Cheap Composition · 4 of 5 · Economics · Beta · 22 min
Where the Value Goes
Who keeps the savings when AI makes desk work cheap, and what to do about it
- As AI makes drafting, analysis and code cheap, value moves to what stays scarce: computing power, someone who answers for the work, a person on site, and owning the customer.
- Whether a firm, its clients or its staff keep the savings depends on competition and on how hard the work is to check, and where customers have few alternatives the owners are likely to gain first.
- The first losses show up in junior hiring, years before pay falls, and today's taxes and safety nets, which rely on wages, were not built for jobs that stop being offered.
In the last week of December 2025 we had a half-finished bathroom in our house in upstate New York. A countertop was booked to go in that week, and I didn't want it pushed into the new year. I could have gone away for a few days. I stayed home. While the work was going on we found a plumbing problem that would have been a disaster if I'd left. By the end of the year I had a beautiful bathroom and no working sink. (The Empty Search)
None of that was composition. In this series composition means putting thoughts together, at any stage of a piece of work, from the first draft to the critique and the plan. A model could have drafted the plan for that bathroom. It couldn't have been in the house when the plumbing turned out wrong.
AI is making composition cheap, and this essay is about who keeps the saving. My answer is that it depends on what is still scarce around the work: the computing power the models run on, someone who answers for the result, a person on site, and owning the customer. Where customers can't easily judge the work or go elsewhere, the price holds and the owners of the firm keep most of the saving first. Where rivals compete on price, customers get more of it. In both cases fewer hours of composition are paid for, and the loss shows first in junior hiring, years before pay falls. One law firm shows how this works.
Where the saving goes at a small law firm
A small law firm reviews standard commercial leases for small businesses. The numbers are made up, and follow one lease at one firm with the same partner throughout.
Before AI, an associate, a junior lawyer, spends three hours drafting the review at $100 an hour, so $300. Then a partner, a senior lawyer who co-owns the firm, reads the draft for mistakes before it goes out under the firm's name. That's what I mean by checking. It takes her half an hour at $300 an hour, so $150. Some mistakes get through anyway, and on average they cost the firm $50 a lease. The client pays $1,000. The $500 left over is the firm's margin, which pays the rent and the owners. The firm does 100 leases a month.
Now a model drafts the review for a few dollars and the associate spends half an hour on it, so drafting costs about $55. The table shows where the $245 saved can end up.
| Per lease | Before AI | 1. The firm keeps the saving | 2. Clients get it | 3. Checking eats it |
|---|---|---|---|---|
| Drafting (associate, plus the model) | $300 | $55 | $55 | $55 |
| Checking (partner's time) | $150 | $150 | $150 | $375 |
| Mistakes that get through (average cost) | $50 | $50 | $50 | $70 |
| Price the client pays | $1,000 | $1,000 | $450 | $1,000 |
| Firm's margin per lease | $500 | $745 | $195 | $500 |
| Leases a month | 100 | 100 | 250 | 100 |
| Firm's margin a month | $50,000 | $74,500 | $48,750 | $50,000 |
In the first ending the price holds and the firm keeps the saving. Its monthly margin rises by about half.
In the second, a rival offers the same review for $450 and the price follows. Small businesses that never paid for a review start to, so the firm does 250 a month. Its total margin ends up about where it was, and the clients keep the rest. This is the ending with the most work for lawyers, and cheap work that brings in new clients is the best hope for wages. Even here the associate's hours on leases fall, from 300 a month to 125, while the partner's rise from 50 to 125. To keep the associate as busy as before, the firm would need 600 leases a month.
In the third, the partner doesn't trust a draft she didn't watch being made. Checking takes her 75 minutes instead of 30, and new kinds of mistake get through a little more often. The saving is gone, because the associate's hours have moved to her. She has reason for caution. In a study described further down, consultants using a model were less likely to get a task right when it had been picked to be beyond the model. It could go the other way if partners learn to check machine drafts fast, or software does part of the checking.
Three questions decide which ending a firm gets.
- Can the client tell a good review from a bad one? If not, they pay for the name, and the price holds.
- Can a rival offer the same work? Where licences are plentiful and clients compare, the price falls toward the new cost.
- Is checking a draft cheaper than writing it? Only if the partner can find the mistakes, and she learned to do that by writing drafts herself.
In law, medicine and audit the first answer is usually no, and licences slow the second. So I expect the first ending more often than the second, at least at first. That's a forecast, and the example only shows what decides it.
What stays scarce when composition gets cheap
When composition gets cheap, value moves to what is still scarce. I count four scarce things.
Compute. The chips, data centres and electricity that run the models. On their guidance after the third quarter of 2026, Amazon, Microsoft, Alphabet and Meta expected to spend about $720 to $740 billion between them on capital projects in 2026 (Yield Theory), not all of it on AI, and in July Amazon, Microsoft and Alphabet all said demand outran supply. The firms that sell the chips, the buildings and the power get paid first. Whether the companies buying them earn it back is taken up below.
Trust. Someone who answers for the work, with a licence, a job on the line, an insurer or a name that can be sued. When anyone can produce a plausible contract or diagnosis, that is the scarce thing. I expect fewer people doing the work, and each person who answers for it to be worth more.
Presence. A body in an unpredictable place, like a plumber, a nurse or a teacher. My bet is that pay rises for this work, because it can't be done from a distance and robots aren't ready for a stranger's house. I could be wrong about this. Much of it is paid for from public budgets, and people leaving desk jobs may retrain for it, which would hold pay down. Robots are also being built for it. The robot maker 1X takes orders for a humanoid called NEO at $20,000 and promised US deliveries in 2026, and by early October 2026 no delivery to a paying customer had been announced. In September 2026 the robots of another maker, Figure, tidied, folded towels and made beds in 30 rented homes they had never seen, and succeeded 56 percent of the time in the company's own test. Once a robot can be in the house when something goes wrong, the part of a plumber's pay that is for turning up would start to go to the robot's owner. I think that's some years off.
Distribution. Owning the customer, through a brand, a client list or a platform. Cheap output floods every channel, so owning a trusted one is worth more, and every company building AI models wants a consumer product.
Compute and distribution mostly take capital, and both favour whoever is already large. That's the road to concentration, the third essay's future in which a few firms and governments hold the technology. Trust and presence are positions a person can take without much capital. So where customers have few alternatives, I expect owners to get most of the gain first, then the people who answer for the work and the people who are there.
What is getting cheaper
A draft, an analysis, a piece of code or a first read of a scan each used to cost a professional's hour. Epoch AI, a research group that tracks AI progress, found in September 2026 that the price of reaching a fixed level of AI performance has fallen about thirteenfold a year since 2023.
In the US, professional and business services, finance and insurance, and private education make up about 23 percent of GDP (BEA, via FRED). Across large rich economies the same sectors produce roughly a fifth to a quarter of output.1 Much of that work is composition.
William Baumol, an economist at Princeton and later New York University, explained in the 1960s why such work keeps getting dearer. A Beethoven string quartet needs the same four players and the same time it did in 1826, but the players' pay has to keep up with everyone else's. He called it cost disease (Chicago Booth Review). Now the part of these sectors that is composition is getting the productivity growth it never had.
Gains on single tasks are measurable. Erik Brynjolfsson, who directs the Stanford Digital Economy Lab, and two colleagues found that an AI assistant helped customer-support workers solve 14 percent more problems an hour on average, and novice and low-skilled workers 34 percent more. At Boston Consulting Group, consultants using GPT-4 finished 12 percent more tasks, 25 percent faster, at higher quality. But on a task picked to be beyond the model, they were 19 percentage points less likely to get it right. Cheap drafts still need someone to catch their mistakes, and that's where the cost can come back, as in the lease example's third ending.
Why the statistics don't show it yet
Economy-wide numbers barely show it yet. In February 2026 a group of economists published a survey of nearly 6,000 executives in the US, the UK, Germany and Australia. Nine in ten said AI had made no difference to their firm's employment or productivity over the past three years. Daron Acemoglu, an Institute Professor of economics at MIT who shared the 2024 Nobel prize, estimated in The Simple Macroeconomics of AI (2024) that AI would add no more than 0.66 percent over ten years to total factor productivity, the growth in output that more workers or machines don't explain. The first sign the other way came in March 2026, when Alex Imas, a professor of behavioural science and economics at Chicago Booth who keeps a running survey of the studies, wrote that the US totals were "showing signs of AI productivity gains".
In 1987 Robert Solow, an MIT economist who won that year's Nobel prize, wrote that "you can see the computer age everywhere but in the productivity statistics." The economic historian Paul David answered with electricity, which paid off only decades later, when factories were rebuilt around a small motor on each machine. I think AI's gains are stuck inside firms built around people doing the composition, and I expect new firms to do much of the rebuilding.
Jobs and pay
So far the clearest signal is who gets hired. On payroll data through June 2026, the Stanford Digital Economy Lab found that employment of 22-to-25-year-olds in the most AI-exposed jobs was about 19 percent below where it would be had it kept pace with peers in less-exposed jobs. It comes from fewer hires, and the authors say it shows a pattern without proving a cause.
Call centres show it most starkly. The data firm Revelio Labs reported on 6 October 2026 that global call-centre employment had fallen year on year for eight quarters running, again through less hiring. In the US, call-centre jobs were down 30 percent in August 2026 from November 2022, when ChatGPT came out (BLS). The cause is unproven. The US decline began in 2016, with offshoring and older automation, and in the Philippines call-centre employment rose every year to 2025.
Fewer people produce the work. Desk work is splitting into two jobs. One is producing the work, such as drafting and analysis, which machines make cheap. The other is answering for it, which gets more valuable. The partner in the lease example is worth more. But firms need fewer lawyers overall, partners included, because one person can answer for far more machine output than anyone could make by hand. In every ending of the example, the leases take fewer hours of lawyers' time than before (350 hours a month before AI, and 100, 250 or 175 after). I expect the change to show first in who gets hired and promoted, and in pay only years later.
Junior jobs go first, and the loss compounds. Seniors learn their skills as juniors. Matt Beane, who teaches technology management at the University of California, Santa Barbara, spent two years watching surgeons train as robotic surgery spread. With a robot the senior surgeon could operate alone, so residents got "ten to twenty times less practice", and many graduated without crucial skills. In The Skill Code (2024) he finds the same pattern in more than 30 professions. A profession that stops hiring 25-year-olds in 2026 has few 35-year-old experts in 2036. Each firm hopes to hire seniors someone else trained, so no single firm pays for the shortage.
Owners get a bigger share. The labour share is the part of what an industry produces that is paid to its workers. The rest goes mostly to owners and lenders. David Autor of MIT and four colleagues found that in US manufacturing the share paid out as wages and salaries fell from about 41 percent to about 25 percent between 1982 and 2012, mostly because sales moved to a few very productive "superstar" firms that pay out less of what they produce as wages. In US businesses outside farming, the labour share was 52.8 percent in the second quarter of 2026, the lowest since the series began in 1947 (BLS). A link to AI hasn't been shown. My forecast is that the share in law, finance, consulting and administration falls as it did in manufacturing, because more of the work is done by machines someone owns.
So will my salary go down? Not necessarily. The share is a ratio, so it can fall while pay rises, as long as output rises faster. If most of your work is composition and nobody needs you to answer for it, I'd expect slower raises and fewer openings before any cut. If you answer for the work, or own part of it, I'd expect you to do better. So far the evidence on pay is thin. A study of 25,000 Danish workers in exposed jobs found no effect of AI chatbots on earnings or hours after two years.
How far could it go? Anton Korinek, an economics professor at the University of Virginia, and Donghyun Suh modelled it in Scenarios for the Transition to AGI (2024), where AGI means AI that can do any task a person can.2 If there's always a long tail of tasks only people can do, wages can keep rising. If machines reach whatever ceiling human abilities have, wages fall toward what an equivalent machine costs.
Why firms replace workers instead of helping them. In The Turing Trap (2022), Brynjolfsson argues that building AI to imitate people steers it toward replacing them, when tools that let people do new things would create more value. The trap is about power as well as pay. While know-how lives in people, they can bargain. Once it's built into machines someone else owns, workers "lose economic and political bargaining power", and any sharing of the gains is up to the owners. US tax law tilts the choice too. Acemoglu and two colleagues put effective US tax rates above 28.5 percent on labour and about 5 percent on equipment and software. For a firm, the useful question about each AI project is whether it replaces someone on today's task or lets them do something new.
Why firms may get smaller while the biggest ones grow
Ronald Coase, a British-born economist who won the 1991 Nobel prize, asked in The Nature of the Firm (1937) why firms exist at all, if markets can coordinate everything. His answer was that using the market costs something. You have to find people, negotiate, write contracts and make sure they're kept. A firm grows until doing one more thing inside costs as much as buying it outside.
AI agents lower both costs. An agent is a model given a goal and some tools, which it uses on its own. My bet is that agents lower the cost of buying from outside more, because finding suppliers, describing the job and writing the contract are the kind of composition they're good at. Then firms buy more from outside, and what they keep inside needs fewer people. A two-person company could buy its bookkeeping, contracts, marketing and customer support from outside firms that run largely on agents.
At the other end the biggest firms grow, because compute and distribution get cheaper per customer the larger a firm is. So I expect many tiny firms, a few enormous ones, and fewer mid-sized firms in between. For workers that means more jobs in tiny firms, with more ownership and less security, or in giant ones, with less say.
Will the data-centre spending pay off?
About $720 billion of building at four companies in one year rests on a promise of transformation. So did railways in the 1840s, electricity around 1900 and fibre-optic cable in the 1990s. Many of those builders went bust, and the infrastructure paid off later, mostly for its users.
Much of the money goes round in a circle. Nvidia signed a letter of intent in 2025 to invest up to $100 billion in OpenAI as OpenAI installed Nvidia systems. Anthropic, which makes Claude, agreed to buy $30 billion of Microsoft's cloud computing while Microsoft and Nvidia agreed to invest up to $15 billion in it.3 In January 2026 about 45 percent of Microsoft's $625 billion of signed but undelivered contracts came from OpenAI. More of the money is borrowed, with more than $130 billion of bonds sold by the big cloud companies in 2026 by mid-September (NPR). Revenue at the companies that make the leading models is growing fast (TechCrunch), and the spending makes sense only if that growth lasts for years.
The third essay calls the future in which AI stops getting more capable the wall. If that happens, this is a repeat of the fibre-optic bust, and the losers are whoever financed it. If capability keeps improving and paying customers keep up with the building, it's more like the electrical grid. In both cases the data centres would still be there, even if the shareholders and lenders who paid for them lost money.
Taxes and the safety net
Rich countries raise roughly half their taxes from work. Across the countries of the OECD, the Organisation for Economic Co-operation and Development, social security contributions and personal income tax each raised about a quarter of tax revenue in 2023 (OECD).
If AI moves income from wages toward the owners of compute and distribution, as I expect, more of it ends up where tax is lighter, and taxes on work would pay for a smaller share of public services. At the same time more people may need help. Unemployment insurance was built for people who lose a job in a downturn and find a similar one when it ends. It does little for someone whose kind of job stops being offered. So two things may need to change: where tax falls, so that it depends less on wages, and what the safety net pays for. These are the main proposals, with the main objection to each.
- Tax capital and labour more evenly, as Brynjolfsson and Acemoglu propose, for example by ending fast write-offs for equipment. The objection is that it makes investment dearer.
- Progressive taxation, general or aimed at AI companies. Dario Amodei, Anthropic's chief executive, proposes it in The Adolescence of Technology (January 2026). It's hard to design well, as he concedes.
- A robot tax. Bill Gates, Microsoft's co-founder, proposed in 2017 taxing a robot that replaces a worker at about the worker's rate. A machine that replaces work is hard to tell from one that helps (Bruegel).
- Public ownership. Sam Altman, OpenAI's chief executive, proposed in 2021 an American Equity Fund, funded by a 2.5 percent yearly levy on large companies and private land and paid out to every adult. Senator Bernie Sanders proposed in June 2026 a one-time 50 percent tax, paid in shares, on large AI companies. Critics say a state can't fairly regulate a company it owns.
- Cash for everyone. Elon Musk, Tesla's chief executive, backs a "universal high income". In a three-year US trial of $1,000 a month, recipients worked one to two hours a week less. The objections are cost and less work.
- Wage insurance. It tops up pay for displaced workers who take a lower-paid job. Brookings proposed a federal version for AI displacement in September 2026, and a study of a US trade programme found it paid for itself. It only helps people who find new work.
None of these has yet been adopted as an answer to AI.
The promise of abundance, and where I differ
Several of the people running AI companies predict an age of abundance. Altman wrote in Three Observations (2025) that "the price of many goods will eventually fall dramatically". He also wrote that the balance between capital and labour "could easily get messed up, and this may require early intervention", and Amodei's The Adolescence of Technology warns of an unemployed or very-low-wage "underclass".
The obvious criticism is that the loudest forecasts of abundance come from people who sell the technology and need capital to build it. The writer Brian Merchant argues that "when OpenAI needs an infusion of cash, or wants to seal a deal, out come the promises of AGI." The same caution applies to Amodei, whose company also needs capital on this scale, and to me, since I build with AI. Acemoglu and his MIT colleague and fellow laureate Simon Johnson argue in Power and Progress (2023) that the gains from technology get shared only when institutions force it.
One way to judge it without taking sides is to sort the claims by what has been measured. The price of AI itself is collapsing. The prices households pay showed no sign of it by mid-2026. The Minneapolis Fed found "flat-to-rising" software prices since generative AI arrived, and the Richmond Fed found the AI build-out pushing inflation up.
Mathematics may show what's coming. On 6 October 2026 OpenAI posted 719 AI-written manuscripts from an unreleased model, about 42 percent of their main results verified by Lean, a program that confirms each step of a proof is valid. Nature reported mathematicians "up in arms", reactions split (Fortune), and OpenAI withdrew at least three of the solutions within days. It's the lease example on the scale of a whole field. Writing the proofs got cheap, and finding out whether they were right became the slow part.
I differ from the abundance essays on three things.
What gets cheap. Cheap composition lowers the price of whatever is mostly composition. Everything else gets dearer by comparison, such as a person's time, a licence, care, and land near people and power. Pedro Santa-Clara, a finance professor at Nova School of Business and Economics in Lisbon, argues that this is how the gains reach workers, because the human tasks that remain get paid more. That holds for whoever does them. Someone whose work was the composition now sells something cheaper, and still pays for a nurse, for a lawyer to answer for a contract, and for rent. None of this says property is a good investment. A thing getting dearer isn't a reason to buy it, since the price may already reflect that.
Order. Several of the proposals for sharing the gains end in a payment. I expect the losses to start years before any payment does, with junior hiring. A payment can replace lost income, but it can't give someone back the years as a junior in which they would have become an expert.
Timing. Altman's fund and Amodei's progressive tax both assume governments can tax the gains once they arrive. The second essay argues that governments tend to act late, after something has gone wrong. Meanwhile I think who owns the compute and the customers is being settled now, before most of the gains arrive, and ownership is hard to change once it's settled. That worries me more than how big the gains will be.
What to do
None of this is investment advice. I'm not qualified to give it.
If you work for a living or run a business, the fifth essay has the detail. In short, move toward work you answer for, such as a licence, a client who asks for you by name or a decision with your name on it. Where you can, own part of the work you do, such as a client list, a share in the firm you work for or a small business of your own, since I expect more of the gains to go to owners. That kind of ownership can lose value too. And if you hire, keep training juniors, or in ten years your profession runs short of experts.
Questions to ask a minister, a chief executive or a board.
- Skills. Where will the next experts come from if firms hire fewer juniors, and who pays to train them?
- The safety net. Is it built for a kind of job disappearing, or only for downturns?
- People. What share of the work is composition, and what is the plan for the people doing it?
- Tax. How will public services be paid for if pay falls as a share of the economy?
- Ownership. Will ordinary people own any of the gains, through pensions, funds or shares?
- Infrastructure. Who owns the compute the country runs on, and who pays for the power?
- Who decides. Are you one of the people deciding this? If not, who is, and can the rest of us see what they decide?
Open questions
- Do the gains show up fast or slow? Several years of productivity growth clearly above trend, and executives reporting effects in their own firms, would settle it.
- Does new demand hold wages up? It did for accountants when spreadsheets arrived, and didn't for switchboard operators. The labour share in professional services, and young people's hiring in exposed jobs, will tell.
- Is the data-centre spending more like the electrical grid or the fibre-optic bust? AI revenue catching up with the spending would point to the grid. The four biggest spenders' combined spending falling year on year, or lenders charging much more for data-centre debt, would point to the bust.
- When can a robot handle a stranger's house? Confirmed deliveries, and independent tests of robots doing chores without a remote operator, would tell.
- Do governments change tax and the safety net in time? History says they usually act late. A tax commission, or wage insurance for AI displacement in a major economy, would be the first sign otherwise.
When a sector's composition gets cheap
People ask when the singularity is coming, meaning the point at which AI overtakes us, as if it were one date for everyone. The first essay argued that for an individual it's the day the thing you're best at becomes cheap. For a sector, it's the day its composition becomes cheap and the value it held moves to whatever is still scarce. For most desk work that day falls inside this decade. So the useful question, for a person, a firm or a government, is how much of what it does is composition, and how much rests on compute, trust, presence or distribution. Moving toward those takes years, and it can start now.
That winter I brushed my teeth in the kitchen. (The Empty Search)
- Sums of OECD national accounts for finance and insurance, professional and technical services, administrative services and education (OECD Data Explorer). The UK comes to about 28 percent. The OECD doesn't publish this total itself. ↩
- Korinek has been a member of Anthropic's Economic Advisory Council since April 2025. Anthropic makes Claude, which this essay was written with. ↩
- This essay was written with Claude. ↩
New writing by email: subscribe to Aperture/I, my newsletter.
Share this essay: ·
Think I'm wrong? Follow on X · LinkedIn · RSS
Ask this essay
Ask about this essay, or about me and the site. Answers come only from the essay and the site, and questions are logged without names so the answers can get better.
Colophon. Drafted 19 September 2026 with Claude. Numbers: 2026 capital spending of about $720 to $740 billion at the four largest cloud companies (their guidance after the third quarter of 2026); Stanford Digital Economy Lab employment data through June 2026; a February 2026 survey of nearly 6,000 executives; US Bureau of Labor Statistics figures for the second quarter of 2026. The lease example is illustrative. Not investment advice. Revised since with Claude and other AI reviewers, most recently on 11 November 2026.
Sources · 29
- On the price of AI: Epoch AI, The plunging price of thought (Luke Emberson and David Roodman, September 2026).
- On the size of the sectors: US Bureau of Economic Analysis, value added by industry as a share of GDP, second quarter of 2026, via FRED; OECD national accounts, Table 6A (accessed 11 October 2026).
- William J. Baumol, "Macroeconomics of Unbalanced Growth" (American Economic Review, 1967), for the cost disease, with Chicago Booth Review (May 2017) and Princeton's memorial notice.
- Erik Brynjolfsson, Danielle Li and Lindsey Raymond, Generative AI at Work (NBER Working Paper 31161, April 2023, the version whose figures are used here; the version published in the Quarterly Journal of Economics in 2025 puts the average gain at 15 percent), and Brynjolfsson's Stanford profile; Fabrizio Dell'Acqua and colleagues, Navigating the Jagged Technological Frontier (HBS Working Paper 24-013, September 2023).
- Ivan Yotzov, Jose Maria Barrero, Nicholas Bloom and colleagues, Firm Data on AI (NBER Working Paper 34836, February 2026, revised March 2026).
- Daron Acemoglu, The Simple Macroeconomics of AI (NBER Working Paper 32487, May 2024; full text); on his Nobel prize, CEPR (14 October 2024).
- Alex Imas, What is the impact of AI on productivity? (January 2026, updated March 2026); his Chicago Booth profile.
- Robert M. Solow, "We'd Better Watch Out" (The New York Times Book Review, 12 July 1987, page 36), scan of the page, with Brad DeLong's transcription; on his Nobel prize, nobelprize.org.
- Paul A. David, "The Dynamo and the Computer" (American Economic Review, May 1990), via AcaWiki.
- Stanford Digital Economy Lab, Canaries in the Coal Mine, August 2026 update.
- On call centres: Revelio Labs, After a Decade of Growth, Global Call Center Employment Is Shrinking (6 October 2026); US Bureau of Labor Statistics, all employees in telephone call centres (series CEU6056142001, through August 2026); Fortune on the Philippines (August 2026).
- On robots for the home: 1X, NEO order page (accessed 11 October 2026); BotInfo (5 October 2026); Figure, Helix 2.5 in 30 unseen homes (17 September 2026).
- On capital spending: Yield Theory's hyperscaler tracker (updated 5 November 2026), summing company guidance after the third quarter of 2026; Fortune on Amazon's demand (30 July 2026).
- Matt Beane, "Shadow Learning" (Administrative Science Quarterly, 2019), via UC Santa Barbara (August 2020), and The Skill Code: How to Save Human Ability in an Age of Intelligent Machines (Harper Business, 2024), via UC Santa Barbara (April 2024).
- David Autor, David Dorn, Lawrence F. Katz, Christina Patterson and John Van Reenen, The Fall of the Labor Share and the Rise of Superstar Firms (Quarterly Journal of Economics, 2020).
- US Bureau of Labor Statistics, Productivity and Costs, Second Quarter 2026, Revised (USDL 26-1434, 3 September 2026), for productivity, real pay and the labour share.
- Anders Humlum and Emilie Vestergaard, Large Language Models, Small Labor Market Effects (NBER Working Paper 33777, 2025).
- Anton Korinek and Donghyun Suh, Scenarios for the Transition to AGI (March 2024; also NBER Working Paper 32255); Korinek's homepage; Suh on RePEc.
- Erik Brynjolfsson, The Turing Trap: The Promise & Peril of Human-Like Artificial Intelligence (Daedalus, Spring 2022).
- Daron Acemoglu, Andrea Manera and Pascual Restrepo, Does the U.S. tax code favor automation? (Brookings Papers on Economic Activity, March 2020).
- On the circular deals, the backlog and the debt: Nvidia (22 September 2025); Microsoft on Anthropic (18 November 2025); The Register on Microsoft's backlog (29 January 2026); NPR's The Indicator, via WUNC, on bonds (18 September 2026); TechCrunch on Anthropic's revenue (17 August 2026).
- Ronald Coase, The Nature of the Firm (Economica, November 1937); on his Nobel prize, nobelprize.org.
- On the tax base: OECD, Revenue Statistics 2025 (December 2025); Bruegel on taxing robots (March 2017); Sam Altman, Moore's Law for Everything (March 2021); AP, via News4JAX, on Senator Sanders's proposal (17 June 2026); Fox Business on Elon Musk and universal high income (April 2026); Eva Vivalt and colleagues on the OpenResearch guaranteed-income trial, NBER Working Paper 32719 (July 2024, revised August 2026); Guy Ben-Ishai and Neil Thompson on workforce policy for AI (Brookings, 15 September 2026); Benjamin Hyman, Brian Kovak and Adam Leive on wage insurance, NBER Working Paper 32464 (May 2024).
- Sam Altman, Three Observations (February 2025).
- Dario Amodei, The Adolescence of Technology (January 2026).
- Brian Merchant on Sam Altman's essay The Gentle Singularity, Blood in the Machine (18 June 2025); Daron Acemoglu and Simon Johnson, Power and Progress (PublicAffairs, 2023).
- On prices: Federal Reserve Bank of Minneapolis, Jeff Horwich on AI, productivity and prices (3 August 2026); Federal Reserve Bank of Richmond, John O'Trakoun's Macro Minute on AI and inflation (30 June 2026).
- On mathematics: OpenAI's math repository (6 October 2026); Nature (7 October 2026); Fortune (7 October 2026).
- Pedro Santa-Clara, The Weakest Link and the Oldest Theorem (August 2026); his Nova SBE profile.
Changes · 7 since publication, 10 drafts before publication, latest 11 Nov
- 11 Nov 2026 · Shortened from about 6,100 words to about 4,800, and reordered so the answer comes first. The opening now says who keeps the savings, and the law-firm lease example follows straight away. The four scarce things and where the value goes are one section. The first signs in hiring and the Turing Trap are now part of Jobs and pay. Cut: the passage on what repeated harm from AI agents would do to data-centre spending, the paragraph on two economists who expect slow gains, and several quotations in the sections on data centres, taxes and abundance. The customer-support study is quoted from the version the essay links (14 percent on average, 34 percent for novice and low-skilled workers). Capital spending is given on the four companies' guidance after the third quarter of 2026. Dates are written out in full, and the subtitle is new.
- 11 Oct 2026 · Revised from my 29 reading notes. Evidence for each economic claim, with who every economist is; the lease example's terms and columns explained, and its figure redrawn as stacked bars; call-centre hiring; what a falling labour share means for pay; new sections on the Turing Trap, the build-out in the loop, the shrinking firm and tax options; abundance weighed even-handedly, with its critics and OpenAI's mathematics papers; home robots; plain questions for leaders.
- 7 Oct 2026 · Composition now means the making of thought at every stage, not just the first version, as in the first essay.
- 5 Oct 2026 · More personal and about a seventh shorter. It opens on a renovation that needed someone in the house and has a new passage on accountability at the top of the market: that value goes to fewer people, I think. The partner is worth more, but fewer partners are needed, because one person can now answer for far more machine output than anyone could make by hand. Every figure, study and quote is unchanged.
- 30 Sep 2026 · Marked as beta while I revise it.
- 27 Sep 2026 · Altman's and Amodei's positions are quoted rather than characterised. Fewer juniors being hired is described as a pattern in the data rather than a proven cause. The investment section is now 'Themes I'm watching'.
- 27 Sep 2026 · Line edit for style: fewer colons, set-piece contrasts and fragments, plainer sentences. The argument, facts, numbers and sources are unchanged.
10 drafts before publication, 19 to 24 September 2026. AI models reviewed and fact-checked the drafts.