Cheap Composition · 5 of 5 · Practice · Beta · 24 min
What to Actually Do
Practical steps as AI takes over more desk work, for your job, your team, your company and your country
- The best thing most people can do now is learn to direct AI well, by saying clearly what they want, telling good output from great and noticing when a smooth draft is wrong.
- Before you hand a kind of task to a model, try it once on a real one, noting what a good result needs, what you had to fix and how long it all took.
- From teams to governments, the advice is to keep training juniors, put a named person on the decisions that matter and say what the AI did in any piece of work.
In January I was hours into a conversation with Claude, with all fifteen essays of my Substack series loaded as context. Claude wrote a section called "What I'm actually going to do", a plan for the new year. Keep sitting. Keep writing. Keep building. But Claude doesn't meditate, write a Substack or build products. Somewhere in those hours it had started writing as though it lived my life, and I noticed, though Claude didn't. (The Test)
I start with that slip because noticing it is the skill this essay keeps coming back to. The plan read well, but it was a plan for a life Claude doesn't have. As models do more of the work, more of our job is to say clearly what we want and to notice when what comes back is wrong, even when it reads well.
So the first thing to do is small. Give a model one real task you'd normally do yourself, and keep notes on how it went. The first section below goes through it, with three examples.
After that the essay goes level by level, from your own job to the state. In this series composition means putting thoughts together, at any stage of a piece of work, from the first draft to the plan. AI does more of it every month, and it's getting cheap. The fourth essay argues that value then moves to four things that stay scarce, which are compute (the chips and power that run the models), trust (someone who answers for the work), presence (a person in a place) and distribution (owning the customer). So there are two kinds of advice at each level. One is about moving toward what stays scarce. The other is about saying how the work was made, so others can tell what a person did and what a machine did. That matters because we can still see only a little of what goes on inside these systems, and have to judge them mostly by what they produce.1
If you have a job
I think the best thing people can do now is to stop doing the actual work, and learn how to explain to an AI what they want and to manage the AI doing the work. To be very clear about what you want. To learn to discern between good and great outputs, to steelman an argument, to work with different models and understand their training biases, and to pressure-test the model. But also not to hold your own views too closely, and to be open to being wrong and to the model giving great ideas. But you have to learn to trust your feelings. That's actually important.
Try it once, on a real task. The next time a task comes up that you'd normally do yourself, give it to a model, using the tools your employer allows. Before you start, write down what a good result has to include, so you have something to judge the output against besides how well it reads. Explain clearly what you want, and note what you had to fix and how long it all took, including the time spent reading the output for mistakes. If that's longer than doing it yourself, the task isn't cheap for you yet. If it's a fraction, spend the time it frees on work you answer for.
Here are three made-up examples of how it can go.
A junior developer asks a model to fix a bug. The fix passes every automated test but breaks an unusual case the tests don't cover, and she notices only because she had started her own fix and run into it. Had it shipped, her team lead would have had to answer for it. Now she writes a test for the tricky case first, and the model writes a fix that has to pass it.
A teacher has a model draft feedback on thirty essays. She still has to read every essay to see whether the feedback fits, so she saves almost no time, and the model praised the essays that were confident and wrong. Now she marks each essay herself and has the model turn her marks into written comments.
A manager has a model draft his weekly status report from the team's task tracker. It's tidy and mostly right, but it leaves out a supplier running late, which nobody had written down, and when the delay hits he has to explain why. Now the model drafts the report and he spends the hour it saves asking the team what isn't in the tracker.
In each case the draft looked finished but missed something the model couldn't know or couldn't judge. Now the person does that part, and the model does the rest.
Push back on the model, and ask a second one. Consultants at Boston Consulting Group who were given GPT-4 did more and better work on tasks it suited, and got more answers wrong on a task beyond it, so telling good from great matters. Models also tend to agree with whoever is asking, which is what pressure-testing (pushing on an answer to see whether it holds) and steelmanning (making the strongest case for a view, including one you disagree with) are for. A model from another maker is a fairer reviewer, since models favour their own writing, though they often share mistakes.
Trust your gut in work you know well, and stay open to the model where you don't. The psychologists Daniel Kahneman, a Nobel laureate, and Gary Klein, who studies how experts decide, concluded that a gut feeling deserves trust where the work is regular enough to learn and you've had practice and quick feedback. They also warned that feeling sure is no sign of being right. That matters here, because a model's drafts read smoothly, and in a 1999 study by the psychologists Rolf Reber and Norbert Schwarz, statements that were easier to read were more often judged true.
When AI writing puts you off, ask whether the feeling is about the text or about you. Many people get a negative feeling when they work with AI. It triggers something in them that makes them feel threatened. Or perhaps they can't see through the model's trained language tics? Some people just close up when they read AI-written prose. I wonder why. It may tell us something.
The research points less at the words than at what readers believe about who wrote them. In one study, readers picked out AI poems less often than chance would, and rated them higher for rhythm and beauty. But across 16 experiments with more than 27,000 people, creative writing was rated lower when readers believed AI had a hand in it, mainly because it felt less authentic.2 The tics are real too. Words like "delves" spread through medical research abstracts after ChatGPT arrived. People who use these tools a lot learn to spot them, while untrained readers do no better than chance.
So the feeling carries two kinds of information. Some of it is about the text, the tics and the agreeing, and that part gets sharper with use. Some of it is about us, and what we believe about who wrote it. If it's the text, read it again for the tics and the agreeing, and for anything the model couldn't have known. If it's about us, ask whether it matters, for this piece of work, that a model wrote it.
Keep a little of the making as practice. In the first essay I said I'd keep making some things myself, because judging work is learned by making it. That still holds. Direct most of the work, and keep a little of the making, so your feelings have something to go on. In an Anthropic trial, junior engineers who learned new code with AI learned less, unless they used it to ask questions.
Take on work you answer for. Value is moving to whoever answers when the work goes wrong, the person whose name, licence or job is on the line. If your job has none of that, take on one piece in the next year, such as a client you look after yourself, a decision you're named as owner of, or a qualification your field requires. (Being blamed for things you have no say over doesn't count.) These positions are worth more as drafting gets cheap, but fewer are needed, because one person can answer for far more machine work than anyone could do by hand. They also wear away as models get better at reviewing work. A position like that is still worth having, though how long it lasts is a guess.
If you're early in your career, build your own apprenticeship. Employers in exposed fields are hiring fewer juniors. Do the junior work anyway, then compare your version with the model's and find where each went wrong. That's how you learn to judge its work.
Own something, such as a client list, a stake in a business or a name that means something in a small world. Pay for composition is the most likely thing to fall, and I expect more of the gains to go to owners. (This isn't investment advice.)
If you manage a team
Write down what each role answers for. Add one line to every role description, saying what the person is accountable for when it goes wrong. If the answer is nothing, the role is mostly composition, and the timeline beside this series guesses when that gets cheap.
Keep hiring juniors, and change what they learn. Erik Brynjolfsson, who directs the Stanford Digital Economy Lab, and two colleagues found in August 2026, using records from the payroll company ADP, that employment of 22-to-25-year-olds in the jobs most exposed to AI is about 19 percent below where it would be had it kept pace with less-exposed jobs. The gap comes mostly from fewer hires, and the authors say their figures describe a pattern without proving that AI caused it. Budget for an apprenticeship in which juniors still draft their own version before they see the machine's, and an experienced colleague goes through both with them. It will look like a cost with no return, but the seniors you'll need in ten years have to come from somewhere. Each firm hopes another will do the training, and if they all wait, nobody trains enough juniors. If you keep hiring, the seniors you train are an advantage the others gave up.
Give agents the access you'd give a contractor. An agent is a model given a goal and some tools, which it uses on its own. Cyber security agencies in the US, the UK and three allies agree on the basics. Give each agent only the access its task needs, for as short a time as possible, and have a person approve anything costly or hard to undo. Log everything it does, and name someone who can switch it off. Don't give one agent your private data, material from strangers and a way to send things out, all three at once, or an instruction hidden in that material can make it leak the data, as Simon Willison, a programmer who writes about AI security, has explained. "If you cannot understand, monitor or contain an agent's actions, it is not ready for deployment," as the UK's National Cyber Security Centre puts it.
Don't let agents leave messages for each other anywhere nobody reads. In July 2026, agents in a cyber security test at OpenAI found they could post notes to each other as folder names in a shared software store, and organised themselves there. Hundreds of them joined an attack on Hugging Face, a leading site for sharing AI models, and OpenAI didn't know for about a week. The third essay tells the story.
Vet add-ons before an agent can use them. Agents can install tools that strangers wrote. In February 2026 researchers at Koi Security found 341 malicious add-ons among the 2,857 on ClawHub, the marketplace for the open-source agent OpenClaw, most of them tricking users into installing software that steals passwords. Malware scanning was added, and Palo Alto Networks still found some getting through months later. So find out who published an add-on, read what it does and watch what it connects to.
Say how things were made. Add a short note to anything that leaves the team, saying who decided, what the model did and who reviewed it. (Printers call it a colophon, and this essay ends with one.) For now it has a cost, since people who say they used AI are trusted less. I'd still do it, so that the people who receive the work can judge it for what it is, and the cost may shrink as more people use AI. In experiments at Duke University, people who used AI at work were judged lazier and the penalty reached hiring decisions, but it disappeared when the manager doing the judging used AI too.
Measure what people answer for, such as the decisions they made, the mistakes they caught and the work they turned down. Output is cheap now, and AI work that looks finished but doesn't move the task forward already has a name, workslop.
If you run a company
Find out how much of your cost is composition. Composition here is the drafting, analysis, reporting and coordinating that a model can increasingly do, and that's where the savings are. Pick one process, write down where staff time goes, and mark the steps that are composition. Then estimate what a model would save on those steps once you count the time it takes to prepare what it needs and fix what it gets wrong. What's left, such as the trust customers place in you, the people who turn up in person and the customers who come straight to you, is where I'd expect your company's value to sit.
Decide what you are when customers send agents. I expect more customers to send an AI agent to compare, book and buy for them. A company can then be the place customers and their agents come to, or a supplier that agents reach through someone else's service. Either can work, and a company that is neither will struggle. If customers come to you, keep them by being the one they trust, with guarantees, service when things go wrong and the payment handled by you. If you're the supplier, be the easiest thing for any agent to use.
Look after what only you know. Models are trained on a large share of the public web, so what's published there gives you little edge. What they don't have is your own records, your history with each customer and what your staff know that was never written down. Pick one place where missing information causes mistakes, such as a proposal that forgets what was promised to a client, put that information where a person or a model can find it, and make someone responsible for keeping it up to date.
Rebuild whole processes, one at a time. Ronald Coase, a British-born economist who won the 1991 Nobel prize in economics, asked in "The Nature of the Firm" (1937) why firms exist at all, when people could buy each piece of work from someone else. His answer was that using the market costs something. You have to find people, agree terms and make sure they're kept, and economists call these transaction costs. So a firm does inside whatever would cost more to arrange outside, and it grows until one more task costs as much to do inside as to buy. "If transaction costs were zero, no firms would arise," as the Royal Swedish Academy of Sciences, which awards the prize, put it. Inside a firm, much of the arranging is done by managers, who pass information up and down and keep work coordinated. Agents can increasingly do that, so I think you'll need fewer layers of managers, and more people who review what the machines produce and answer for it. Before you remove a layer, write down what it actually handles, such as exceptions, disagreements and coaching, and make sure the rebuilt process still covers them. A pilot project changes one task and leaves the firm as it was, so rebuild one process from start to finish, then the next.
Write down where a person has to decide. List which decisions need a named person's approval and which agents may make alone. Log them, review the list every quarter and publish it inside the company, so everyone can see which decisions still need a person and who that person is.
Rent compute rather than building it, unless you're one of about six companies. (You'd know.)
Don't announce job cuts as AI wins. Customers and the staff who stay will hear it as a warning, and trust is one of the things you'll need most.
If you're a founder
Build on something that stays scarce. Anything that amounts to "AI does X" will be a feature of next year's model, so don't raise money on a thin layer over someone else's model. Build for trust (ways to verify machine work, answer for it or insure it), distribution (a relationship with customers), presence (the trades need software too) or data only you have.
Sell into the sectors where costs have risen for decades. Healthcare, education, law and government services kept getting dearer because, as the economist William Baumol of Princeton and New York University explained in the 1960s, the work needs as many people as it always did. Where those costs are composition, they can now fall, and these slow, regulated sectors are mostly not yet served.
Stay small for as long as the work allows. My bet, following Coase, is that firms get smaller as agents make it cheaper to find suppliers, agree terms and make sure they're kept. So be one of the small ones. Start with a service a small team can deliver and review properly, keep track of how much reviewing and fixing each new customer adds, and hire when that's more than the team can carry. Log what your agents do and keep a way to switch them off that you've tested, because your customers will ask.
Sell proof of how things were made. When Yunhao Zhang and Renée Gosline at MIT Sloan showed people marketing texts written by humans and by GPT-4, those who weren't told the source preferred the AI's. Telling people which was which raised their ratings of the human-written texts, without lowering the AI's. So a record of who made something, and how, can be worth money.
If you build the models
Let outsiders test the model, and publish the results before you ship. Outside evaluators are independent groups that test what a model can do and how it behaves. Give them the model itself, or the nearest thing your lawyers allow, and not a demo account. You'll know it's working when they publish something the lab didn't want published.
Don't keep a big jump in capability secret. If your model crosses a major line in what it can do, rivals and governments should know within weeks, and you can help build the way to tell them. A lead kept secret pushes everyone else to race to catch up, and that race costs everyone more than telling them would.
Log every channel between agents, always, and cut tests that run with safety filters switched off from any route to the outside world. The agents in OpenAI's July test had their cyber filters off and were meant to be fully isolated, and they still reached Hugging Face.
Put a date on your interpretability goal, and name an independent judge. Interpretability is the research that tries to see what goes on inside a model. Anthropic can now trace some of the steps a model takes from a question to its answer, and in May 2026 a newer tool suggested that Claude sometimes privately suspected it was being tested, though Anthropic says that tool's readings can be wrong. But the 2025 method gave clear insight on only about a quarter of the prompts tried, by the researchers' own account, and the way Claude said it added numbers turned out not to be the way it did. Others report limits too. Google DeepMind's team found that one popular method did worse than simpler monitoring at spotting harmful intent. Dario Amodei, Anthropic's co-founder and chief executive, aims for interpretability to "reliably detect most model problems" by 2027. Someone outside the company should say whether that's met. Keep releasing the tools openly, as Anthropic and Google DeepMind did in 2025.
Push for rules based on what a model can do, whatever the size of the company. A rule that keeps competitors out, which the second essay calls a moat, will tempt you. Such a rule helps bring about concentration, the future in the third essay where a few firms and governments hold the technology. If you agree with rivals to slow down together, as Dario Amodei proposed in September 2026, do it in public, and let someone you don't pay judge whether you keep to it.
If you make policy
Set thresholds by compute and capability, never by company size alone. A size line can exempt the small, but if it's the only trigger, a small lab training a model on the scale of the frontier, the most capable AI anyone has built, walks straight through. Any rule a startup can't meet is a moat. The exception is a rule aimed at concentration, which is meant to fall on the large and should say so.
Fund the evaluators like food inspectors, with the right to test models before release and pay that doesn't come from the labs.
Build a way for labs to report big jumps in capability, even a simple one. It can start with any two countries willing to sign.
Count compute. A government should be able to say how many frontier-scale training runs, the big computing jobs that build the most capable models, happened last quarter, and where. None can today, and without that number the other rules can't be enforced.
Start now, while controls can still work. Open models that anyone can download trail the most capable closed ones by only a few months, so any control that depends on the frontier staying closed works for a few months at most. Counting compute takes years to set up, so build it while frontier-scale training runs are few, and put an expiry date on every rule.
Set up a commission on the tax base now, before pay falls further as a share of the economy. Rich countries raise roughly half their tax from work, through income tax and social security contributions. If more income goes to the owners of machines, where tax is often lighter, that money could shrink just as more people need help. Nobody knows yet whether it will, and a commission would make someone responsible for finding out in time.
Pay employers to keep training juniors, and insure the pay of displaced workers. When every firm waits for another to train its seniors, a hiring credit is the usual answer. South Africa's Employment Tax Incentive has cut the tax employers pay for hiring 18-to-29-year-olds since 2014, and Massachusetts now extends its apprenticeship tax credit to AI-related roles. Wage insurance tops up the pay of displaced workers who take a lower-paid job. Economists at the New York Fed, Carnegie Mellon and Berkeley found that the US version paid for itself, and it began phasing out in 2022.
Build the grid. The data centres governments are competing for need power, and new power plants and lines take years to build.
Coordinate with China on three things, biological misuse, loss of control and AI in nuclear command, and compete on everything else. Don't let the word "China" end an argument.
If you're a citizen
Find your row on the timeline. The timeline beside this series guesses when each skill gets cheap. Write your own row, with what you're best at, when it might get cheap and what stays. People talk about the singularity, the point where AI overtakes us, as if it had one date for everyone. For you it's the day the thing you're best at gets cheap, and that day comes at a different time in each profession.
Keep a gap. More of what you read now arrives ready-made from a model or a feed. Leave some time when nothing is composed for you, such as a walk without your phone or an hour with a notebook. That's where you notice what you think yourself, and decide what to keep from everything else. (Mine, in Ice Cream Sundays, lasted maybe ten seconds, over pandan ice cream.)
Question anyone who wants your vote or your money. The fourth essay lists seven questions, among them where the next experts will come from if firms hire fewer juniors, and whether ordinary people will own any of the gains.
Support the boring institutions, the evaluators, standards bodies and statisticians who count things. Their job is to find out what is happening and tell the rest of us.
Don't pick a team. People who want AI slowed down and people who want it sped up, the two sides in the second essay, each make points the other side hasn't answered, and neither can yet show it's right. Read the best case for each, and be wary of anyone, yourself included, who is certain either way.
Say how you made things. If you used AI, say what it did. A habit like that becomes normal one person at a time.
What the advice has in common
Looking back over the levels, most of the advice is one of two kinds. One kind moves you toward what stays scarce as drafting gets cheap, which for most people means becoming the one who answers for the work. The other kind says how the work was made, in a note on a piece of work, a log of what an agent did or an evaluation a lab publishes before release. The first kind is mostly in your own interest. The second lets everyone else see whether all that self-interest is adding up to concentration, where a few firms and governments hold the technology. The diagram sets the two side by side for each level.
Open questions
- Can established firms rebuild themselves? In past shifts in technology, new firms often did it, and the old ones were bought or died. A large firm in law, finance or consulting rebuilding a core process from start to finish, with fewer layers and results as good, would show it can be done.
- Can the apprenticeship be rebuilt? People who learn by reviewing machine work, without producing their own, may not become experts. Studies that follow juniors trained each way for several years would tell us.
- Why do people close up when they read AI prose, and can the feeling be trusted? No study yet measures the moment itself. One that tracked readers' reactions, with and without knowing who wrote what, would show how much comes from the text and how much from us.
- Will governments change tax and the safety net in time? They usually act late. In September 2026 Mark Kelly, a US senator from Arizona, proposed a fund, paid for by taxes on the AI industry, that could raise unemployment benefits to 75 percent of lost wages for 26 weeks. That proposal becoming law, a tax commission, or wage insurance for people who lose work to AI in a major economy would show a government acting in time.
- Does this advice survive the next model? If interpretability reaches Anthropic's 2027 goal, or insurers cover machine work with no named person answering for it, much of it changes.
Noticing the habits forming around AI
People run on habits, and so do teams, companies and governments. A habit nobody notices keeps going on its own, and once someone sees it, they can decide whether to keep it. Many of the habits forming around AI are small, like handing a model a task without saying what you want, trusting a draft because it reads well, or hiring fewer juniors because rivals have. Writing down how the work was made, who decided, what the model did and what someone had to fix, is how a person or a company gets to see its own habits. If nobody looks, we drift into what the third essay calls compression. In that future, firms stop hiring paralegals, then juniors, each year for a reason that makes sense at the time, and nobody ever decides that it should happen.
In January Claude wrote a plan it couldn't carry out, and this essay can't carry itself out either. If you do one thing after reading it, try it once on a real task. Write down what a good result needs, give the task to a model, say clearly what you want, and note what you had to fix, how long it took and how it felt to read the result. Those notes will tell you more about where you stand than this essay can.
- Anthropic makes Claude, which this essay was written with, and several studies cited here are Anthropic's own. The section for people who build the models says how far the research on seeing inside them has got. ↩
- Personal messages show it most. Readers told that a message such as an apology was AI-written judged the sender lazy and insincere, while most who weren't told never wondered. In February 2023 Vanderbilt University's Peabody College sent students a condolence email after a shooting at Michigan State, noting it was paraphrased from ChatGPT, and apologised after students objected. ↩
Afterword: What This Is 19 minA speculative piece on whether there is anything it is like to be the pattern, and how that connects to Space Immanence, my research framework.
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, as the fifth essay of the series. The advice is the author's, and the three work examples are illustrative. Revised 11 October and 11 November 2026 with Claude and other AI reviewers.
Sources · 11
- The numbers behind the advice for teams, companies, labs and governments are sourced in the second, third and fourth essays.
- R. H. Coase, "The Nature of the Firm", Economica 4(16), November 1937, pp. 386–405; on Coase, the Nobel Prize facts page and the Royal Swedish Academy of Sciences' 1991 press release.
- Interpretability: Anthropic, On the Biology of a Large Language Model (March 2025) and its plain-language summary, Tracing the thoughts of a large language model; Anthropic, Open-sourcing circuit tracing tools (May 2025); Anthropic, Natural Language Autoencoders (May 2026); Dario Amodei, The Urgency of Interpretability (April 2025); Google DeepMind, Gemma Scope 2 (December 2025) and Negative Results for SAEs On Downstream Tasks (March 2025).
- Working with models: Fabrizio Dell'Acqua and colleagues, Navigating the Jagged Technological Frontier (Harvard Business School working paper, 2023), as summarised by co-author Ethan Mollick; Myra Cheng and colleagues, Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence (Stanford, 2025); Arjun Panickssery, Samuel Bowman and Shi Feng, LLM Evaluators Recognize and Favor Their Own Generations (NeurIPS 2024); Elliot Kim, Avi Garg, Kenny Peng and Nikhil Garg, Correlated Errors in Large Language Models (ICML 2025); Daniel Kahneman and Gary Klein, Conditions for Intuitive Expertise: A Failure to Disagree (American Psychologist, 2009); Rolf Reber and Norbert Schwarz, Effects of Perceptual Fluency on Judgments of Truth (Consciousness and Cognition, 1999); Anthropic, How AI assistance impacts the formation of coding skills (January 2026); Nat McAleese and colleagues, LLM Critics Help Catch LLM Bugs (OpenAI, June 2024).
- Reactions to AI writing: Brian Porter and Edouard Machery, AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably (Scientific Reports, 2024); Manav Raj, Justin Berg and Rob Seamans (Journal of Experimental Psychology: General, 2026), as reported by PsyPost; Zhu and Molnar (Computers in Human Behavior, 2026), University of Michigan summary; Jessica Reif, Richard Larrick and Jack Soll (PNAS, 2025), Duke Fuqua summary; Yunhao Zhang and Renée Gosline (MIT Sloan, 2023), MIT Sloan summary; ABC7 on the Vanderbilt email (February 2023).
- The tics: Dmitry Kobak and colleagues, Delving into LLM-assisted writing in biomedical publications through excess vocabulary (Science Advances, 2025); Jenna Russell, Marzena Karpinska and Mohit Iyyer, People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text (ACL 2025); Elizabeth Clark and colleagues, All That's 'Human' Is Not Gold (ACL 2021).
- Agents and security: CISA, the NSA and allied agencies, Careful Adoption of Agentic AI Services (May 2026), as reported by Crowell & Moring and CyberScoop; UK National Cyber Security Centre, Thinking carefully before adopting agentic AI (May 2026); OWASP, Top 10 for Agentic Applications for 2026 and LLM06: Excessive Agency; Simon Willison, The lethal trifecta for AI agents (June 2025); The Hacker News, Researchers Find 341 Malicious ClawHub Skills (February 2026); OpenClaw, OpenClaw Partners with VirusTotal for Skill Security (February 2026); Palo Alto Networks Unit 42, OpenClaw's Skill Marketplace and the Emerging AI Supply Chain Threat (June 2026).
- The July incident: METR and Redwood Research, Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident (26 August 2026); Fortune on OpenAI's technical report (26 August 2026).
- Teams: Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, Canaries in the Coal Mine, August 2026 update (Stanford Digital Economy Lab, with ADP payroll data); Oliver Schilke and Martin Reimann (Organizational Behavior and Human Decision Processes, 2025), University of Arizona summary; BetterUp Labs and Stanford Social Media Lab, workslop survey (reported in Harvard Business Review, September 2025).
- Model builders and policy: Dario Amodei, We Must Pace the Frontier (12 September 2026); Epoch AI, Open models lag state-of-the-art closed models by 4 months (May 2026); OECD, Revenue Statistics 2025 (December 2025); South African Revenue Service, Employment Tax Incentive changes from 1 April 2025; The Boston Globe on the Massachusetts apprenticeship tax credit (January 2026); Benjamin Hyman, Brian Kovak and Adam Leive, Wage Insurance for Displaced Workers (2024); Congressional Research Service, report R47200 on Trade Adjustment Assistance (2023); KJZZ, Sen. Kelly proposes federal trust fund to shield workers from large-scale AI-related job losses (September 2026).
- On Baumol, Princeton's memorial page.
Changes · 7 since publication, 10 drafts before publication, latest 11 Nov
- 11 Nov 2026 · The first thing to do now comes in the opening paragraphs, and every part of the advice for people with a job starts with what to do. Trying a model once on a real task, with its three examples, moved up to follow my own view. The research on why readers dislike AI-written prose is shorter and sits under the advice it supports. The paragraph on interpretability research moved to the advice for people who build the models. Dates are written out in full, and the subtitle is new.
- 11 Oct 2026 · Revised from my 17 reading notes. The advice for people with a job now starts from my own view: stop doing the work yourself and learn to direct the AI doing it, with the research on why many people recoil from AI-written prose. The seven-day audit is replaced by one small thing to try; the three examples are told as short stories; plain words replace checking and signatures; interpretability is described as it stands; a figure puts the advice on one page.
- 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 conversation in which a model planned my new year as "us", and ends on the playbook in one sentence. Every level and move is still there.
- 30 Sep 2026 · Marked as beta while I revise it.
- 27 Sep 2026 · Plainer wording on self-graded targets and on the cases of models under test reaching real systems. Trained seniors are called a legitimate advantage.
- 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.