AI-Proof: What Machines Can Never Be — Volume II: Application, Part IV: Conscious Creation, Chapter 14: The Commodity Trap

7 hours ago

41 min read

Volume II turns from the person to the things people build. After the Part IV overview, this first chapter of Part IV asks what becomes of a product when anyone can rebuild it.

← Previous: Part IV: Conscious Creation (overview) · Series hub · Next: Chapter 15: Building Real Moats in an AI World →

Chidi’s company took seven months to die and an afternoon to rebuild. Whether a product survives the next copy is one question. Whether it deserved to exist is another, and the market does not answer it.

In Brief: For readers who build, fund or work on products made on someone else’s AI model. Chapter 14 opens the book’s second volume, on what people build. A founder in Lagos watches a product built on someone else’s AI model die in seven months; a founder in Recife learns from one customer’s phone call what her customers were paying for; a cloth trader in Balogun Market tests a newcomer from Aba with a question about the rains. The Tirukkuṟaḷ on enterprises that eat their capital, Locke and Nozick on what labour makes yours, and the Īśa Upaniṣad on whose wealth is anyone’s help answer what would show that a product deserves to exist, and not only that it has survived.

When anyone can rebuild what you made in a fortnight, on a machine you do not own, surviving does not show that it deserves to exist; that is judged by what it does to the people it serves, to those who pay for it, and to those who pay without being asked.

Opening: The Copy in an Afternoon

Chidi Okonkwo asked the machine to build his company again.

It was January 2025, a month after he had closed it. He sat at the same rented desk in a co-working space in Yaba, the part of Lagos where young technology companies gather near the university. He described his old product to an AI coding tool in plain words: a clean screen, a set of instructions for a language model, a page that took subscriptions. Then he watched. By the end of the afternoon he had a rough copy that worked. It drafted a simple contract for a shop in Ikeja. It wrote an advertisement for a hair salon. It answered a question about registering a business name. The payment page did not yet take money. The copy lacked nothing the subscription had paid for. What it lacked was the company, and that was the part Chidi had never priced.

The company had taken seven months to die.

It began in April 2024. Chidi was twenty-nine. He had studied computer science at the University of Lagos and spent two years at a payments company on Victoria Island. He left to build something of his own, and he was good at building: fast, careful with code, able to ship over a weekend and have customers by Monday. His product took a large language model made by another company, through the connection that company sells to developers. It wrapped the model in a clean screen and in instructions written for small-business owners in West Africa. It drafted contracts, marketing copy and financial projections, and it replied to basic legal questions. Four hundred people paid for it in the first three weeks. Two thousand were paying by July. In August a Lagos venture fund invested.

He checked the numbers many times a day. He refreshed the dashboard the way his mother checked the weather: the information would not change his plans, but the act of checking was itself a comfort. Revenue rose. Few users left. Each new user cost less to win. Everything pointed the same way.

The numbers did not tell him that his whole product could be described on one page, which a competent team would read in an afternoon. In May 2024 the company whose model he used made its best model available free, with limits, in its own app.¹ Few of his users noticed at first. More noticed each month. In September an update to the model changed how it read his instructions, and he spent two weeks repairing replies that his users had already learned to get elsewhere. In October a team in Nairobi launched a near copy, better in three African languages. It had taken them twelve days.

The numbers told him he was winning. He was evaporating. By November people were leaving faster than he could find new ones, and each new one cost three times what it had in the summer. His two employees, a designer and a support coordinator, were friends from university. They sat across the desk and waited for him to say what all three of them knew. He closed the company in December.

He was not the first. Chapter 6 told of Adaeze Nwosu, in the same city. Her writing assistant for Nigerian English shrank the same year, as the model under it learned to do what she sold. Chidi had heard about her. He built anyway.

The print edition told Chidi’s season as a story about his temperament. “Chidi loved the speed,” it said. “Not as a strategy. As a feeling.” That is part of the story. The rest belongs to others. The fund that invested in August asked him each month for one thing, which was growth. It did not ask the questions this chapter will ask. A fund of that kind expects most of its bets to fail and a few to pay for all the rest. For the fund, Chidi’s company was one bet among many. For Chidi, it was seven months of his life and a year’s savings. For the designer and the support coordinator, it was a job each, and they lost it. The users lost almost nothing. They moved to the free app, and most of them were better served there.

So in January, with a working copy on the screen, Chidi asked himself the question that his speed had been outrunning since April. What did I build? An interface. Some prompts. A subscription page. A layer of convenience around someone else’s intelligence. When the intelligence improved, the layer shrank. When its maker offered its own screen, the users left. When a competitor used the same intelligence to build a similar layer, the difference vanished. Platforms build their own features, eventually, without malice, and often without noticing the small companies they absorb.

A company that a machine can rebuild in an afternoon sells a commodity: something that is the same whoever sells it, so that buyers choose on price. For the buyer, that is good news. Water is Adam Smith’s example: nothing is more useful, and it will buy almost nothing.² The trap is the seller’s. It is the trap of building a thing whose whole worth is something anyone can sell. And it is the trap of taking fast growth as proof that the thing is worth something.

The copy on Chidi’s screen raised a harder question than whether such a company can last. The commodity trap is usually told as a story about survival: how to avoid being copied, absorbed or undercut. This part of the book has a different word in its title. It is about building things that deserve to exist. The two are easy to confuse and they come apart. A product can survive by holding its users against their will, and a product can die after serving them well. So this chapter asks: When anyone can rebuild what you built, on a machine you do not own, what would show that it deserves to exist, and not only that it has survived?

The print edition offered a programme for survival. It told founders to “build value that resides not in the intelligence but in the consciousness”, the “human judgement, domain knowledge, and trust” that the intelligence “serves but cannot replace”. It called the knowledge of an old trader “a platform around consciousness”. This edition withdraws that claim, for two reasons. Consciousness cannot be scarce among people: every founder has it, including those whose companies die. And the print edition’s own story refutes it. It proposed to capture the trader’s judgement in a database, which would make it exactly the kind of thing a machine can learn.

What this chapter holds is narrower. Surviving is not the test. The test is what a thing does to the people it serves, to those who pay for it, and to those who pay without being asked. It is measured against what they would have without it. The market settles a different question: whether enough people will pay. Along the way the chapter asks what, in a product built on a rented machine, a builder can call his own. It is less than founders hope, and it asks more of them.

14.1 Why Wrappers Die

To say that most AI startups will die is to make a prediction, and nobody knows how many such companies there are. What can be said is narrower. One kind of company dies, and for a reason that can be stated.

Chapter 6 named it a thin wrapper: a product whose main value is access to something the underlying model already does. When the model improves, the wrapper is replaced by the layer it was built on. The reason fits in one line: as of 2026, a difference that can be written down is a difference a model can be taught. Chidi’s difference was written down. It was his instructions to the model, and his screen.

There is a simple test, and this chapter is its home. The fortnight test asks: can a competitor rebuild your product in a fortnight, using the same model? If so, the product you can see is a wrapper. Whatever business there is lies in what the copy would lack. The Nairobi team needed twelve days. As of 2026 a fortnight is generous; the machine on Chidi’s desk needed an afternoon. The test does not measure how hard you worked. Chidi spent weeks on his instructions. A competitor matched the result with different instructions. The effort was real. It was no defence.

Why, then, do capable and well-warned people keep building wrappers? Because a wrapper is what the market rewards first. Users want help now. Investors want growth they can see. And for its first months a wrapper looks exactly like a real business. It has revenue. It has users who value it. The numbers are right, and each year they stay right for a shorter time. Speed is the commodity trap’s perfect bait. It produces the look of progress while it uses up the time that anything slower would need.

Some companies that began as wrappers did not die. They gathered records nobody else had, fitted themselves into the daily work of a business, or became the place a community went. The wrapper was the entry point, not the business. The business was what grew around it. What can grow is the subject of the next sections. The point here is about time. Growing anything takes months, and a wrapper’s months are short.

The Tirukkuṟaḷ, the Tamil book of conduct, has a chapter of ten couplets on deliberation before action. It is counsel to a king about his undertakings, not to a founder, and I borrow two of its sentences. In V. V. S. Aiyar’s translation: “There are enterprises that tempt with a great profit but which perish even the capital itself: the wise undertake them not.” And: “Decide not upon any action except after careful deliberation: it is a fool who undertaketh first and sayeth in his heart, I shall think afterwards.”³ Chidi undertook first. He meant to think afterwards, and afterwards came in January.

Set the first couplet beside the fund, though, and it is less simple. The Kuṟaḷ speaks to one king with one treasury: do not start what can eat your capital. The fund’s whole design is to start many such things. It knows that most will eat their capital, and counts on the few that do not to pay for the rest. Both can be reasonable. They are reasonable for different people. The fund could afford ten companies like Chidi’s; Chidi could afford one. The couplet’s prudence belongs to whoever carries the loss, and the fund’s odds belong to whoever spreads it. A founder who takes the fund’s odds as his own has accepted a risk that was priced for someone else.

Chapter 9 promised that this chapter would return to a pause. It told how speed removed “the pause in which doubt used to arrive”. That was when Priya Chandran, a test lead in Bengaluru, spent two weeks redesigning tests for a part of the product about to be retired. The pause Chidi lacked was the same pause, placed before building instead of before testing. The Kuṟaḷ puts thinking before undertaking. Speed reversed the order: it became cheaper to build and see than to think and decide. A copy in an afternoon makes building cheaper still, so the pause will not come back by itself. It has to be built. One part is a set of questions asked before the code (14.3). Another is a habit of listening that does not wait for the numbers to fall (14.2). It also has to be paid for. A fund that asks only for growth each month has bought the speed and left the pause to the founder’s conscience. Those who choose a change and take its gains should carry most of its cost. The same holds for the clock: whoever sets it should pay for the pause.

The users who left Chidi for the free app were right to leave. It was cheaper and, for most of them, better. This book has told of a price app in Cebu routed around by an assistant, and of a free platform that served Vietnam’s fishing communities better than one scientist’s sensors. Chidi’s case is plainer than either. His product was better at nothing except being first. When the commodity becomes free, the buyer gains. A seller who offered nothing else has no claim against that gain.

14.2 The Phone Call That Changed the Product

In Recife, Isabella Cavalcanti had built a wrapper too.

Her product made documents for MEIs. An MEI, a microempreendedor individual, is Brazil’s simplest registration for a very small business run by one person: a cleaner, a hairdresser, a mechanic. Her product drafted their service contracts, explained their tax forms and replied to their questions, for forty reais a month. It ran on someone else’s model, with instructions she had written for Brazilian law. By September 2024, three months after launch, it had twelve hundred subscribers. Her growth looked like Chidi’s. So did her danger.

Since August she had kept one habit. Every week she telephoned ten subscribers and asked them one question: what would you do if this product disappeared tomorrow? Most said they would find something similar. That was the reply she had expected and feared. It meant that her product could be swapped for another, and so that it was dying on a timeline she had not yet worked out.

In September she telephoned Cláudia, who runs a small cleaning service in Jaboatão dos Guararapes, the city on Recife’s southern edge.

“I would call my cousin’s husband,” Cláudia said. “He is a lawyer. He is expensive. But he knows the rules.”

Isabella asked which rules.

“The rules for Jaboatão,” Cláudia said. “The rules for a cleaning service. Not the general rules. The specific ones. The ones that changed last month. Your product knows them. The free version does not.”

Cláudia had said exactly what she meant. Her reply named her real alternative, and it was not the free app. It was a person: a lawyer in the family. He costs money, he knows the local rules, and he would have to face his wife’s cousin if he got them wrong. Earlier in this book, a bank’s compliance officer asked Diane, a software founder in Kigali, whom he would call if her system broke at the end of the month. Cláudia already knew whom she would call. Isabella’s product was competing with that lawyer.

So Cláudia was paying for two things that the machine’s fluency did not supply by itself in 2024. Someone kept up with what had changed. And someone would be there when a document went wrong.

Isabella’s instructions had been written for the national rules, which she knew well. Some of what Cláudia needed was national too. Since September 2023, an MEI who issues an invoice for a service has had to issue it through a single national system.⁴ Her city was another matter. A cleaning service also lives under the rules of its own city, and Jaboatão is not Recife. The instructions Isabella had tuned by hand happened to fit Cláudia’s case that month. They would not stay that way. Rules change, and a set of instructions does not read the news.

She stopped improving her instructions and started building what the call had shown her. She hired a paralegal to read, every week, the changes in tax and labour rules and in the rules of the cities around Recife. She built a small database of those rules and gave it to the model with each request. A contract drafted on a Thursday then followed that week’s rules, not the model’s memory of an older year. By January 2025 the database covered the cities of the metropolitan region. It did not cover the country. It was small, and it was current.

Then she took one more step. Brazilian law reserves legal advice to lawyers.⁵ Her product did not call itself advice. Cláudia used it as advice, and so did most of the twelve hundred. Isabella also paid a lawyer to review every template and to approve each change. That did not make her company a law office, or every contract right. Whether a product used as advice stays within that law is a question for Brazilian lawyers, and this chapter does not settle it. But if a contract was wrong, someone qualified would have to explain it and put it right, and that someone had a name.

Notice what this cost, and who paid. The paralegal’s week and the lawyer’s hours came out of the forty reais. Isabella’s ten calls a week came out of her own evenings. Every practice uses someone’s hours, and here they were hers, because she owned the company. None of it was free, and none of it grew as fast as the subscribers.

The machine has a share of this work, and it is growing. As of 2026, part of the paralegal’s job is something a general model with search can do. Asked about a city’s new licence rule, it will often find the decree, sometimes before the paralegal does. What Isabella offers is upkeep that someone pays for every week, and a named person who answers when the upkeep fails. That gives her no advantage she can keep. Any competitor willing to pay a paralegal and a lawyer can do the same. It is a cost she chose to carry, because her customers had begun to rely on her.

And a cost carried for customers does not pay for itself. A cheaper rival could offer the same screen with no paralegal and no lawyer. Most subscribers would not see the difference until a contract failed, and by then they would have chosen. The market might reward that rival. If the weekly upkeep is only a comfort, customers have the right to choose without it. If it is necessary, because a wrong contract or tax form can lose a cleaner her clients or bring a fine, then the question is who should pay for it. It could be the customers, or the companies that sell the models. It could be the public, through a rule that holds every such product to the same standard. Isabella’s choice does not settle that question. It only shows what the upkeep costs.

It would be easy to tell her story as an escape, and to use her as proof that the escape works. That would judge a decision by how it turned out, the error called outcome bias.⁶ Isabella’s product is alive in October 2026. It may not be next year. If a large company builds the same upkeep for the whole country, she will face Chidi’s question at a larger scale. Her survival does not prove that her choice was right. Nor does Chidi’s collapse, by itself, prove that his choice was wrong. What tells against it is what he could have known before he began, and the next section shows how little it would have cost to ask. What makes Isabella’s choice worth studying is the choice itself. She asked her customers what they would do without her, and she believed them.

14.3 Six Questions

The lesson gathers into six questions. They are a discipline of self-interrogation, not a checklist. The replies change as the product changes, and a product that passes in January can fail in July. Chidi wrote the questions on a whiteboard in Yaba in January 2025, beside the screen with the copy on it. Here they are, with what he wrote, and with one question changed.

If the model underneath disappeared tomorrow, what would remain? The screen would remain, and a screen is easy to make. So ask further. What records do you hold that exist nowhere else? What would it cost a customer to move? What would survive a change of technology? Chidi wrote: nothing. Isabella, by then, could have written: a current database of the rules for the cities around Recife, a paralegal’s week, and a lawyer’s approval.

Can a competitor rebuild the product in a fortnight, using the same model? This is the fortnight test. Founders fail it because they confuse how hard it was to build their product with how hard it is to build something just as good. Chidi did not need to write anything. The reply was on the screen beside him.

Does the product become more useful to each user as more people use it? Some products do, because each user adds something the others can use. Economists call this a network effect, and the next chapter takes it up. Chidi’s did not: each user got the same replies they would have got alone.

If the model’s maker built your feature tomorrow, who would stay, and why? Users who stay usually do so for one of three reasons. The product does something the platform does not; or switching away is too costly; or the relationship is something a product announcement cannot replace. The second reason needs a warning. A user who stays because leaving is costly has not chosen the product. He is held by it. That can keep a company alive for years, and it is still no reason for the company to exist. A builder who counts it as one has begun to confuse surviving with deserving. Chidi wrote: nobody.

When the product is wrong, who answers? This question replaces the print edition’s fifth, which asked for the product’s “judgement layer” and promised that “Deep judgement is a moat”. Depth is not a moat, and this book does not sell answering as an advantage. The question here is plainer. Your product will be wrong sometimes. When it is, who explains, who undoes what can be undone, and who puts right what can be put right? A wrapper usually answers for nothing. Its terms of service disclaim its output, and the terms of the model underneath disclaim it too. Chidi’s product had drafted contracts for small shops. He had never asked what happened when one was wrong. He wrote: nobody. Then he looked at the word for a while.

Would the product be worth building if the model were free? Chapter 5 showed the price of a given level of machine performance falling fast.⁷ As of 2026, for many tasks the model is close to free already. If the whole offer is making the machine easier to reach, a free machine removes the offer. If something would remain, then the product uses the machine, and does not merely consist of it. Chidi wrote: no.

What he wrote made a short column: nothing; an afternoon; no; nobody; nobody; no. He did not wipe the whiteboard.

Look at what the six ask. Five of them ask whether a product can last. Only the fifth asks who answers when it is wrong. A product could pass all six and still not deserve to exist. Deserving needs two more questions, and they come at the end of this chapter, with the people who raise them.

Two things about these questions. First, they are better asked before building than after a collapse. Asked in March 2024, they would have cost Chidi an afternoon. Asked in January 2025, they cost him nothing more, because he had already paid. That is the promised pause, made into a habit. Second, most people who make products do not decide what gets made. An engineer, a designer or a product manager inside a company may see that the thing they are building fails all six. A company should not ask its people to question the purpose of their work while it keeps from them the decisions that set it. These questions belong first to whoever decides. Everyone else can ask them of that person, in writing, before the work begins.

Receipts, please

The sixth question leans on a figure about falling prices, and the figure is narrower than it sounds. Epoch AI, a group that tracks the costs of machine learning, published it in March 2025.⁷ For each score on six public tests of models, such as questions in science, mathematics and code, it found the cheapest model that reached that score. Then it followed that model’s price over time. The price of reaching a fixed score fell by between 9 and 900 times a year, depending on the score.

Three limits follow. The figure is the cheapest way to reach a level some model has already reached. It says nothing about the price of the strongest model at any moment. The tests measure set tasks and people’s votes between replies, and a founder’s task may not be among them. And the authors note that the fastest falls came in the most recent year they measured, so it is less clear that those falls will last.

My reading. The figure supports the sixth question for a product that needs only last year’s level of skill. A product that needs the best model there is pays that model’s current price, and that price is the first dependency in 14.5.

14.4 Feature vs. Infrastructure

The distinction is old. A feature solves one problem and sits on someone else’s platform. Infrastructure is what other people build on, or rely on every day without thinking about it. Earlier in this book the word meant something larger: AI itself becoming part of the roads and pipes of daily life, below notice. Here the word is smaller. It says where a product sits: as a stop on someone else’s road, or as a road that others travel.

Most products built on AI models are features. They solve a real problem in a particular way, and the solution is useful. But they add a layer that the owner of the platform underneath can add for itself whenever it chooses. It is tempting to draw a rule from a few famous companies that became platforms. The rule says that the companies which survive a change of platform are the ones that become platforms themselves. Such a rule is drawn from the survivors. The many companies that tried to become platforms and failed are missing from it. And being a feature is no disgrace. Many good things are features. When a platform absorbs one, its users usually gain. A founder whose product becomes a feature of something larger can sell it, merge it or close it well, and nobody need be worse off for that.

Infrastructure is harder and slower. Isabella’s database is a small example. It serves very small businesses around Recife as a road they travel each week without seeing it. A road, once built, is used even when the engines change. Isabella can change the model under her product next year, and the road of current rules will still be there. The Chinese philosopher Zhuangzi has a line that fits, though he wrote it about how names come to fit things: “A road is made by people walking on it.”⁸ Isabella’s road was made that way. Cláudia’s call marked its first stretch, and every question her subscribers have asked since has shown where it needed to go.

A road also brings a duty that a feature does not. When Chidi’s product died, his users lost almost nothing. If Isabella’s closes, Cláudia loses the place where her rules were kept, and she may not know it has gone until a contract fails. Infrastructure makes people depend on it. For the builder, that dependence is easy to count as value. It is also a debt. A builder who has become a road for others owes them more than a warning, in proportion to how much they rely on it. He owes them the date each rule was last checked, a way to take their own records with them, and notice before the road closes.

The print edition said that holding the short term and the long term together, shipping quickly while building slowly, “is an act of consciousness”. It is simpler and harder than that. It is a budget. Someone must pay for the slow part while the fast part pays the bills. Whether a founder can afford it depends less on his awareness than on whose money he has taken. It also depends on what that money asks of him each month.

14.5 Building Your House on Someone Else’s Land

A rented model brings three dependencies.

The first is price. Your margin, what is left of each payment after your costs, is set by the model maker’s prices. Those prices have fallen fast. They are still somebody else’s decision, and a change in price, or in the limits on use, can change in a day what your product costs to run.

The second is capability. Your product can do what the model can do. When the model improves, every product built on it improves too, so the improvement sets you apart from no one. When a new version behaves differently in the place your product depends on, your product changes with it, and you cannot repair the model. Chidi spent two weeks in September repairing replies that an update had changed.

The third is strategy. It is the most dangerous, because it is the least visible. The model’s maker has plans it does not share with you. When it decides to offer what you offer, you learn of it with everyone else. This is old. In 2002 Apple built a search tool, Sherlock 3, into a new version of its system. It did much of what a small company’s program, Watson, had done since 2001.⁹ Developers for Apple’s computers and phones still call this being “Sherlocked”. In June 2024 a news report listed apps whose work Apple had just announced it would build into its phones, among them a writing assistant.¹⁰ Chidi’s case was the same. He was Sherlocked by the company whose model he rented. A model maker may say that it will not compete with the developers who build on it. Even then, a statement is not a contract, and a company whose life depends on a competitor’s restraint is built on a promise it cannot enforce.

The practical defence is to treat the model as a part that can be replaced, as a well-built program treats its database. Build around your own records, your own rules and your users’ work, and change the engine when a better one arrives. Isabella did this: her database fed whichever model she used. But changing engines protects you from the second dependency, and much less from the third. Some model makers now publish a model’s weights, the numbers that make it work, so that anyone may run it on their own computers.⁷ A founder who runs such a model removes much of the third dependency, because nobody can withdraw it or change its terms. He also changes the first: he pays for computing power, which many firms sell, instead of one maker’s price. He may give up some capability, since the model he can run may not be the best there is. None of this helps with the commodity trap. Every competitor can run the same model, so a wrapper around it is as easy to copy as before. If your product is only the engine made easier to use, a better engine owned by someone else still leaves you with nothing.

That leaves a question that consciousness was never going to settle. When the model is rented, the screen can be copied in an afternoon, and the instructions fill a page anyone could write, what in such a product is yours?

The best-known reply in Western philosophy is John Locke’s. In his Second Treatise of Government, published in 1690, he argued that a person owns his own labour, and so comes to own what he mixes it with. “The labour of his body, and the work of his hands, we may say, are properly his. Whatsoever then he removes out of the state that nature hath provided, and left it in, he hath mixed his labour with, and joined to it something that is his own, and thereby makes it his property.” He added a condition, which philosophers call his proviso. This holds “at least where there is enough, and as good, left in common for others.”¹¹

Chidi had mixed his labour with something. He had written instructions for weeks. But Locke’s argument is about things held in common, which no one yet owns: fruit on a wild tree, water in a stream. The model belonged to someone. Labour spent on another person’s property makes it no more yours than painting a rented house makes it your house. The American philosopher Robert Nozick, writing in 1974, pressed Locke harder with a joke that is also an argument: “If I own a can of tomato juice and spill it in the sea so that its molecules (made radioactive, so I can check this) mingle evenly throughout the sea, do I thereby come to own the sea, or have I foolishly dissipated my tomato juice?”¹² Chidi had poured his tomato juice into someone else’s sea. Neither Locke nor Nozick was writing about software, and I do not claim otherwise. What their arguments expose is an assumption: that a builder’s labour entitles him to everything his product uses.

Locke’s argument also has a history that a chapter set partly in Lagos should not leave out. Scholars have shown that Locke was involved in the affairs of England’s American colonies. Some argue that his theory let him treat Native Americans’ hunting land as empty and open to be taken.¹³ I borrow his question, and refuse that use of it.

The Īśa Upaniṣad, one of the shortest of the Upaniṣads, opens with a verse that meets the same question from the other side. The verse ends with the words mā gṛdhaḥ kasya svid dhanam, and its translators divide over them. Their division is this chapter’s question. F. Max Müller translated the line: “When thou hast surrendered all this, then thou mayest enjoy. Do not covet the wealth of any man!”¹⁴ On that reading the verse is a rule about other people’s property: do not want what is theirs. Śaṅkara, the great commentator of the Vedānta, offered a second reading. In S. Sitarama Sastri’s translation: “Or, it may be thus interpreted. Do not covet. Why? ‘Whose is wealth?’ is used in the sense of an objection; for nobody has any wealth which could be coveted.”¹⁵ On that reading the question expects the reply “nobody’s”: all this belongs to the Self, and what one covets is unreal. Śaṅkara gives this verse to the renunciant who knows the Self, and the next verse, on a life of work, to everyone else. It is counsel for the renunciant, not a law against owning things.¹⁶

The verse is about the Lord who dwells in all things, and about renunciation. It is not about software, and I borrow only its question. The two lamps also address different things. Locke asks who may keep other people out: what gives one person a right against the rest. The Īśa asks what a person should hold on to at all. But on Chidi’s case they pull against each other. Chidi wanted Locke’s “mine” for something he had only been given to use. The Īśa, on either reading, tells him to enjoy only by giving up, and not to mistake what he uses for his own. Müller’s reading tells him not to covet the model maker’s wealth. Śaṅkara’s second reading tells him that nothing he holds is his to covet.

This book holds a narrower “mine” than Locke’s, and a firmer one than Śaṅkara’s second reading. Chapter 1 met another side of Locke: the philosopher who called a person “a forensic term, appropriating actions and their merit”.¹⁷ To appropriate an act is to take it as your own. What is yours, in a product built on a rented model, is neither the model, nor the screen, nor the instructions, which can all be copied. It is the acts you take as yours: what the product tells the cleaner in Jaboatão, what it does when it is wrong, and what you do then. A founder can own code, a brand and contracts. None of that makes him the author of the model’s abilities, or frees him from answering for what he chose to build. The Bhagavad Gītā names five causes of every act, of which the agent is only one. This book read it earlier to mean the same thing: to own an act is not to claim that you caused all of it, but to answer for the part that was yours.¹⁸ The lawyer’s approval on Isabella’s templates is hers in that sense, because she chose it and pays for it. Chidi’s product answered for nothing, so in this sense nothing in it was his.

This kind of ownership gives no advantage. A competitor can take on the same duty, and a founder who takes it on pays for it. It is closer to a liability in the plain sense of the word: something you owe. Nor does it show that a product deserves to exist; a product can be well answered for and still not be worth having. What it marks is the line between using the machine and hiding behind it.

14.6 What Emeka Knew

Chidi’s uncle Emeka has traded cloth in Balogun Market since 2003. He is Chidi’s mother’s brother, and his shop is on Lagos Island. He buys from suppliers across Nigeria and abroad, and he sells to retailers in Lagos, Accra, Abidjan and Douala. The business has no app and no investor. It has Emeka, in a plastic chair in front of his shop, taking calls and making decisions on which several families’ incomes depend.

In February 2025, a few weeks after the whiteboard, Chidi went to sit with him. He did not go for advice. He wanted to be near a business that had lasted. He meant to stay a week. He stayed two months more.

On the second day a young man arrived from Aba, a city known across Nigeria for its workshops.¹⁹ He was called Tobechukwu. He wanted to sell through Emeka’s network. He sold garments sewn in Aba from Ankara cloth that, like most of the Ankara sold in Nigeria, had been printed abroad.²⁰ He brought samples. The seams were straight, the patterns matched at the joins, and his prices were good. He had references: a cousin in the market, and two retailers in Surulere who had already ordered from him.

Emeka turned the garments over in his hands and looked at the seams. Then he asked questions. He asked about Tobechukwu’s father, and how long the family had been in the trade. He asked who did the sewing, and whether Tobechukwu owned the machines or rented them. And he asked, as if in passing, about the rains: how the road from Aba had been in the last wet season.

These were not small talk. Each question was a probe, but they were probes of different kinds. Two psychologists, Daniel Kahneman and Gary Klein, gave a test for when an expert’s quick sense deserves trust: a world with clear signals, and years of feedback.²¹ Some of Emeka’s questions pass it. A man who travels the road from Aba often will know at once what the rains did to it. If he hesitates, he may not make the journey, or not often. Whether he owns his machines says something about whether he will still be in business next year. These are questions about the supplier, and Emeka had twenty-two years of experience to check them against.

The question about the father is a different kind. It asks about the supplier’s family, not about him. Chapter 4 told of Rajan Subramanian, a cloth maker and exporter in Coimbatore, turning away the young owner of a weaving unit near Erode, whose records were clean. His nephew asked whether he would have seen the same thing if he had known the man’s father. Chapter 7 showed Rajan judging his best supplier, Senthil, against Senthil’s father for years. The Tirukkuṟaḷ (510) warns that “to suspect a man whom thou hast found worthy” leads, like its opposite, “to endless ills”.²² A father’s name is a kind of record. It says that someone in the family once kept their word. But it is a record the newcomer cannot add to. Tobechukwu could deliver on time for ten years and never change who his father was.

Tobechukwu hesitated on the rains.

Emeka did not refuse him. He placed an order far smaller than the samples warranted, as a test. If it came on time and matched the samples, the next order would be larger. If it came late or short, there would be no next order. This is one trader’s way of giving credit: trust extended a little at a time, and withdrawn at once. It has a cost. The test order is cheap for Emeka. For Tobechukwu it is a season of small sales, on the strength of a cousin’s word. A supplier whose father had traded in the market would have been tried with a larger order, or not tried at all. The caution is Emeka’s, and the newcomer pays for it.

Chidi had built his first company to help people like his uncle: owners of small businesses, making hard decisions. But he had heard the problem wrong. He had heard that such people needed better tools. Watching Emeka, he saw that they had tools already, and good ones. The print edition drew a lesson from this. The trouble, it said, was that such tools “cannot compound”: when Emeka retires, his knowledge retires with him. So it had Chidi resolve to capture “the accumulated human judgement of traders”, “captured, structured, and made compoundable”, and it called the result a platform around consciousness.

That plan undoes itself. If Emeka’s judgement can be captured and structured, it can be written down. As of 2026, a difference that can be written down is a difference a model can be taught. A platform built from captured judgement would be the next thing a machine could rebuild in an afternoon. As a defence against copying, it fails. Whether such a record would be worth keeping is a different question.

But look more closely at what Emeka knew. Part of it was a record nobody had kept: which suppliers delivered on time, in which seasons, with which faults. The same is true of much expert knowledge. Part of what passes for deep knowledge can be a record nobody kept, and a kept record can become a rung for someone new. Part of what Emeka knew was not knowledge at all, but kinship: who is whose son. And part of it was not knowing but acting. Emeka places each order with his own money, and when a delivery is short, it is Emeka whom his buyers in Accra telephone.

On the first part, the machine has the better of the argument. A record of deliveries does not ask about anyone’s father. Kept over time, it would let Tobechukwu’s deliveries speak for him, so that a newcomer could build a history where he has no family name to stand on. A late delivery can be a flooded road or an order changed at the last minute, so the record must let the supplier explain it. A machine keeps such a record well, and better than Emeka’s memory, which forgets the small deliveries and remembers the families. Built carelessly, the same tool could do the opposite. If it learned from the traders’ old cautions, it would learn the question about the father too. It would then ask that question of every newcomer at once, where nobody could hear it asked. What Chidi went on to build in the market, from August 2025, belongs to Chapter 15.

Here the chapter’s question meets Emeka’s market. Imagine two products built from it. One keeps a plain record of deliveries, open to the people it describes, so that a newcomer’s work can be seen. The other turns the traders’ old cautions into a score, and sells it to them. Both might survive. Both might grow. The second might even earn more, because it would sell established traders the comfort of their own habits. Its survival would show only that it had found people willing to pay. Every Tobechukwu would pay too, in a way no chart of growth records.

So the chapter can say something definite, though not everything. Surviving is not the same as deserving to exist. Growth shows that some people want a thing. It does not show what the thing does to the people who use it, or to those who pay for it without being asked. The market does not settle what deserves to exist, and nobody else is given that to settle instead. Within the law, whoever decides what gets built still decides. They answer to those who pay without being asked, as well as to those who pay.

Two questions belong beside the six, and the market asks neither. The first is who pays without being asked. Tobechukwu paid for Emeka’s caution with a season of small sales. The designer and the support coordinator paid for Chidi’s speed with their jobs. The second is whether the people a product describes or judges can see what it says about them, leave it, and dispute it. A record of deliveries that Tobechukwu could read and correct would pass. A score sold to established traders, which he could never see, would fail.

Chidi’s first company served its users for a while and harmed almost no one outside it. It need not have been saved: its users were better off with the free app. That does not make every cost of its closing fair. Two people lost their jobs, and Chidi lost a year’s savings. The fund spread its loss across many bets, and most of the users were already better served by the free app. The closing was no loss to its users, and a real one to the people who built it. A product that turned Emeka’s caution into a score might live for years and not deserve to. What else it takes for a product to deserve to exist is a question Part IV keeps asking. The market does not settle it.

What would prove this wrong?

My claim in 14.1 is dated and checkable. As of 2026, a difference that can be written down is a difference a model can be taught. So a product whose only difference is written down dies when the model’s maker offers the same use. Two kinds of evidence would show me wrong.

The first is a thin wrapper that lasts. Take a product whose offer is still only a screen and a page of instructions, two years after its model’s maker gave away the same use. If it has kept most of its paying users at its old price, the trap is weaker than I say. A product that lasted by gathering records nobody else had, or by fitting itself into its users’ daily work, does not count. By 14.1’s definition it stopped being a wrapper, and the test is of the product as it stands.

The second is instructions that do not travel. If a page of instructions tuned by hand for one model could not be matched on another model within a fortnight, that page would be worth more than I allow. Chidi’s afternoon would then be one case, not a rule.

Where to look: the products that platforms absorb each year, like the apps Sherlocked in 14.5, and the price series behind the sixth question.

Ending: One Delivery Is One Delivery

Tobechukwu’s first delivery came in April, two days before the date Emeka had given him. Chidi was there, in the last week of his two months in the market. Emeka opened the bales himself. He counted, held a shirt up to the light at the door, and turned it inside out to look at the seams. Everything matched the samples.

Emeka placed a second order. It was larger than the first. It was still smaller than the samples had warranted in February.

Chidi asked him why.

“One delivery is one delivery,” Emeka said.

Chidi did not know whether that was wisdom or habit, and he suspected that Emeka did not know either. Tobechukwu folded the order slip into his shirt pocket, thanked them both, and went to find the bus back to Aba.

Practice

The afternoon rebuild. Choose one product you make or help to make, or one piece of work a customer pays you for. Describe it in plain words to an AI tool, and ask it to make a rough copy, or to describe exactly how a competitor would copy it in a fortnight. Then list what the copy lacks. Beside each item, write two things: who answers for it when it goes wrong, and whose hours keep it up. Some items will be records you keep, some relationships, some promises. Some lines may stay empty. Then ask one more question: if a free copy appeared tomorrow, would the people you serve be worse off, or better off? Write the answer down, even if it is “better off”. If you decide what gets built, keep the page and read it again in three months. If you do not, give the page to the person who does, and ask for an answer in writing. The exercise will not tell you whether your work will last, and it promises nothing about your job. It tells you which parts of it someone is answering for, and whose time pays for them.

Words to keep

commodity: a product that is the same whoever sells it, so that buyers choose on price.
feature: a product that solves one problem on top of someone else’s platform, and that the platform’s owner could add for itself.
fortnight test: can a competitor rebuild your product in a fortnight, using the same model? If so, the visible product is a wrapper, and whatever business there is lies in what the copy would lack.
infrastructure (in this chapter): what other people build on, or rely on every day without thinking about it; a road, not a stop on someone else’s road.
kasya svid dhanam: words from Īśa Upaniṣad 1; read by Müller as “of any man” (“Do not covet the wealth of any man”), and by Śaṅkara, in a second reading, as the question “Whose is wealth?”, expecting the reply that nobody has wealth to covet, since all belongs to the Self: counsel for the renunciant, not a law against owning things.
mā gṛdhaḥ: “do not covet” (Īśa Upaniṣad 1).
MEI (microempreendedor individual): Brazil’s simplest registration for a very small business run by one person.
model dependency: a product’s reliance on a model it does not own, for its price, its abilities and its future.
proviso: Locke’s condition that labour makes property only “where there is enough, and as good, left in common for others”.
Sherlocked: developers’ word for having your product’s work built into the platform it runs on.
thin wrapper: a product whose main value is access to something the underlying model already does (Chapter 6).

Sources

OpenAI, “Hello GPT-4o,” 13 May 2024 (openai.com): “We are making GPT‑4o available in the free tier”. The real event used for the composite’s timeline; Chidi’s company is a composite.
Adam Smith, An Inquiry into the Nature and Causes of the Wealth of Nations (1776), Book I, Chapter IV, paragraph 13 (Glasgow edition I.iv.13): “Nothing is more useful than water: but it will purchase scarce any thing”; quoted in Chapter 5.
Tirukkuṟaḷ 463 and 467 (chapter 47, “Deliberation Before Action”), trans. V. V. S. Aiyar, The Kural or the Maxims of Tiruvalluvar (Madras, 1916; Standard Ebooks). Tamil of 463: ākkam karuti mutaliḻakkum ceyviṉai ūkkār aṟivuṭai yār; of 467: eṇṇit tuṇika karumam tuṇintapiṉ eṇṇuvam eṉpatu iḻukku.
Comitê Gestor do Simples Nacional, Resolução CGSN nº 169, 27 July 2022, as amended by Resolução CGSN nº 172/2023 (start postponed from 3 April to 1 September 2023); Nota Técnica NT 2023.001, “Emissão da NFS-e pelo MEI” (gov.br/nfse): from 1 September 2023, MEIs issue service invoices (NFS-e) only through the national system.
Lei nº 8.906, de 4 de julho de 1994 (Estatuto da Advocacia e a Ordem dos Advogados do Brasil), Art. 1º, II: “as atividades de consultoria, assessoria e direção jurídicas” are among the “atividades privativas de advocacia” (planalto.gov.br).
Jonathan Baron and John C. Hershey, “Outcome bias in decision evaluation,” Journal of Personality and Social Psychology 54, no. 4 (1988): 569–579; see Chapter 11.
Ben Cottier, Ben Snodin, David Owen and Tom Adamczewski, “LLM inference prices have fallen rapidly but unequally across tasks,” Epoch AI data insight, 12 March 2025: declines of “9x to 900x per year” depending on the performance milestone; see Chapter 5. On published weights: DeepSeek-R1 (January 2025), whose README states, “This code repository and the model weights are licensed under the MIT License”, and that the series “support commercial use” (github.com/deepseek-ai/DeepSeek-R1, section 7). For the receipts in 14.3: for each score on six benchmarks (among them MMLU, GPQA Diamond, MATH-500, HumanEval and the Chatbot Arena rating), the insight takes the lowest-priced model reaching it and tracks that price, in US dollars per million tokens, for models from November 2021 to February 2025; the authors add that the fastest falls came in the most recent year, so it is less clear that they will persist (epoch.ai/data-insights/llm-inference-price-trends, checked 4 November 2026).
Zhuangzi, chapter 2, “Discussion on Making All Things Equal” (道行之而成), trans. Burton Watson, The Complete Works of Chuang Tzu (New York: Columbia University Press, 1968). The same line is quoted in Chapter 3.
Karelia Software, “Watson FAQ” (karelia.com): Watson first released in 2001; Sherlock 3 first released with Mac OS X 10.2 (“Jaguar”) in 2002, overlapping much of Watson’s function; Karelia “was not involved in any aspect of Sherlock 3”.
Bobby Allyn, “Apple just made your app obsolete? You’ve been ‘Sherlocked’,” NPR, 17 June 2024. The report’s apps include a writing assistant (Grammarly); its dating of the term to “the 1990s” is loose.
John Locke, Two Treatises of Government (London, 1690), Second Treatise, chapter V, “Of Property”, §27 (Project Gutenberg).
Robert Nozick, Anarchy, State, and Utopia (New York: Basic Books, 1974), pp. 174–175, in the section “Locke’s Theory of Acquisition”.
Alex Tuckness, “Locke’s Political Philosophy,” Stanford Encyclopedia of Philosophy (substantive revision 6 October 2020), section on property, citing James Tully, An Approach to Political Philosophy: Locke in Contexts (1993), and Barbara Arneil, John Locke and America: The Defence of English Colonialism (1996). Locke’s Essay is cited in Chapter 1.
Īśa Upaniṣad 1, trans. F. Max Müller, The Upanishads, Part I, Sacred Books of the East 1 (Oxford: Clarendon Press, 1879), p. 311. Sanskrit: īśā vāsyam idaṃ sarvaṃ yat kiñca jagatyāṃ jagat / tena tyaktena bhuñjīthā mā gṛdhaḥ kasya svid dhanam.
Śaṅkara, commentary on Īśa Upaniṣad 1, trans. S. Sitarama Sastri, The Upanishads and Sri Sankara’s Commentary, vol. 1 (Madras: V. C. Seshacharri, 1905; a reprint of the first volume), p. 4. Sastri’s first reading, like Müller’s, takes the words as “anybody’s wealth”, which Śaṅkara glosses as “wealth either yours or another’s”.
Śaṅkara, commentary on Īśa Upaniṣad 1 and 2, trans. Sitarama Sastri (1905), pp. 4–5: “he who knows the Atman should renounce the three-fold desire of son, etc.”; the next mantras are “for the benefit of him who does not know the Atman” (p. 5; OCR corrected).
John Locke, An Essay Concerning Human Understanding, 2nd ed. (1694), Book II, Chapter 27, §26, modernised spelling; see Chapter 1.
Bhagavad Gītā 18.14, trans. Kashinath Trimbak Telang, The Bhagavadgîtâ, Sacred Books of the East 8 (Oxford: Clarendon Press, 1882), p. 123: the five causes of every action; the reading that to own an act is to answer for one’s part is this book’s (Chapter 3).
“Made in Aba: A changing perception,” Blueprint (Abuja), 3 April 2025 (News Agency of Nigeria): Aba as a commercial and industrial hub for textiles, garments and shoes.
Jimisayo Opanuga, “Nigeria loses billions as nearly 90% of Ankara fabrics are imported,” The Guardian (Lagos), 29 May 2025, reporting Consonance Investment, “Who is Dressing 220+ Million Nigerians?”: imported “primarily from China and India”.
Daniel Kahneman and Gary Klein, “Conditions for Intuitive Expertise: A Failure to Disagree,” American Psychologist 64, no. 6 (2009): 515–526; see Chapter 3.
Tirukkuṟaḷ 510 (chapter 51, “Testing of Men for Confidence”), trans. V. V. S. Aiyar, The Kural or the Maxims of Tiruvalluvar (Madras, 1916; Standard Ebooks): “To trust a man whom thou hast not tried and to suspect a man whom thou hast found worthy lead alike to endless ills.” Tamil: tērāṉ teḷivum teḷintāṉkaṇ aiyuṟavum tīrā iṭumpai tarum. Chapter 7 applies 510 to Senthil and 509 to the checklist refusal.

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 founders should “build value that resides not in the intelligence but in the consciousness”, the judgement and trust that the intelligence “serves but cannot replace”; that holding the short and long term together “is an act of consciousness”; that a trader’s decisions affected families “in ways that no algorithm could map and no model could score”; that traders’ judgement, “captured, structured, and made compoundable”, would make “a platform around consciousness”, and that Emeka’s attention was “consciousness applied so consistently and for so long that it had become a form of capital”; that “Deep judgement is a moat”; that a product without network effects “has the same value at ten thousand users as at ten”; that a product a competitor can rebuild in a fortnight means “you have a wrapper, not a business”; that users held by switching that is “prohibitively costly” are a reason a product survives, with no warning that lock-in is not a reason to exist; that Balogun Market’s credit system is “trust extended incrementally, withdrawn absolutely”, without counting what the newcomer pays; that “the companies that survive platform transitions are the ones that become platforms themselves”; that some model makers “have signalled” they will not compete with developers; that Isabella’s product did “something no general-purpose model could replicate” and that, having “escaped it by accident”, she proved the escape: “The same force that killed wrappers strengthened her product”; that Cláudia “could not have said it that way”; and that Chidi’s second company was “the one that survived”. It titled 14.1 “Why Most AI Startups Will Die”. It named the first company ClarityAI and the second TradeGrid (both names belong to real companies); set Isabella in São Paulo and Cláudia in Guarulhos; placed MEIs in “the informal economy”; gave Isabella a database of three hundred municipalities within four months; had Tobechukwu make wax print in an Aba workshop; put Chidi under a ceiling fan; used “what this book calls conscious building”; and ended with Chidi building again, the six questions still on his whiteboard.
What I now hold: a thin wrapper dies because what it adds can be written down, and as of 2026 a machine can rebuild it in an afternoon; the fortnight test is generous, and it shows only that the visible product is a wrapper. Whatever business there is lies in what the copy would lack, and Chidi’s copy lacked the company. Speed is the bait, but the clock is also set by those who fund it: the Kuṟaḷ’s prudence (463) belongs to whoever carries the loss, and a fund’s odds to whoever spreads it; the pause Chapter 9 promised must be built (questions before code, listening before the numbers fall) and paid for by whoever sets the clock. When the commodity becomes free, its buyers gain. What a customer pays for, beyond fluency, may be upkeep and someone who answers; neither is an advantage a builder can keep. Lock-in keeps a company alive and is no reason for it to exist. Running a model whose weights are published removes much of the dependency on its maker’s strategy, but not the commodity trap. Upkeep carried for customers does not pay for itself: a cheaper rival doing less may win, and if the upkeep is necessary the question is who should pay for it. Infrastructure brings a debt to those who depend on it, in proportion to their reliance: the date each rule was last checked, a way to take their records with them, and notice before closing. The six questions stand, with the fifth changed to “When the product is wrong, who answers?”; five of the six ask whether a product can last. Locke’s labour makes nothing yours on another’s property (Nozick’s tomato juice); neither wrote about software, but their arguments expose the assumption that labour entitles a builder to everything his product uses; on Śaṅkara’s second reading nobody has wealth to covet, since all belongs to the Self, and the verse is counsel for the renunciant who knows the Self, not a law against owning things; owning code, a brand and contracts does not make a founder the author of the model’s abilities; what is yours in a product on a rented model is the acts you appropriate and answer for (Chapter 1’s Locke; Chapter 3’s partial authorship), which is a debt, not an advantage. Emeka’s probes divide: some are evidence, one is kinship, and the newcomer pays for the caution. Part of what Emeka knew is a record nobody kept, which a machine keeps better and which can be a rung for a newcomer; built carelessly, the same tool would scale the question about the father. Surviving is not deserving to exist; growth shows that some people will pay, not what a thing does to them or to those who pay without being asked. Neither survival nor collapse proves a choice; it is judged by what could have been known when it was made. The market does not settle deserving, and nobody else is handed it: within the law whoever decides what to build still decides, and answers to those who pay without being asked as well as to those who pay. Captured judgement fails as a defence against copying; whether a record is worth keeping is a separate question. Deserving needs two questions the market does not ask: who pays without being asked (Tobechukwu, the two employees), and whether the people a product describes can see what it says about them, leave it and dispute it; a record of deliveries must let a supplier explain a late one. A closing can be no loss to a product’s users and a real loss to the people who built it; that the product need not have been saved does not make every cost of its closing fair.
What changed my mind: the audit’s finding that the chapter stretches consciousness to cover a database of traders’ judgements; Chapter 5’s finding that consciousness cannot be scarce among people; Chapter 6’s account of wrappers; Chapter 10’s finding that depth is not a moat and that part of it is a missing record; Chapter 11’s warning about outcome bias; Chapter 7’s conviction of the question about the father; the discovery that both company names belong to real companies; the 2025 report on Ankara imports; Brazil’s law on legal advice; the objections that the afternoon copy was not the company, that the closing had not been shown to be fair, that five of the six questions concern survival, that responsible upkeep does not pay for itself, and that a record must let the person it describes explain it; and the fact that, in January 2025, a machine could rebuild a company like Chidi’s in an afternoon.
What survives: Chidi’s season, and his mother’s weather; “The numbers told him he was winning. He was evaporating.”; “What did I build? An interface. Some prompts. A subscription page.”; “Platforms build their own features”; the fortnight test; “speed is the commodity trap’s perfect bait”; “The wrapper was the entry point, not the business. The business was what grew around it.”; Isabella’s ten calls, and Cláudia’s cousin’s husband; the six questions as a discipline, not a checklist; feature against infrastructure, and the road “used even when the engines change”; the three dependencies, and the engine that can be swapped; Emeka’s plastic chair; “He had heard the problem wrong”; “Tobechukwu hesitated on the rains.”; the small order.

Next in the series: Chapter 15: Building Real Moats in an AI World.