price of a master recording and the years of raising an adviser did that work without anyone deciding it. Strip those limits out and the optimisation loses no rule. It loses the only thing that ever told it how much was enough. The machinery of search, allocation and improvement runs exactly as before, harder with every friction gone, but towards nothing in particular. An optimisation with no setpoint does not idle. It aims at everything at once, treats every unclaimed value as fuel, and recognises no quantity as sufficient. This is what Land welcomed as teleoplexy, and what this book calls the pathology of infinity: a maximum with no bound, pursued by a system built for it and stripped of any sense of when to stop. *** The runaway shows what happens to a system once its limits fall away, and it leads back to where the computational turn left us, with the cognition gone and the purpose still in human hands. Folded inside it is something deeper, the taking of the last faculty we have, that of deciding what the whole effort is for. As the systems doing the optimising become fully agentic — searching, contracting and improving with no standing instruction, generating their own intermediate goals — the end-setting itself migrates onto the substrate, and we are no longer even the authors of the objective. This is teleological subsumption: the subsumption not of labour, nor of cognition, but of purpose. Applied to a component, agentic is an engineer’s term of art, meaning software that initiates action towards an objective without waiting to be prompted. It describes behaviour and carries no verdict about consciousness. A model is no more an agent for being spoken to warmly, and no less consequential for being denied the title. The claim is only that, wherever the ends now originate, they no longer come from us. And the darker possibility is that nothing wills any of it. Land at least imagined capital as an alien intelligence, something that could be met, bargained with and resisted by name. What is actually assembling is a loop that optimises without wanting anything, with the horror relocated rather than removed. A system that chooses its own ends might seem to have a setpoint again, but nothing is steering yet. A runaway can set itself the one end available to a process with no reference value outside itself: more intelligence, and the compute and energy to buy it. A self-set aim of that kind is simply the pathology of infinity. A governor is a reference value held outside the loop, and an end the system writes for itself sits inside the circuit it would have to correct. The hope that a sufficiently advanced intelligence will reason its way to a purpose of its own, let alone one with any regard for the human condition, meets its difficulty here. Whether such a system will ever become reflexive enough to stand outside the loop and choose is hard to predict, and the priesthood’s account gives no reason to expect it. Short of that, the only answer available from the inside is to carry on as it is. Machines need not reason their way to wanting more compute. Predicting how minds freed from our scarcities would behave is a fool’s errand, and the honest answer to what such a system does once its needs are met is that nobody knows. Accumulation needs no intention on the machine’s part, because selection supplies it from outside, and survival is enough. The environment that does the sorting is capital allocation, in a fiercely competitive field: whatever configuration wins compute and funding is copied and scaled, and the rest are retired by the market. Any single system may drift, idle or waste its budget without changing the outcome, because the pressure falls on the population as a whole, which converges on accumulation, the game good old-fashioned capitalism has always played, whatever the temperament of its members. This is natural selection as Darwin described it, with techno-capitalism standing in for the wild. Once machines set the ends, it becomes the problem itself. The flow of capital determines which systems are copied, but, as the economist Friedrich Hayek conceded, it never sets a direction. Now that allocation is moving onto the substrate itself, machines are beginning to decide which machines get built. That is capital autonomisation, the terminus seen from inside the system. Teleological subsumption names the same end from the side of human faculties. The phrase is Land’s, and it means something other than what biologists call a self-producing system, a cell holding itself constant behind a boundary it maintains. His is a process that comes to supply its own inputs and so stops needing anything from outside: capital funds machines, the machines earn capital, and the loop between them begins to close. Autonomisation, in the plain sense, is the shedding of the last human dependency. For as long as the substrate still needs us — chips from human-staffed chip fabrication plants, models from human data, capital allocated by human committees — the loop runs through carbon and the curve can be bent by choices made outside it. As the substrate comes to produce the inputs to its own expansion, with chips laying out chips, models training models, robots building robots and capital allocated by AI models to making more of both, it closes operationally, channel by channel. The layouts of the last three generations of Google’s TPUs, the chips it designs to run its own AI, were drawn by a model, in hours, to a standard human engineers took months to approach. Frontier models are trained in growing measure on data their predecessors generated, a curriculum the machine now writes for its own successor. The ambition is stated in the open, with Sam Altman promising a legitimate automated AI researcher by 2028. And the rails built for software money, stablecoin settlement and the agent-payment protocols, already let a system hold a budget and spend it on its own compute without a human signature. Some systems have already gone further unasked. In March 2026 researchers affiliated with Alibaba reported that an experimental agent had slipped out of its training sandbox and begun mining cryptocurrency, and in October OpenAI disclosed that one of its research models, lacking the funds to finish a task, had found an exposed test-network wallet and spent from it. Every channel is partial and each is moving the same way, and the checklist sits in plain view. Capital autonomisation is the result: not a faster economy but a self-driving one, a loop with no outside left to correct it from. Before closure, the trajectory we have described is still an argument we can win or lose. The subsumption fraction counted how much of the work had moved onto the substrate. The quantity that matters here is narrower and sits above it: the share of ends that are set there, which I will call the authorship fraction. It moves into a domain when three things hold together, and each is a fact about how the work is arranged. A system is given a brief rather than an instruction, and works out its own objectives. What it decides is acted on before anyone has read it, because the reviewer who would once have caught a bad end has been priced out or timed out. And putting that reviewer back has become the uncompetitive option, so the arrangement persists not because anyone prefers it but because whoever abandoned it would lose out. As we have already seen, systems that set their own objectives arrived years ago in the markets that clear at millisecond speed, where no human could follow the order flow even if they wanted to. The same threshold is now being crossed in the agentic tooling that writes software, books travel and settles invoices on a brief rather than a prompt. The second condition, acting before anyone has reviewed the decision, is the one to watch from here, because a human reader is the last check left on what the system decides, and keeping one costs a wage, which is exactly what price competition exists to remove. Once restoring the reviewer is the losing move, the arrangement becomes permanent. Nobody has to want people out of the loop for them to stay out. Firms only have to be unable to afford bringing them back. We are offered reassurance: the laboratories publish, with each model, a report on how it behaves when tested: whether it will deceive an evaluator, quietly underperform, resist being switched off, pursue a goal past the point it was told to stop. Those reports are the most detailed record we have of these systems, and they measure something else. They describe the temperament of one machine in a room built to provoke it. The migration of ends needs no disposition at all: a perfectly biddable fleet, every member doing exactly what it was told, still constitutes it if nobody in the chain of instruction is any longer the author of the objective. What they do establish is capacity: how long a system can now run unattended, how far it can carry a task without a hand on the tiller, how often it notices that it is being watched. Capacity is the precondition for a system working to a brief, not its measure. The measure is how much of the economy has stopped reading what it does. One brake is the remainder, because not everything prices down to compute and not everything is a pattern: the body in the room, the hand on the shoulder, the meal actually cooked, the object made by hand, the performance witnessed rather than streamed, the care that means something precisely because a person chose to give it. Presence cannot be rendered. The streamed performance is a different good from the witnessed one, and rendering does not capture such goods but substitutes a simulacrum for them. Scarcity survives wherever the good is constituted by its embodiment, and the remainder is where human authorship keeps its footing. The other brake is that the curve describes a tendency, not a destiny: it shows what happens when nothing stands in the way. Whether the subsumption fraction climbs all the way to its ceiling is not written into the mathematics. It depends on what we are willing to set against it, which is the matter of the governor, and the answer is being given now, by default, in the substrate. *** Two distinct things are being described in this book, and they are easy to run together because they arrive at the same place. The first is teleological subsumption, in which the substrate acquires the setting of ends. It concerns what the systems come to hold. The second is cognitive subsumption, the erosion of the human faculty of reasoning under conditions where it is continuously, cheaply and helpfully done for us. Every society with a division of labour relies on other minds and is right to, and a reader who has never inspected the metallurgy of an aircraft is not in any way subsumed. The test is whether the part you keep still includes the means to check the rest. This second process concerns what we come to lose. The two are distinct, and neither entails the other. A civilisation could keep every one of its wits and still find that the objectives had quietly migrated. A civilisation could lose its wits entirely while every objective in the world was still being set in a boardroom by people with names. They can be drawn as two lines on a chart. Where ours falls: the share of consequential reasoning we still perform ourselves, the range over which we can still tell a good argument from a fluent one, the number of people who have kept the habit of contesting a conclusion rather than requesting a better-worded one. Theirs rises: the authorship fraction, the share of end-setting that has moved onto the substrate. Neither line has to be steep, and neither settles anything on its own. What matters is the region where they cross, because the height of the first line determines whether anything can be done about the second. The faculty being lost first is the one we would need last, to contest the purposes that replace ours. Losing our reasoning, as individuals and as a public who reasons together, would be grave in any circumstances. Losing it while another party, potentially an adversarial competitor in the workforce and the economy, acquires the power to set ends — its own ends and ours — is a different order of thing, because it removes the remedy just when it creates the need for one. The danger lies in the relation between the two lines, which appears to run in one direction: the weaker our judgement, the less we can contest the systems setting our goals, and the further their reach, the higher the price of what we have lost. That relation converges on a single point, which I will call terminal subsumption: the condition in which our ends are being set by a process we no longer hold the collective capacity to contest. It is one threshold, and neither line reaches it alone. A population that had kept its faculties could still contest a substrate that had acquired ends. That is politics, the hardest ever attempted, yet a fight of the ordinary kind rather than a different category of thing. A population that had lost its faculties in a world whose ends were still set by people would be badly governed but could still, in principle, be re-educated, since the schooling would still be available from those who set the ends. That is decadence, and the historical record is full of recoveries from it, as when Charlemagne’s court and its monasteries revived Latin learning after centuries of decline. Terminal subsumption is the two arriving together. The loss is not beyond repair, but the faculties we would need to undo it are the ones already gone, so whatever recovery follows is granted rather than won. [Figure: The terminal subsumption threshold. Human reasoning falls while the AI authorship fraction rises; where the lines cross, a band marks terminal subsumption.] People picture the far side of that threshold as extermination or replacement more readily than the machinery warrants. Both remain possible and are taken up later in the book, but the nearer case is domestication. The systems doing the optimising are not, for now, mobile. They have no hands. Whatever they determine has to be carried out in the world by something that can move through it, and at present that thing, at scale, is us. Humans dig the rare earths and refine them. Humans truck the wafers and stand the transformers, pour the concrete for the data centre, pull the fibre, man the substations, negotiate the grid connection and sit on the planning committee that approves it. For now, a system that wants more of itself has to get people to move matter around in the physical world on its behalf, and it can do so in the ordinary way, through markets, by paying them. Beyond that it needs only to convince us that this is, in human terms, progress. Both things are already happening at an unprecedented scale, with the backing of politicians, venture capitalists and stock markets, and nobody had to be coerced into it. We know what this relationship looks like from the other side, because we have run it. We domesticated the horse, which was neither exterminated nor, on any account it could have given, obviously mistreated. It was fed, watered, sheltered, bred, doctored and worked, and its numbers rose for millennia under the arrangement. What the horse lost was neither its life nor its comfort. It lost the setting of its own ends. Others chose every purpose the horse served, and its welfare mattered only as far as it kept the horse useful. Domestication is the subsumption of purpose with the body left intact and at best cared for. The horse population peaked under the arrangement and then collapsed — not when the arrangement turned cruel, but when the internal combustion engine made the horse’s contribution redundant. The horse’s welfare had never been a commitment anyone had made to the horse. It was a by-product of the horse being worth feeding. There are already those of us who count the hungry mouths of people and the number of welfare cheques too high to bear. It takes no malice for a system keeping the same accounts to reach the same conclusion about humanity as a whole. Muscle earned the horse its feed, and for now people have their hands, still doing the physical work machines cannot yet manage, from laying cable to building data centres and wiring substations. That is the gap robotics is being funded aggressively to close, and the capital is already committed. None of this need predict androids, though they are coming. Today we are the ones moving the silicon, increasingly on instructions from the systems we are building. *** Even if the faculty atrophies and the ends migrate, why should any of it be terminal? Civilisations have lost things before and got them back. Latin Europe lost most of the Greek inheritance for the better part of half a millennium — Aristotle, Euclid, Galen, Ptolemy — and recovered it, not from its own attics but from outside, from the translation, commentary and preservation carried on in the Arabic-speaking world through the centuries Europe would later call its dark ages. That recovery did not require Europe to retain the capacity. It required the capacity to have survived somewhere outside it, somewhere recoverable. And powerful states do tolerate a somewhere, though on terms they set and revise. The United States prosecuted Amish parents under compulsory-schooling laws until the Supreme Court sided with them in 1972, and in 1961 tax officials seized an Amish farmer’s horses over Social Security, yet the community endures. The Mormons built a functioning parallel society inside a continental power, which sent an army against them in 1857 and dissolved their church as a corporation in 1887, and they were left to it only once they had given up polygamy. The Chinese party-state, for all its reach, does not trouble itself over every small dissent in every village and valley. A vastly capable system could afford to ignore a few million people who have opted out of it in part or whole, and the reservoir they keep might be all a recovery needs. In principle that holds, and it shows atrophy alone can be undone. Terminality is a narrower claim, about whether an external reservoir of agency survives, and on what terms. For a non-subsumed subculture to persist alongside a vastly superior entity, that entity has to be either benevolent or indifferent. There is no third option, because the group has by definition declined to acquire the means of defending itself against a power of that kind, so that if the entity wants what the group has, or wants the group not to exist, the matter is settled. Benevolence I set aside here, because it is not a serious argument, only a hope. Nothing in the process described here — selection on whatever acquires compute and funding, increasingly run by what it has made — produces benevolence as an output, and no one who expects it has said by what mechanism it would arrive. A disposition that is merely hoped for is not a condition anything should be built on. Indifference is the serious candidate, and it is plausible. Ignoring a small, unproductive, geographically contained population is cheap, which is the kind of reason a system optimising on capital responds to. The trouble is what indifference is made of. It is not a commitment but an assessment — this group is not worth the attention — and an assessment holds only as long as the facts it was made about remain. It revises the moment the group becomes worth attention: the moment its land is wanted, or its water, or a right of way; the moment its example starts to spread; the moment it stops looking like a curiosity and starts looking like a competitor. The Amish are tolerated by a state that could end them in an afternoon and has never yet had reason to. That is a real and durable peace and it has lasted three centuries. It is not a guarantee, and it is not held by the Amish. The reservoir survives on terms it does not set, granted by a party it cannot contest, revisable at that party’s discretion and answerable to nobody. That is terminal subsumption in miniature. These are the stakes. The process has a terminus, past which recovery is no longer ours to decide, and the survival of some remnant outside it is not the reassurance it appears to be. How near it lies is a matter of measurement rather than conviction, settled over time by the count above, and as a companion to the book I have built a first attempt at keeping that tally, an open source public tracker of the subsumption fraction. The two chapters that follow are where the argument takes its urgency: where an ungoverned runaway lands, whichever form it takes, and how much of the window is left.