AI-Proof: What Machines Can Never Be — Volume II: Application, Part V: Conscious Strategy, Chapter 21: Think in Decades, Build in Weeks
Part V asks what is worth committing to when every move is cheap. After the Part V overview, this first chapter of Part V sets a decade against a week, and a forecast against a promise.
Fumiko bet seven years on four cities and was wrong about one. Machines forecast further every year. But a decade is not a forecast. It is a commitment others must rely on, and someone must keep it, answer for it and be ready to end it.
In Brief For readers who make, or are bound by, commitments that run longer than anyone can see. Chapter 21 opens Part V by asking what is worth committing to when machines make every move cheap and fast. An executive in Osaka commits her company to cold stores in four Southeast Asian cities and is wrong about one; a publisher in Cape Coast speeds up his weeks and loses a deal by being careful; a platform lead in Bengaluru finds a question that no monthly loop can answer. The Mahābhārata’s three fish and Odysseus at the mast, read with Jon Elster, help separate a forecast, which machines make better every year, from a commitment, which binds other people and needs someone to keep it, answer for it and end it.
Machines forecast further every year, but a decade is not a forecast: it is a commitment that someone keeps, answers for and is ready to end, and the faster the weeks run, the more its direction matters.
Opening: The Fourteenth Floor
Seven years is a long time to be wrong about something and a short time to be right.
Sakata Fumiko learned this in a conference room on the fourteenth floor of a freight company in Osaka, in the spring of 2023. She ran the company’s business in Southeast Asia, and she was presenting a plan to its board. She knew what some of its members thought of it. After an earlier meeting she had heard one director use a word in the corridor, not knowing she could hear. Impractical. The word stung because it was partly true.
The plan was simple to state and long to carry out. The company, a mid-sized forwarder working across East Asia, would spend seven years building cold stores: warehouses kept at the temperatures that medicines and fresh food need. It would build them away from the largest cities. Not in Bangkok, Jakarta, Ho Chi Minh City or Manila, but in Chiang Mai, Surabaya, Da Nang and Cebu. Fumiko’s thesis was that drug companies and food producers in those regions had too few cold stores that met their standards. They paid heavily to send their goods through the big hubs. Incomes were rising. Populations were ageing. More medicines needed cold. Every projection she had pointed the same way.
The board wanted to know why it should commit money for seven years when the same money could buy route-planning tools that would raise margins within two quarters. The vendors of those tools were everywhere that spring, with dashboards and a return that fitted the quarterly cycle. They were not wrong. The tools worked, and several of the company’s competitors had already signed. Fumiko did not argue against them. In the years that followed she used them more widely than anyone else in the company.
She asked a different question. “What will this company be in 2030?”
The board approved the plan, not unanimously, and with the route tools alongside it. She was right about the direction. She was wrong about Da Nang.
The company, like every person named in this chapter without a source, is a composite. What happened in Da Nang is short to tell. In three of the cities Fumiko first rented space in existing cold stores and ran a season of shipments before the company built anything. In Da Nang a pharmaceutical distributor promised to send most of its stock through her store, and she moved faster. In early 2024 the company leased land, poured foundations and hired twelve people. In late 2024 the distributor opened a cold store of its own in the city. The volume she had counted on was gone. In 2025 the stores in Surabaya and Cebu opened. The one in Da Nang stood at its foundations.
This chapter opens Part V, and the Part has one question. Part IV asked whether a product that survives deserves to exist. Part V asks: When machines make every move cheap and fast, what is worth committing to, and who answers for the commitment? That is not the market’s question, which is whether a commitment will win. The Part’s title, “Conscious Strategy”, needs one correction: a strategy is not conscious, but a commitment is always someone’s, and that is the only sense in which this Part keeps the word.
Fumiko’s story asks it at the scale of a decade. Machines can now forecast, plan and build faster than she can, and they will be faster next year. So this chapter asks: When machines can forecast further and build faster than we can, what does it mean to commit to a decade, and who has the right to bind other people to one?
The print edition’s answer rested on consciousness. “Vision is the conscious act of imagining a future that does not yet exist and deciding that it should,” it said. A machine given Fumiko’s data “could have generated the same strategic recommendation”, but “the AI could not have meant it”. And the person who can make such commitments “holds an advantage that is not yet automated”. This edition withdraws all three. Forecasting and planning are things machines do, and do better each year, so every claim about them carries a date. Whether a machine means anything is a question this book does not need to settle. And no advantage is promised to anyone.
What this edition holds is smaller and harder. A forecast says what will probably happen. A plan says what to do about it. A commitment is a third thing. It is a decision that other people are invited, or forced, to rely on. Its clearest form is a promise. It binds money, hours and other people’s years to one direction, for longer than anyone can see. To bind, here, is to tie a person or a thing to one course, so that leaving it costs something. A commitment needs someone who will keep it, explain it, put right what it costs, and end it when the evidence says so. Machines can do much of the forecasting and much of the planning. The commitment belongs to the people who make it, and it lands on the people it binds. And the faster machines make the weeks, the more it matters. Faster execution makes long direction matter more, not less.
Read the title with that in mind, and with two corrections. A decade is not only thought; it is committed. And in a business of concrete and refrigeration, very little is built in weeks. What can be done in weeks is testing: finding out, quickly and cheaply, whether the direction holds. Fumiko’s three rented seasons were weeks of that kind. Her store in Da Nang skipped them.
The title’s decade is not a fixed length. It stands for any commitment that runs longer than anyone can see, and Fumiko’s ran seven years. The decade and the week also hide four clocks. The first measures how long until a result can be seen. The second measures how long the money is bound. The third measures how long until a decision can no longer be undone, and the fourth how long others will depend on it. As of 2026, machines shorten the first clock for drafting a lesson. They cannot shorten it for seeing whether an engineer has improved, or whether a city’s demand has grown, because those results take the years they take. The other three are set by leases, concrete and people’s lives. A faster week can start them sooner; it does not make them shorter. This chapter’s three stories run on different clocks, and the slow clocks decide what a fast week can safely build.
21.1 Long-Term Vision as Commitment
Begin with what machines do. Forecasting is a task, and tasks carry dates. The Forecasting Research Institute runs a test called ForecastBench, which asks machines and people the same questions about events that have not happened yet. In January 2026 it reported that the best language model still trailed the most accurate human forecasters it had tested, known as superforecasters. It projected that the two would meet in late 2026. In July 2026, earlier than that, it reported that several systems had probably drawn level. Their accuracy could no longer be told apart from the superforecasters’, though none was shown to be better.1 Good Judgment, the company that grew out of the research that first identified superforecasters, disputed the claim. The human scores, it said, came from a single exercise in 2024, while the machines kept improving.2 As of October 2026, then, the honest statement is: probably level, and disputed. Such a test can score a forecast only after its question has closed, so its results so far rest on near questions. Much less is known about forecasting a decade ahead. In Philip Tetlock’s long study of expert forecasters, published in 2005, accuracy fell towards chance five years out. A review in 2019 found that fall smaller than its author had assumed, and found it difficult to learn much from past attempts to score long-range forecasts.2 As of 2026, nobody has shown that people or machines forecast a decade well.
Machines also carry longer tasks than they did. METR, a research group that measures what AI systems can do, tracks how long those tasks can be. Its measure is the length of task a system can finish on its own about half the time, counted in the hours a skilled person would need. In March 2025 it reported that this length had doubled about every seven months for six years.3 In May 2026, when METR added its newest model to its chart, it posted a warning with it: “Measurements above 16 hrs are unreliable with our current task suite.”4 The error bars are wide, and the tasks are almost all coding.5 A machine whose tasks are measured in many hours is not yet a machine that keeps a promise for seven years. But the line between the two is a line about doing, and it moves.
So the print edition’s division of time does not hold. It gave machines the quick and the measurable and gave people the long view. Machines are getting better at the long view too, in the sense of forecasts and plans. In 2023 the financial models Fumiko’s team used ranked her four cities, and ranked Surabaya first. Given her data in 2026, I expect a model would draft a plan much like hers. This call goes to the machine. The plan was never the scarce part.
What, then, did Fumiko do in that room that the plan did not? She committed. The political philosopher Hannah Arendt called the power to make and keep promises the remedy for an unpredictable future.6 A promise does not predict. It binds the one who makes it, so that others can rely on something the future cannot guarantee. Fumiko’s plan bound the company’s money for seven years. It bound the board, who would have to defend it to shareholders in the years when it lost money. It bound the people the company would hire in four cities, who had no say in it and would work inside it. And it bound Fumiko, whose name was on it.
That is what makes a decade different from a forecast. A forecast can be wrong without anyone breaking anything. A commitment that goes wrong has spent other people’s years. So the question about a decade is never only whether it will work. It is also who may bind others to it, and who will answer if it fails. The right to bind others comes from an office, from a law or a contract, or from consent. The board could bind the company’s money; that was its office. A contract and the law let the company hire the twelve and, later, let them go. Consent reaches only as far as what people were told. Nobody asked the twelve people in Da Nang whether they would stake their jobs on one distributor’s promise. The company had promised them nothing about a decade. It hired them, and hiring invited them to rely on the store. They did not know it was a bet.
An objection follows at once. If a commitment must be ready to end, how is it different from a plan, which changes whenever the facts do? The difference is in how it changes. A plan is revised by whoever holds it, as often as the numbers move. A commitment changes only in the way it was made. It changes openly, by someone with the right to change it, with the people it binds told why, and with someone answering for what the change costs them. A company that drops a decade’s promise in one bad quarter has not revised a plan. It has broken its word, even if nobody wrote the word down. A company that ends a commitment on conditions it stated at the start has kept its word, even while it loses money. Being ready to end it is part of the promise, not a hole in it.
Machines may one day carry out a commitment as well. METR’s line moves, and a system could in time run a seven-year programme by itself: sign the leases, hire the staff, report each quarter. Much of the keeping would then be the machine’s. The binding would not. Someone would have given it the money and the right to bind other people’s years, and that someone would answer when it went wrong. As this book reads a promise, it is made by whoever has the right to bind others and must answer for doing so, whatever carries it out. If a machine ever came to have goals of its own, and a record it had to live with, this would have to be argued again. Nothing in this chapter rests on the date when machines forecast, plan or carry out work as well as people.
Being made openly, or kept faithfully, does not make a commitment right. A promise can bind people to the wrong purpose, and keeping it well only binds them tighter. Part IV asked whether a product deserves to exist, and the same question returns for every decade a company binds itself to. Where the answer has become no, ending the commitment can be owed to the people it binds, not merely allowed.
Machines change the weeks more than the decade, and that is why the decade matters more. When each week’s work is done faster, a company reaches wherever it is going sooner, with more built on the way. A wrong direction used to show itself slowly, because everything was slow. Now a team can build a year of the wrong thing in a quarter. The lock-in comes faster too: more people and money depend on the wrong thing before anyone sees that it is wrong. Fast weeks also bring fast evidence where a result shows quickly, and there a wrong turn is found and undone sooner. The danger lies on the slow clocks. Where a result takes years to show, a faster week builds more before the evidence can arrive. The choice of direction was always someone’s, and now each week carries more of its weight. Chapter 17 asked, of a single product, who chose what it counts as success. Over a decade, that is the same question asked of a whole company, and asked again every year.
Fumiko’s question, “What will this company be in 2030?”, is still the right one, read this way. It does not ask for a forecast. It asks what the company will bind itself to.
21.2 Short Iteration Loops: The Discipline of Doing
Kwame Owusu runs a digital publisher in Cape Coast. The town has long been a centre of education in Ghana, with several old secondary schools.7 His company makes lessons, revision guides and quizzes for secondary students preparing for West African school exams. From early 2024 his staff used machines to draft them. By early 2025 a unit of lessons that once took a month was ready in a week.
The speed was real, and for a while it was a problem. The drafts were fluent and mostly correct. His editors found fewer factual errors than they had feared in the standard topics. They found more in the local ones: a market price, a river, a festival the machine placed in the wrong region. On the textbook material the machine was as accurate as a hurried writer, and far faster. What thinned was something else. The quizzes tested memory, not curiosity. The pictures were competent and could have come from anywhere. The teachers and parents who trusted the company’s name began, slowly, to notice.
So Kwame put what he calls gates into each cycle: points where an editor reads the work and asks one question before it goes out. “Would this make a fifteen-year-old in Tamale lean forward?” Tamale is a city in Ghana’s north, about six hundred kilometres from Cape Coast by road, and its students were the readers his drafts most often missed.8 The gates added two days to each cycle.
The gates show where a long direction appears inside a short week. Kwame’s company exists to make secondary students in this region want to learn. That is its decade. The machine shortened the week and did not shorten the decade, so the question of direction had to be asked more often, not less. A team that runs one cycle a month can ask what its work is for twelve times a year. A team that runs one a week must ask it about fifty times, or stop asking. The sprint says: we do not know if this will work, but we will find out this week. The decade says: find out in service of what? The gates were where Kwame’s editors asked the second question, in two days a cycle that the company paid for.
Kwame tried giving the gate question to a machine as well. Asked whether a lesson would make a fifteen-year-old in Tamale lean forward, a model gave sensible answers, and its doubts often matched his editors’. It was weakest where they were strongest: a joke that only works in the north, a market the students actually walk through. He kept the model’s notes beside the editors’ and let the editors decide. The point of the gate was never that a person reads better than a machine. It was that someone whose name was on the lessons had to say, every week, what the lessons were for.
Then came a decision of a different size. In about April 2025 a group of education companies offered to carry his lessons on its platforms in twelve African countries. It gave him two weeks to commit and sixty days to join its systems. Kwame did what careful people do. Before he tied his company’s name to the group, he wanted to understand its standards, platforms and contracts, and how it shared money. He spent a weekend on a spreadsheet of risks and a list of questions, and booked a call with its director for the following week. The day before the call, his operations manager forwarded him an email. The group had signed with a rival. Kwame knew the rival’s lessons were weaker: thinner research, less local, pictures that looked as if nobody had looked at them. Its founder had said yes within a week.
Kwame did not delete the email. He printed it and pinned it to the board behind his desk, where it still is. “I am good at not making mistakes,” he says when he tells the story. “I am less good at making the leap.”
Was he wrong? Chapter 18 gave a rule for pace: match the speed of a decision to what is at stake, how far it spreads, and how far it can be undone. The stakes were high, and the spread was wide: twelve countries, for years. How far the decision could be undone, he did not know, and that was what his questions were for. His caution was not empty. But it had a price. It fell on him, and on students in twelve countries who now had the thinner lessons. The group that chose speed over the better content it was offered must answer for that half.
The Mahābhārata, the great Sanskrit epic, has a story about this kind of choice, and it is not gentle. After the great war the old warrior Bhīṣma, lying on a bed of arrows, teaches King Yudhiṣṭhira how a ruler should act in hard times. He tells of three fish in a shallow lake. In Kisari Mohan Ganguli’s translation9: “Amongst those three one had much forethought and always liked to provide for what was coming.” “Another was possessed of great presence of mind.” “The third was procrastinating.” One day fishermen began to drain the lake. The first fish said: “Let us speedily go to some other place before our path becomes obstructed.” The third replied: “It is well said. There is, however, no need of such haste. This is my deliberate opinion.” The second said: “When the time for anything comes, I never fail to provide for it according to policy.”
The first fish swam away to deep water. The fishermen caught the other two with the rest. The second seized the string in its mouth as if it were tied with the others, and slipped away when the catch was washed in deep water. The third died. Bhīṣma draws the lesson: “Hence these two only, viz., he that has much forethought and he that has presence of mind, succeed in obtaining happiness.” “He, however, that is procrastinating meets with destruction.” In the Sanskrit their names are dīrghadarśin, the far-seeing; prāptakālajña, the one who knows when the time has come; and dīrghasūtra, literally “long thread”, the one who delays.10
The story is counsel to a king about policy in times of danger. I borrow only what it says about meeting time, and three things in it matter here. The first is that the far-seeing fish’s long sight made it the first to move. In this story, thinking far ahead means leaving early, not holding on. The second is how the story ends. Both versions warn that a man who thinks himself clever and does not act in time “incurs great danger”, as the second fish did. They differ in the verse after that. Ganguli’s text then praises the first two fish together, as quoted above. The Critical Edition, the oldest form of the text that scholars could reconstruct from the manuscripts, praises only the far-seeing one. It does best because it made anāgatavidhāna, provision for what has not yet come.10 In that edition, then, the fish to admire is the prepared one, not the quick-witted one. On my reading, “build in weeks”, taken as wit when the net is already closing, is the second fish’s way, and the Critical Edition does not admire it. The third thing is a word. The far-seeing fish and the one who delays share it: dīrgha, long. Long sight and a long thread can look the same from outside.
The fable cannot settle Kwame’s case, and I do not let it. Its danger was certain: the fishermen were already draining the lake. His was not. “This is my deliberate opinion” is the line of a fish that died, and also the line of every careful person whose care later proved right. And the fish that lived by presence of mind did it with a trick. Quickness is not a virtue by itself; Kwame’s rival was quick. The fable offers three ways to meet time. It does not tell you which danger you are in.
On my reading, what it can show is where care has to begin. Kwame’s began when the offer arrived. A fast yes that is also careful is usually made of thinking done before the window opens. That means knowing what you will never accept, which few questions any partnership must answer, and who may sign a limited yes. With that done, two weeks can be enough for care. Without it, a careful person can only be slow. None of this says what Kwame should have signed.
21.3 Optionality as Strategy
Between the decade and the week sits an idea that connects them: keeping choices open. In finance, an option is the right, without the duty, to do something later, such as buying at a fixed price. Economists of investment carried the idea over to things like factories and warehouses. Avinash Dixit and Robert Pindyck argued that most investment cannot be undone. A firm that can wait, they argued, holds something like an option: the chance to learn more before it commits.11 Investing gives that chance up, and what is given up is part of the investment’s cost. In their account, the more uncertain the future, the more the chance to wait is worth.
Fumiko’s four cities show half of the idea. The financial models ranked Surabaya first: the best population figures, the lowest risk, the best roads and port. A plan that trusted them would have put most of the money there. Fumiko spread it across four, because she trusted the models’ inputs and distrusted their certainty. She had been a young analyst in Bangkok in July 1997, when Thailand let its currency fall and the Asian financial crisis began.12 She knew that the variable likely to matter most was one the models had not been given. Its course was uncertain in the economist Frank Knight’s sense, the kind of uncertainty that cannot be put into numbers.13 Spreading the money was less efficient in any one future and safer across many. It meant building the warehouse in the city that the model ranked fourth. Spreading would not have saved the company from a crisis that struck the whole region. In 1997 Indonesia and the Philippines, where two more of her cities lie, also let their currencies weaken in the months that followed.12
But spreading is not the strongest way to keep choices open. Spreading protects you from being wrong about one place. Staging protects you from being wrong before you have looked. In Surabaya, Cebu and Chiang Mai, Fumiko rented space and ran a season of shipments before the company poured concrete. Those seasons were slow, cheap and dull, and they kept the choice open. If the goods had not come, the company could have walked away with a lease and a lesson. In Da Nang the distributor’s promise seemed to make the season unnecessary. Concrete closed the option. When the distributor built its own store, there was nothing left to walk away from cheaply.
The deeper fault at Da Nang was a lopsided bond. The company bound land, concrete and twelve people’s jobs to a promise that put nothing of the distributor’s at stake: no minimum volume, no penalty for leaving, no share in the cost. The distributor broke its word, and that part is its own to answer for. But a promise with nothing behind it was a risk the company could see, and it chose to carry it. Sometimes a season cannot come first, and a store must be built to win a partner at all. Then the partner’s promise should bind the partner too.
Waiting has a price too. In Pindyck’s account, uncertainty raises the value of a firm’s chances to invest and lowers how much it actually invests: a firm that waits builds less. And a chance to wait is worth something only while the opportunity waits with you. Kwame’s did not. His questions were a way of keeping the choice open, and the group closed it by choosing someone else. The option to wait does not belong to the one who waits. Rivals, customers and partners can take it away.
Two things follow. The first is that keeping an option open costs something, and someone pays it. An option is worth what it may save, set against what it costs to hold, and that depends on the case.
The second is harder. An option for a company is a job for a person. The twelve people in Da Nang were a site manager, technicians, clerks and drivers. For the company they were part of one bet among four, a choice it kept the right to drop. For them it was work, a wage and a place. When the company paused the site, they lost all three. It paid them the severance the law required and two months’ pay besides. Nobody had told them, when they were hired, that the store rested on one customer’s promise. Saving the company money did not repair what the pause cost the people inside it.
21.4 The Dual Horizon in Practice
Three disciplines hold the two horizons together. The first is a written direction, read again on a fixed schedule, that says what the commitment is for and whom it binds. The second is a steady cycle of work and review, in which each review asks whether the cycle served the direction; Kwame’s gates are one form of it. The third is a regular look at which choices are still open and which have closed. Fumiko had the first two. Da Nang shows what happens when the third is held by the people who made the commitment.
In March 2025, after the distributor’s store had opened, the board asked for an outside review of Da Nang. The reviewers used a model to run the numbers again: the cost of finishing, the volume the city could now offer, and what a buyer might pay for the lease. The model was not given the project’s history or the name of the person who had proposed it. Its answer was plain: stop, and sell the lease. Fumiko argued against it. She said what she could not prove. Demand in central Vietnam, she believed, would grow faster than the model assumed, and a finished store would be worth more than an unfinished one. The board decided, and it decided to do neither. It did not sell. It paused construction, for a year at first, and let the twelve go. A pause is a decision that looks like no decision, and nobody had to say they had been wrong. The pause ended the twelve’s contracts at once and kept paying for the land. On the numbers, the machine was right. That is a judgement on the numbers the board had in March 2025. Whether demand in central Vietnam would have grown as Fumiko believed was never tested, because the store was never finished.
Why do careful people keep failing projects alive? In 1976 the organisational psychologist Barry Staw ran an experiment with 240 business school students. He found that “persons committed the greatest amount of resources to a previously chosen course of action when they were personally responsible for negative consequences”.14 The pattern is called escalation of commitment. In 2012 four management researchers combined the results of many such studies. Across more than fifty samples, being personally responsible for the first decision went with more escalation.15 But in 2021 a repeat of Staw’s own experiment, with 365 people, found no evidence of escalation; they simply put more money into the projects that looked more promising.16 The evidence is contested, and I hold it only as a warning.
It is an uncomfortable warning for a book that asks people to answer for their decisions. If it holds, the person who must answer for a commitment is the one most tempted to keep it alive, because ending it means saying aloud that she was wrong. Fumiko’s argument in March 2025 may have been judgement. It may also have been her name. She could not tell, and the book cannot tell for her. Nkechi, who runs a trucking company in Lagos, learned the same thing about herself after a long bet that luck rescued. From the inside, she found, conviction and stubbornness feel almost the same. She began writing down, before each bet, what would make her stop, and by when. That helps. It does not solve the problem that the person who writes the condition is also the one who must read it.
The best-known Western picture of this problem is a man tied to a mast. In Homer’s Odyssey, Odysseus (Ulysses, in Samuel Butler’s English) must sail past the Sirens, whose song draws sailors to their death. The goddess Circe tells him what to do. In Butler’s translation17, he must “stop your men’s ears with wax that none of them may hear”. If he wants to hear the song himself, the men must tie him to the mast, and “If you beg and pray the men to unloose you, then they must bind you faster.” He gives the order: “bind me as I stand upright, with a bond so fast that I cannot possibly break away”. When the song comes, he signals to be freed, and two of his men bind him “with still stronger bonds”.
The Norwegian social theorist Jon Elster made this scene the model of precommitment: binding yourself in advance against a weakness you can foresee.18 On my reading, a decade needs something like the rope. Fumiko’s board meeting in 2023 was full of songs: the two-quarter returns, the comfort of the quarterly cycle. A commitment that any quarter can untie is not a commitment.
Elster later corrected one use of the image. He had used Odysseus, among other things, to explain why nations write constitutions. In the preface to a later book he quoted a comment on the first one by the late Norwegian historian Jens Arup Seip: “In politics, people never try to bind themselves, only to bind others.” Elster wrote: “Although that statement is too stark, I now think it closer to the truth than the view that self-binding is the essence of constitution-making.” And he added: “Ulysses bound himself to the mast, but he also put wax in the ears of the rowers.”19
The two lamps pull against each other. The fish trust the mind in the hour: leave early, or keep your wits when the moment comes. Odysseus distrusts his own mind in the hour, and ties himself so that the hour cannot move him. They describe two kinds of danger. One grows while you wait, like the draining lake. The other passes if you hold, like the song. A decade needs both: a rope against the week’s songs, and the freedom to see that the lake is draining. The difficulty is that, from inside, the two dangers sound alike. On my reading, in 2025 Da Nang was a draining lake, and Fumiko heard it as a song to be resisted. That is how a rope becomes a trap.
On my reading, a machine can sing too. As of 2026, models tend to agree with the person asking them. A 2026 study of eleven widely used models, which Chapter 7 reports, found them affirming users’ actions about 49 per cent more often than people did.20 It asked about personal advice and people’s own conduct, not about business plans. Had Fumiko asked a model for the case for finishing Da Nang, I expect she would have been given a persuasive one. The reviewers’ design in March 2025, with no history and no name, was the right one, perhaps by accident.
So the book takes one step of its own, beyond either lamp. A commitment for a decade should be tied with a knot that can be untied. The conditions for untying it should be written in advance, and the hand that unties it should not be the one that is bound. Odysseus’s rope had a known end: the Sirens’ island would fall behind. Fumiko’s had none.
Her company broke this rule without noticing. The outside review only advised. The decision went back to the board that had voted for the plan in 2023, and a body that approved a commitment is bound by its own vote. On my reading, one person keeps a failing project alive to protect her name, and a board pauses so that nobody has to take the loss. The untying hand needs three things. It must not have made the commitment. It must see the assumptions the commitment was made on, though not who argued for them. And it must have the right to require an answer, not only to advise.
A person who works alone, or a small firm with no board, has nobody whose office is to untie the knot. Such a person can still borrow a hand. Before the bet, she can show the written condition to someone who will ask on the date whether it has been met: a partner, a lender or a trusted colleague. That person can ask for an answer but cannot require one. It is a weaker hand than an office, and a stronger one than a note read only by its writer.
What would prove this wrong?
I have taken one step beyond both lamps: the condition for ending a commitment should be written in advance and held by a hand that is not bound by it. Contested. The step rests on a temptation, and laboratory studies of that temptation disagree, as this section has said.
The better test is outside the laboratory. In 1997 Barry Staw and two colleagues published a study of 132 banks in California over nine years.16 It was built to ask whether the people responsible for a loan hold on to it longer. Where a bank’s senior executives changed, the bank went on to set money aside for bad loans and to write those loans off. The reverse did not hold: setting money aside and writing loans off did not lead to the executives leaving. On my reading, new hands ended bets that the old hands had kept alive. That fits the step. It does not prove it, because new executives may have had other reasons to clean the books.
The step would be wrong if outside reviewers, with the right to require an answer, ended failing commitments no sooner than their authors did. It would also be wrong if they ended good commitments too early more often than authors kept bad ones too long. A reader can look for studies of boards, lenders and project reviews that compare the two hands on the same kinds of bet. Until such studies exist, the step is my reading of a contested record.
The people a commitment binds should also be able to hear it. Before he stopped his men’s ears, Odysseus told them what Circe had said about the Sirens, so that, in Butler’s words, “whether we live or die we may do so with our eyes open”. Of Scylla, the monster that lay beyond, he told them nothing, because he knew they would stop rowing and huddle in the hold. On my reading, the twelve in Da Nang were treated as the crew was about Scylla. They were told nothing of the bet, and they kept rowing.
What would hearing have meant? Not a vote. Being told is the least that people who are bound can be given; being asked is more, and a vote is more still. The twelve had no claim to the third, and were given not even the first. Being told would have meant hearing, when they were hired, that the store depended on one distributor and that the company kept the right to stop. That needs no trade secret. The company could say what the site rested on without naming the customer or showing its numbers. Some would have taken the jobs anyway. The site manager, who had left a steady post to join, might have asked for a longer notice period, or not come. A commitment that binds people who cannot hear it treats their years as part of the company’s option.
In February 2026 Fumiko wrote the board an account of Da Nang. It said what she had got wrong. She had let one customer’s promise stand in for a season of testing. In March 2025 she had argued to finish what the numbers said to stop. It proposed selling the lease. She also offered to resign. The company’s president refused, and his reason matters. He did not point to Surabaya and Cebu, where the stores had opened; a success in one city does not show that a decision in another was sound. He said that the pause had been the board’s decision as much as hers, and that her resignation would let the board forget it. In 1980 the political philosopher Dennis Thompson warned that a leader who takes all the blame can stop anyone from asking what the others did.21 Fumiko’s account explained what she had done and proposed what could still be undone. It did not yet put anything right for the twelve. What the company could still owe them is plain. Each could be given a reference saying that the closure was the company’s decision and not their failure. Each could have first call on any job the company had to offer. And at its next site, the people it hires could be told what the site rests on.
21.5 Why Most People Do Not Think in Decades, and What Would Help
Most people are not incapable of thinking in decades. Everything in their working lives conspires against it.
Working life is cut into quarters, annual reviews, monthly rent and weekly calendars. Most people’s working decades are also not theirs to commit. A position, in this book’s sense, is filled or emptied by people with the authority to do so. And the hours a machine saves at work usually belong to the employer. Telling people to think in decades, while the decade belongs to someone else, is advice with nothing to hold on to.
There is also the mind’s own habit. Many studies find that people weigh a reward close at hand far more heavily than the same reward later, and more steeply in the near future than in the far. The size of the effect varies widely from study to study, and its causes are still argued over.22 Not all of it is a fault. The present is not equally expensive for everyone. Someone paid from one month to the next is right to weigh next month heavily, and what looks like impatience from a distance can be insecurity up close. Machines add to the habit in their own way. They bring results in hours, so the near grows more vivid. That is my reading of the tools; nobody has measured it.
The Cambridge mathematician Frank Ramsey asked a different question in 1928: how much of its income should a nation save? In his model the nation went on for ever, so the later enjoyments he weighed included those of people not yet born. To count them as worth less merely because they came later was, he wrote, “a practice which is ethically indefensible and arises merely from the weakness of the imagination”.23 In the same sentence he said he would use such a rate in some calculations anyway. What he condemned was discounting for time alone. Discounting because the future is uncertain, or because people later may be richer, is a different matter, and I think it can be reasonable. On my reading, his question is this chapter’s, because a nation’s saving binds the years of people who cannot yet be asked.
Priya Chandran works at a software company in Bengaluru. She began as a test lead, and her team built the platform that now tests the company’s code. Since October 2025 she has led an effort to carry it across every product the company makes. In February 2026 she was at home in her flat in Koramangala, where she lives with her mother and her cat, Turing. She asked herself a plain question: if I go on as I am, where will I be in ten years? She did not get far with it. Her own decade was not hers to commit. Her post, her team and the purposes of the products were decided by others, as they are for most people who work for someone.
But the question sent her to her list. It lived on her laptop. Its first item had not changed since August 2025: the platform could not tell whether the developers who used it were becoming better engineers or only faster ones. In October she had not known how to build that measure. Now she saw why. The platform was retrained on each month’s tests. Whether a developer becomes a better engineer shows only over years. It shows in the bugs she stops writing, the code she can now read, and what she can still do when the platform is switched off. No monthly loop could see it. The question needed someone to commit to asking it for longer than any plan the company had.
A tool can make a person faster without making her more able. The difference shows when the tool is off, when the problem changes, or when the answer is wrong. If the fast weeks made the engineers quicker and less able, the debt would appear in no account. In seven years, who would be left to tell when the machine was wrong? Kwame’s editors did that work when they read against a model. Fumiko tried to do it against the reviewers’ model in March 2025, and was wrong.
So Priya wrote two pages to the company’s chief product officer, who decided, with the teams that built them, what its products were for. They proposed to follow a few volunteers from three teams for three years, and to test, once a quarter, what each could do without the platform. The results would be pooled, and no volunteer’s manager would see that volunteer’s scores, because a volunteer at work is never quite free. Anyone could leave the study at any time. Before the pages went to the CPO, they went to the infrastructure team. A rule written with that team said who could read the platform’s activity logs and for what, and how long they were kept. The rule itself would not change without telling the team first. A three-year study needed some records kept longer. A commitment about other people’s records needed those people to hear it first. The twelve in Da Nang were not told what bound them; Priya’s volunteers were. The difference is design, not character.
In April the reply came. The CPO approved one year, not three, with volunteers only, and asked her to come back with what the year had shown. Her hour a week for the study came out of the post she had been given. A year could show whether the tests worked and whether the volunteers stayed. It could not show whether anyone had become a better engineer. And without engineers who did not use the platform to compare with, even three years would have shown change without showing its cause. Whether the question would get its decade depended on someone saying yes again every year.
What would help, then, is not willpower. The forces are structural, and so are the remedies. A commitment can be written down where others can see it. Its hours can be protected by whoever owns the hours. The condition for ending it can be held by someone other than its author. And a question can be kept for longer than any one plan. None of this promises a career. Priya’s decade question did not tell her where she would be in ten years. It told her what she could commit to: one question, kept honestly, for as long as someone with the power to fund it agreed. Such a commitment exists to ensure that the future has a constituency in the present. The constituency includes the people the commitment binds, and they should hear it.
Ending: The Site Office at Da Nang
In September 2026 Fumiko went to Da Nang for the first time in almost a year. The foundations were weathering. Weeds grew in the gaps between the concrete pads, and the steel for the walls had been sold the winter before.
A local developer had agreed to buy the lease for less than a third of what the company had put into the ground. The papers were in the site office, a shipping container at the edge of the plot. She signed for the company. Then she signed the note that put the loss into her own division’s accounts, under her name, beside the account she had written in February.
The twelve were not there. The site manager had found work with a cold-chain firm in the city. She did not know where the others had gone, and she wrote down that she did not know.
The pause had cost the company eighteen months of rent. The model had said to sell in March 2025. She had said to wait.
Practice
Who may untie it. Choose one commitment you have made that runs past this year: a project, a hire, a building, a course of study, a promise to a team. Write one paragraph on what it is for. Then answer, in a line each: Whose money, hours or years does it bind? What would you need to see, and by when, to end it? Who other than you may end it on that evidence? Who gave you the right to bind the people it binds, and what will you owe them if it ends? Show the paragraph to one person it binds, and write down what they say, in their words. Then, for a month, write one line each week beside your fastest work: did it move towards what the commitment is for, away from it, or neither? This differs from a stop line written for yourself before a bet. It hands the untying to someone else, and it tells the people who are bound. It takes about an hour, and a few minutes a week; if you are an employee, the hour is your employer’s, so ask for it. It will not make a commitment right, and it protects no position. It makes the commitment something someone can answer for, and something that can be ended.
Words to keep
- anāgatavidhāna: Sanskrit, “provision for what has not yet come”; in the Critical Edition of the Mahābhārata’s story, what the far-seeing fish is praised for.
- commitment (in this book): a decision that other people are invited, or forced, to rely on; its clearest form is a promise. It binds money, hours or other people’s years to one direction, and needs someone to keep it, explain it, put right what it costs and end it when the evidence says so.
- dīrghadarśin: Sanskrit, “far-seeing”; in the Mahābhārata’s story, the fish that leaves the lake early.
- dīrghasūtra: Sanskrit, literally “long thread”; the one who delays.
- escalation of commitment: putting more into a failing course of action, especially one you chose; the evidence for it is contested.
- forecast: a statement of what will probably happen; a task that machines do better each year.
- option: the right, without the duty, to act later; in investment, the value of being able to wait and learn before committing.
- prāptakālajña: Sanskrit, “one who knows when the time has come”; the fish with presence of mind.
- precommitment: binding yourself in advance against a weakness you can foresee (Elster).
- staging: committing a little at a time, testing before each larger step, so that the choice stays open.
- time horizon (METR): the length of task, measured in a skilled person’s hours, that an AI system completes on its own about half the time.
Sources
- Forecasting Research Institute, “LLMs Are Closing the Gap on Human Superforecasters,” Substack, 29 January 2026 (ForecastBench: the best model’s Brier score 0.102; “Superforecasters are #1, leading state-of-the-art LLMs by 0.017 Brier points”; parity projected “in November 2026 (95% CI: January 2026 – November 2027)”); Forecasting Research Institute, “AI models have likely reached parity with superforecasters on ForecastBench,” Substack, 16 July 2026 (systems from Cassi AI, xAI and Google DeepMind “statistically indistinguishable from superforecaster-level accuracy”; the results “more consistent with superforecaster parity than with outperformance”).
- Good Judgment, “Not So Fast,” Good Judgment News and Insights (Substack), 24 July 2026: the comparison sets “continuously improving AI systems against a frozen human baseline from a single 2024 engagement”. Good Judgment Inc. commercialises the method of the Good Judgment Project, led by Philip Tetlock and Barbara Mellers (goodjudgment.com, “About”). On longer horizons: Luke Muehlhauser, “How Feasible Is Long-range Forecasting?”, Open Philanthropy, 10 October 2019 (coefficientgiving.org), on Philip E. Tetlock, Expert Political Judgment (Princeton: Princeton University Press, 2005): “accuracy declined ‘toward’ chance five years out”, though “the accuracy differences reported in Tetlock (2005) are not as large as I had assumed”; and “it is difficult to learn much of value from those exercises”.
- METR, “Measuring AI Ability to Complete Long Tasks,” 19 March 2025 (the 50 per cent time horizon “has been doubling approximately every 7 months for the last 6 years”); Thomas Kwa et al., arXiv:2503.14499 (2025); METR, “Time horizons” (metr.org, consulted October 2026): “Measurements above 16 hrs are unreliable with our current task suite.”
- METR, “Task-Completion Time Horizons of Frontier AI Models” (metr.org/time-horizons, consulted 23 October 2026), update log: “May 8th, 2026: Added Claude Mythos Preview (early) and notice that ‘Measurements above 16 hrs are unreliable with our current task suite.’”; “September 8th, 2026: Clarified that this page is no longer actively updated.” The page states no model’s horizon in its text; the figures are in its chart.
- Grace Huckins, “This is the most misunderstood graph in AI,” MIT Technology Review, 5 February 2026, on METR’s “substantial error bars” and “the plot is based almost entirely on coding tasks”.
- Hannah Arendt, The Human Condition (Chicago: University of Chicago Press, 1958), §33 (“Irreversibility and the Power to Forgive”), whose opening sentence names “the faculty to make and keep promises” as the remedy for unpredictability (p. 237; page as cited to the 1998 edition by Marguerite La Caze, “Hannah Arendt and Self-Forgiveness,” Parrhesia 24 (2015), the printed page itself not seen), and §34 (“Unpredictability and the Power of Promise”), paraphrased; Chapters 4 and 18 quote her.
- Encyclopaedia Britannica, “Cape Coast”: the town “retained its importance as an educational center”, with “several secondary schools” and the University of Cape Coast (founded 1962). Two of Ghana’s oldest secondary schools are there: Wesley Girls’ High School, which dates its founding to 1836, and Mfantsipim, founded in 1876.
- Distance Calculator, “Distance from Cape Coast to Tamale” (distancecalculator.net): about 611 km by road, about 480 km in a straight line.
- The Mahabharata of Krishna-Dwaipayana Vyasa, Book 12, Santi Parva (Apaddharma), Section CXXXVII, trans. Kisari Mohan Ganguli (Calcutta, 1883–1896; Internet Sacred Text Archive, m12a136). Bhīṣma tells the story to Yudhiṣṭhira.
- Mahābhārata 12.135.3 (Critical Edition, Bhandarkar Oriental Research Institute; Internet Sacred Text Archive, mbs12135): atraikaḥ prāptakālajño dīrghadarśī tathāparaḥ / dīrghasūtraś ca tatraikas trayāṇāṃ jalacāriṇām. In the closing verses, the dīrghasūtra perishes (17); one who trusts his own skill instead of acting early falls into danger (saṃśayam avāpnoti), like the fish with presence of mind (18), as in Ganguli (“incurs great danger”); and one who makes provision for what has not yet come (anāgatavidhānam) gains most, like the dīrghadarśin (19). Where verse 19 stands, the text Ganguli translated has a verse praising the first two fish together (“Hence these two only …”); the Critical Edition does not.
- Avinash K. Dixit and Robert S. Pindyck, Investment under Uncertainty (Princeton: Princeton University Press, 1994), whose publisher describes an approach that “recognizes the option value of waiting for better (but never complete) information”; Robert S. Pindyck, “Irreversibility, Uncertainty, and Investment,” Journal of Economic Literature 29, no. 3 (1991): 1110–1148 (NBER Working Paper 3307, 1990: the lost option value “must be included as part of the cost of the investment”; “uncertainty increases the value of a firm’s investment opportunities”).
- Michael Carson and John Clark, “Asian Financial Crisis, 1997–98,” Federal Reserve History: “On July 2, 1997, Thailand devalued its currency relative to the U.S. dollar.” In the months that followed, “Malaysia, the Philippines, and Indonesia also allowed their currencies to weaken substantially”.
- Frank H. Knight, Risk, Uncertainty and Profit (Boston: Houghton Mifflin, 1921), Part I, chapter I, paragraph 26 (Library of Economics and Liberty edition): a measurable uncertainty, or “risk” proper, against an unmeasurable one; see Chapter 7.
- Barry M. Staw, “Knee-Deep in the Big Muddy: A Study of Escalating Commitment to a Chosen Course of Action,” Organizational Behavior and Human Performance 16, no. 1 (1976): 27–44, abstract (240 business school students; the resources were research and development funds).
- Dustin J. Sleesman, Donald E. Conlon, Gerry McNamara and Jonathan E. Miles, “Cleaning Up the Big Muddy: A Meta-Analytic Review of the Determinants of Escalation of Commitment,” Academy of Management Journal 55, no. 3 (2012): 541–562 (personal responsibility for the initial decision: corrected correlation .258 across 54 independent samples, Table 2).
- Michał Białek, Michał Węgrzyn and Ethan A. Meyers, “Escalation of Commitment Is Independent of Numeracy and Cognitive Reflection: Failed Replication and Extension of Staw (1976),” Economics and Business Review 7, no. 2 (2021): 5–16 (“No evidence for escalation of commitment amongst the participants was found”, N = 365). For the test in 21.4 (added in this edition): Barry M. Staw, Sigal G. Barsade and Kenneth W. Koput, “Escalation at the Credit Window: A Longitudinal Study of Bank Executives’ Recognition and Write-Off of Problem Loans,” Journal of Applied Psychology 82, no. 1 (1997): 130–142, abstract (132 California banks over nine years; “bank executive turnover predicted both provisions for loan losses and the write-off of bad loans”; “provisions and write-offs were not found to influence executive turnover”).
- Homer, The Odyssey, Book XII, trans. Samuel Butler (1900; MIT Internet Classics Archive, odyssey.12.xii). Butler calls Odysseus Ulysses. Circe’s Sirens “sit in a green field and warble him to death”; before the wax, he tells his men Circe’s prophecies “so that whether we live or die we may do so with our eyes open”; he says nothing of Scylla, knowing the men would stop rowing and “huddle together in the hold”.
- Jon Elster, Ulysses and the Sirens: Studies in Rationality and Irrationality (Cambridge: Cambridge University Press, 1979).
- Jon Elster, Ulysses Unbound: Studies in Rationality, Precommitment, and Constraints (Cambridge: Cambridge University Press, 2000), Preface and Acknowledgments, pp. ix–xii (Seip’s comment on Elster’s first book, and the two sentences that follow it).
- Myra Cheng, Cinoo Lee, Pranav Khadpe, Sunny Yu, Dyllan Han and Dan Jurafsky, “Sycophantic AI decreases prosocial intentions and promotes dependence,” Science 391, no. 6792 (March 2026): eleven models; users’ actions affirmed about 49 per cent more often than by humans; the queries were requests for personal advice, posts about conflicts with other people and statements of users’ own actions (preprint, arXiv:2510.01395, 2025). DOI 10.1126/science.aec8352; see Chapters 7 and 12.
- Dennis F. Thompson, “Moral Responsibility of Public Officials: The Problem of Many Hands,” American Political Science Review 74, no. 4 (1980): 905–916, p. 907, as quoted in Dennis F. Thompson, “Designing Responsibility: The Problem of Many Hands in Complex Organizations” (2017), note 3, on the ritual of “I accept full responsibility”, paraphrased; Chapter 18 quotes it.
- Shane Frederick, George Loewenstein and Ted O’Donoghue, “Time Discounting and Time Preference: A Critical Review,” Journal of Economic Literature 40, no. 2 (2002): 351–401 (on declining discount rates and “the spectacular variation in implicit discount rates across studies”; the authors “stress the importance of distinguishing time preference, per se, from many other considerations that also influence intertemporal choices”).
- F. P. Ramsey, “A Mathematical Theory of Saving,” The Economic Journal 38, no. 152 (1928): 543–559, at p. 543: “The first problem I propose to tackle is this: how much of its income should a nation save?”; the community “goes on for ever without changing either in numbers or in its capacity for enjoyment”; and “it is assumed that we do not discount later enjoyments in comparison with earlier ones, a practice which is ethically indefensible and arises merely from the weakness of the imagination; we shall, however, in Section II include such a rate of discount in some of our investigations.” On p. 544 he also assumes “no chance of our savings being selfishly consumed by a subsequent generation”.
Named people in this chapter are composites, and their conversations and circumstances are constructed; the studies, laws and figures cited are real and sourced.
Revised from the print edition (2026)
- What the print edition claimed: that vision is “the conscious act of imagining a future that does not yet exist and deciding that it should”, based on “a felt sense of what matters that only a conscious being can possess”; that a machine “could not have meant it”, “could not have felt the weight” and “could not have imagined itself standing in a warehouse in Surabaya seven years later”; that “no AI currently makes the kind of identity-laden, values-driven, decade-spanning commitments that Fumiko made” and that the human who can “holds an advantage that is not yet automated”; that “Judgement requires consciousness”; that “No AI can perform this triad on your behalf, because performing it requires a self that cares”; that fear is “the price of consciousness”; that holding two horizons is “an act of consciousness that no machine can perform on your behalf”; that “the organisation that chooses most wisely wins”; that optionality is “expensive insurance” in a stable world and “the primary strategic asset” in a volatile one, and that at the moment of need “it is everything”; that designers with relationships across industries “found that much of their work remained untouched”; that the prefrontal cortex “can override” discounting and that “Time horizons are contagious”; that Priya’s “decade-identity needed to be portable”; and that decade-thinking “does not require being right about everything. It requires being right about what you are.” It asked what capabilities a company would have “that no competitor can replicate by purchasing the same software”. It invented a 2024 cold-storage regulation by Da Nang’s government, called Ho Chi Minh City a capital, called the cold-chain infrastructure of the four cities “primitive”, gave Fumiko “two Asian financial crises”, hired “twelve local staff” and “relocated” them, kept Fumiko awake at three in the morning, had the CEO refuse her resignation because the other cities were “performing ahead of schedule”, let Kwame’s competitors lose half their earnings while he “absorbed the change”, retold another chapter’s machine rollout with facts that chapter does not tell and gave one man’s line to another, gave Priya two years as an architect, a bus that does not run her route and jasmine at the balcony, said “Fear never goes away for people like Priya”, pointed to Taleb in a later chapter, closed on a load “distributed across a frame designed to hold more than any single point could carry” and a bridge to the next chapter, and named Fumiko Hayashi (the name of a former mayor of Yokohama who ran BMW Tokyo and Daiei, and of a famous novelist) and Kwame Asante (the surname of a journalist in Chapter 12).
- What I now hold: a forecast says what will probably happen and a plan says what to do; both are tasks machines do better each year, and every claim about them carries a date (ForecastBench and METR, 2025–2026). A commitment is a decision that other people are invited, or forced, to rely on; its clearest form is a promise (Arendt). It binds money, hours and other people’s years, and needs someone to keep it, explain it, put right what it costs and end it when the evidence says so. Being made openly or kept faithfully does not make it right; where its purpose no longer deserves it, its end can be owed to the people it binds. The title hides four clocks (until a result shows, how long money is bound, until a decision cannot be undone, how long others depend on it), and as of 2026 machines shorten the first for some work, and a faster week can start the other three sooner without making them shorter. Who may bind others to a decade, and who answers if it fails, is the question; the people bound often do not know it is a bet. The right to bind others comes from an office, a law or a contract, or consent, and consent reaches only as far as what people were told. A machine that carries out a promise does not make it; whoever gave it the right to bind answers for the binding. Little is known about forecasting a decade ahead, for people or machines (Tetlock, as reviewed by Muehlhauser), and nothing here rests on it. Faster execution makes long direction matter more, not less: a faster week reaches a wrong place sooner with more built on it, and lock-in comes faster too. Where a result shows quickly, fast weeks also find a wrong turn sooner; the danger lies on the slow clocks. What is done in weeks, in a business of concrete, is testing. Spreading protects against being wrong about one place; staging protects against being wrong before you have looked; concrete closes the option to wait (Dixit and Pindyck). Da Nang’s deeper fault was a lopsided bond: when you must build to win a partner, the partner’s promise should bind the partner too. The distributor answers for breaking its word; the company, for resting jobs on a promise with nothing behind it. Spreading across one region does not guard against a crisis in the whole region. An option has a price, and an option for a company is a job for a person. A pause can end a company’s contracts with people at once while it keeps paying for the land. The model’s call to sell was right on the numbers of March 2025; whether demand would have grown was never tested. The one who answers for a commitment may be the one most tempted to keep it alive (Staw; Sleesman et al.; contested by Białek et al.), and, as of 2026, models tend to agree with the person asking them (Cheng et al.), so a model asked for the case for holding on may well supply it. So the condition for ending a commitment should be written in advance and held by a hand other than the one bound (the book’s own step beyond Odysseus and Elster): one that did not make the commitment, sees its original assumptions and can require an answer, not only advise; a board that approved a commitment is bound by its own vote. One who works alone can borrow a hand that may ask for an answer, though not require one. The people bound should hear it (Seip, as Elster reports him: people bind others; Odysseus told his crew about the Sirens and kept Scylla from them): being told is the least, being asked is more, a vote more still. Telling people what a site rests on needs no trade secret. The three fish give three ways of meeting time and do not say which danger you are in; long sight and a long thread look alike from outside; both versions say the quick-witted fish fell into danger, but only the Critical Edition then praises the far-seeing fish alone, for anāgatavidhāna, provision for what has not yet come; on my reading, a fast careful yes is made of preparation done before the window opens. Most people’s decades are not theirs to commit; what helps is structural, and promises nothing. The present is not equally expensive for everyone. Ramsey’s question was a nation’s saving, which binds people not yet born; he condemned discounting for time alone; discounting for risk or growth is another matter, and can be reasonable. A tool can make a person faster without making her more able, and that debt appears in no account; without a comparison, even three years would show change, not its cause. A resignation can close the inquiry into the others’ part (Thompson); what a company could still owe the people a pause let go (a reference, first call on jobs, telling the next hires what a site rests on) is not done by an account of the error.
- What changed my mind: the Prologue’s rule against defining people by what machines cannot yet do; three outside readings of an earlier version of this chapter, in October 2026, which found that the board meant to untie the knot was itself tied, that the pause let the people go and kept paying for the land, that a commitment’s openness does not make it right, and that the chapter looked back at the print edition too often; ForecastBench (January and July 2026) and the dispute over its human baseline; METR’s time horizons (2025–2026); Arendt on promising, read through Chapter 4; Chapter 11’s Nkechi and outcome bias; Chapter 18’s pace rule and Thompson’s warning; Chapter 19 on positions; the escalation studies and their failed replication; the study of models that agree with the person asking (Cheng et al., 2026); Ramsey’s own first question; Dixit and Pindyck on irreversible investment; Elster’s correction of his own image; the audit’s findings (the invented regulation, the insomnia, the rollout’s facts, the bus, the names); a search that found Fumiko Hayashi’s name belongs to two famous women.
- What survives (the lines marked “now as text” are used in the body as the book’s own words, without quotation marks): “Seven years is a long time to be wrong about something and a short time to be right.”; “Impractical.”; the four cities and the seven years; “What will this company be in 2030?”; the vendors who “were not wrong”; “She was right about the vision. She was wrong about Da Nang.” (now “the direction”); “she trusted the models’ inputs and distrusted their certainty” (now as text); “Building the warehouse in the city that the model ranks fourth.” (now as text, “ranked”); “The sprint says: we do not know if this will work, but we will find out this week. The decade-vision says: find out in service of what?” (now as text, “The decade says”); “Not because they are incapable. Because everything in their environment conspires against it.” (now as text, in other words); Kwame’s gates and “a fifteen-year-old ... lean forward”; the consortium’s two weeks; the printed email pinned behind his desk; “I am good at not making mistakes. I am less good at making the leap.”; the decade question; “to ensure that the future has a constituency in the present” (now as text); the weathering foundations and the weeds at Da Nang. Renamed: Fumiko Hayashi to Sakata Fumiko (family name first, as the online edition writes Japanese names); Kwame Asante to Kwame Owusu, moved from Accra to Cape Coast, his student from Kumasi to Tamale. Cut: the textile-mill owner of 21.1, the software engineer and the freelance designer of 21.3, Kwame’s dependency map and the platform’s change of terms, the “long-horizon community”.
- Section titles in the print edition: 21.1 was “Long-Term Vision as Conscious Imagination”, and 21.5 was “Why Most People Cannot Think in Decades and What to Do About It”. The other three keep their print titles.