# AI-Proof: What Machines Can Never Be — Volume III: Transformation, Part VI: The Inner Revolution, Chapter 30: Anti-Fragility in an AI World

By Nithin Palal · 21 November 2026 · 43 min read
Section: Artificial Intelligence · Series: AI-Proof: What Machines Can Never Be
Source: https://nithinpalal.com/ai/ai-proof/chapter-30-anti-fragility-in-an-ai-world

> A tester laid off as machines took over testing is told it is a chance to grow. Taleb’s word for things that gain from shocks fits the machines better than him. A person gains from a shock only with slack, and much of it is lent.
*Part VI looks at the person. Chapter 29 asked what the fear of obsolescence is afraid of. This second chapter asks whether a person can gain from the shocks machines bring to work, and who pays for the slack that the gain needs.*

← Previous: [Chapter 29: Fear, Ego, and Obsolescence](https://nithinpalal.com/ai/ai-proof/chapter-29-fear-ego-and-obsolescence) · Series hub · Next: [Chapter 31: The Moral Responsibility of Builders](https://nithinpalal.com/ai/ai-proof/chapter-31-the-moral-responsibility-of-builders) →

**A tester laid off as machines took over testing is told it is a chance to grow. Taleb’s word for things that gain from shocks fits the machines better than him. A person gains from a shock only with slack, and much of it is lent.**

**In Brief: For readers who have been told that a shock to their work is a chance to grow. Chapter 30 asks whether a person can come out of the disorder machines bring to work stronger, and what that strength is made of. A test lead in Bengaluru who grew through the change meets a cousin who did not; a pharmaceutical site in Hyderabad finds that two sound decisions can wreck a promise; an engineer finds that the company’s new safety is his new exposure. Nassim Taleb’s antifragility, read against Nietzsche’s military school of life and a verse of Bhartṛhari on obstacles, shows that the gain from disorder often goes to the whole while the parts pay, and that redundancy is built before it is needed, by those with slack to build it.**

**What gains from disorder is often the whole, while some of its parts break; a person gains from a shock only with slack she seldom made alone, so redundancy is built before it is needed, and whoever imposes a shock answers for it and may not call the harm a gift.**

## Opening: The Café Across the Road

The café was across the road from Priya Chandran’s office in Koramangala, in Bengaluru, and her cousin Naveen was already there when she came down.

It was March 2026. Priya had worked at the same software company for nearly ten years. She had begun as a test lead, with a team of twelve. In October 2024 a machine ran in eleven minutes a cycle of tests that had taken her team a week. Her team then built the platform that now tests the company’s code, and since October 2025 she had led the work of carrying it into every product the company makes. She lived with her mother, and a cat named Turing, in a flat a short walk away.

Naveen was her mother’s sister’s son. For about twelve years he had tested software by hand at an IT services firm across the city. In February the firm had let him go, with some others, and given him two months’ pay. He asked whether her company was hiring testers.

It was not. She knew why, because the reason was her job. She told him so, and he nodded, as if he had known before he asked.

“You were lucky,” he said. He did not say it bitterly. He said it the way a man reads out a figure.

She could not tell whether he was right. She had worked hard, and she had made some good choices. She had also been kept on, paid every month while she learned, and given time that other people were not given. Which part of where she sat was hers, and which part had been lent to her?

By the account the print edition of this book gave, Priya is its proof. It said that the disruption “that had nearly destroyed her was the thing that had forced the transformation”, and that what she had learned to do “is not a personality trait. It is a design principle. And like any design principle, it can be learned.” It built a whole chapter on that hope. This chapter asks: **When machines unsettle a person’s work, can she come out stronger, and if so, what is that strength made of, and who pays for it?**

The print edition’s answer was large. “Anti-fragility is not only a property of well-designed systems,” it said. “It is an inherent property of consciousness itself.” This edition withdraws it, for four reasons that the chapter will show. First, in Priya’s building the thing that gains most from each failure is a machine. Second, the evidence that suffering makes people grow is mostly what people say about themselves afterwards, and its meaning is disputed. Third, the same chapter’s last section said that the conscious self was “never breakable”, and a thing that cannot break does not gain from breaking. Fourth, suppose growth were a property of every conscious person. Then a man who did not grow would have failed a test, and the people who decided his loss would have set it.

What this edition holds is smaller. Some people come out of a shock stronger, under conditions. The conditions are mostly slack: money, time, a household that can carry a bad year, an employer that keeps paying, a public floor to stand on. They include choices kept open and shocks small enough to survive. Much of this is lent, and some of it is paid for by others. Redundancy is built in calm times, by those who can afford to build it. A person may make meaning of a harm, and nobody may demand that she does.

The title needs an argument too. In the sense of the man who made the word, “anti-fragility in an AI world” describes the machines, the companies that run them and the industry better than any worker inside it. Whether a person can have it is this chapter’s question. The answer is partial.

## 30.1 What Grows Under Pressure

Nassim Nicholas Taleb was once an options trader. An option is a contract that gives the right, without the duty, to buy or sell something later at a set price. He became a writer on risk and chance, and in 2012 he published *Antifragile*, a book about a word he said the language lacked. “Some things benefit from shocks; they thrive and grow when exposed to volatility, randomness, disorder, and stressors,” he wrote. There was no word for the opposite of fragile, so he made one: “Let us call it antifragile.”¹ He set it beside two older ideas. “Antifragility is beyond resilience or robustness. The resilient resists shocks and stays the same; the antifragile gets better.” A glass breaks when it is dropped. A stone resists. A muscle that is worked, and then rested, grows.

The muscle shows the conditions at once. It grows from a dose it can survive, with time to recover. Too much load, or no rest, and it tears. Not all suffering teaches. Taleb knew this. He likes to write in cafés, against a little background noise, and he compares such noise to the small doses of harm that can strengthen a body. Then he adds that “these small distractions, like hormetic responses, act up to a point”.² Antifragility is not a property a thing has whatever happens to it. It is a gain from certain shocks, in certain amounts, with time between them. The range differs from one thing to another, and from one person to another. I take the muscle as a picture, not a measurement. A person with savings, health and people around her can usually take a larger blow and still recover. A person with none of those has a narrower range, and the same blow can carry her past it.

Several words are easy to blur here. To be resilient is to take a shock and go on working. To recover is to return to where one was. To adapt is to change to fit new conditions, which can leave a person worse off. To grow is to gain a capacity. Antifragility, in Taleb’s sense, is a gain that comes from the shock itself. Each word also needs a subject. The same shock can make a system antifragile, an organisation merely resilient and a person poorer, so the first question about any gain from disorder is whose gain it is.

The fourth chapter of Taleb’s book has a title that answers an older sentence. In 1888 Friedrich Nietzsche wrote, in Anthony Ludovici’s translation, under the heading “From the military school of life”: “That which does not kill me, makes me stronger.”³ Taleb called his chapter “What Kills Me Makes Others Stronger”. His example is a street of restaurants. Each restaurant is fragile: many fail. The restaurants of a city, taken together, get better, because the weak ones close. The gain belongs to the trade, and the cost falls on the cook whose place shut. The trade’s gain is real, because diners eat better. It does not cancel the cook’s loss, and it does not say who should carry it. Taleb says plainly that a system can gain from disorder at the expense of its parts. He does not leave that as a description. In his Prologue, in a passage on gains taken at the expense of others, he had already written: “At no point in history have so many non-risk-takers, that is, those with no personal exposure, exerted so much control.” Then he set down what he called “the chief ethical rule”: “Thou shalt not have antifragility at the expense of the fragility of others.”¹ The last part of his book returns to it. Nietzsche speaks for the survivor. Taleb counts the restaurants that closed.

India has its own survivor’s verse. It is in the *Nītiśataka*, the “hundred verses on conduct” that tradition gives to the poet Bhartṛhari. C. H. Tawney put it into English verse in 1877:⁴

> “Cowards shrink from toil and peril, Vulgar souls attempt and fail; Men of mettle, nothing daunted, Persevere till they prevail.”

The Sanskrit says a little less and a little more.⁵ Its second half, *vighnaiḥ punaḥ punar api pratihanyamānāḥ prārabdham uttamaguṇā na parityajanti*, says only that those of the highest quality, struck back by obstacles (*vighna*) again and again, do not give up what they have begun. A prose translation of 1886 keeps that limit.⁶ “Till they prevail” is the translator’s promise, not the poet’s. Sri Aurobindo’s version, which he called “freely rendered into English verse”, adds a picture of its own: “Buffeted and borne down, rise up again / And, full of wounds, come on like iron men.”⁷ In Aurobindo’s picture, the best of Bhartṛhari’s three kinds of people come out wounded and still going. In Tawney’s, they win. In the Sanskrit they neither win nor bleed: they do not give up. In Taleb’s three kinds, that is staying power, not gain.

Bhartṛhari wrote about a person’s own undertakings, not about an industry, and the application that follows is mine. Read beside Nietzsche, the verse can sound like the maxim, but it praises something smaller. It praises not giving up. It does not say that the obstacles make anyone better. Taleb’s chapter moves the question. In Priya’s building, what gains most from every failure is not a person.

It is the platform her team built. Each month it is retrained on that month’s tests, including the defects that customers found after release. A failure that escapes in March becomes something the April model has seen. It does not learn by itself. People find each defect, record it and decide to retrain on it, so its gain is a loop of human work. It also learns only what is recorded. A defect that a customer met and never reported, or a customer who simply left, teaches it nothing. It learns from every product team at once, and no tester’s career pays for its lessons. A tester learns from her own mistakes, one at a time, and pays for each. Someone does pay for the platform’s lessons, though. Every defect it learns from in this way was first found by a customer. It might be a bill charged twice, a form that would not submit, or a page that failed on the day it was needed. The system gains from those failures. The customers who met them gained nothing, and most of them never knew that their bad afternoon had become a lesson for a machine. That is Taleb’s restaurant street again, at the scale of one company. The gain gathers in the whole, and the cost is spread thinly across people outside it. As I read Taleb’s rule, it does not forbid learning from a failure that has already happened. It forbids an arrangement that makes others carry the failures you gain from. A company that tested less because its machine learns from what customers find would break it. Whether Priya’s company has done so, its own count of the defects customers find can show. If antifragility were a mark of consciousness, the platform would have the better claim.

The print edition also said that machines catch “the easy bugs” and miss “the bugs that had no pattern”, which were left for a person’s judgement. That was already out of date. In November 2024 Google’s open-source security team reported that test programs written or improved by AI had found 26 new flaws in widely used open-source software. One, a low-severity flaw, was in OpenSSL, a library that encrypts much of the internet’s traffic.⁸ It had “likely been present for two decades”, and it “wouldn’t have been discoverable with existing fuzz targets written by humans”.⁹ A fuzz target is one of these test programs. A fuzzing tool runs it on a stream of generated inputs, looking for one that breaks the software. What a person checks moves as the machines move.

So what grew in Priya? Something did. Between October 2024 and early 2026 she changed what she did, how she thought about it and what she was trusted with. Some of that was her own: she asked questions nobody had asked her to ask, said yes when the post was offered, and did the work. Much of the rest was given. She had a salary that did not stop while she learned. She had years of standing when the change came, so people listened to her. She lived in a household of two that could carry a hard year. Her manager, her director and the company’s chief product officer made decisions that she did not make. None of that makes her growth false. It makes it partly lent.

Naveen’s word was not wrong, either. Some of what carried her she could not have arranged. She did not choose the company she happened to join, the director she happened to have or the year the machine arrived. Nor did she choose that her team would be asked to build the platform, rather than be replaced by someone else’s. Luck is the name for what helped a person and was not chosen by her. To say so takes nothing from what she did with it. It only stops her story from being used as proof that anyone who tries as hard will land where she landed. Some of it was not luck but an employer’s choice. Her company put her inside its plan for the change and paid her while she learned. Naveen’s firm left him outside its plan. Those were decisions, and people made them.

The same change that carried her moved others out. In the year to March 2026, the five largest Indian IT services firms together lost 7,389 people.¹⁰ In the year before that they had added 12,718. The report set the fall against changing demand and a turn to AI, and the count does not say how many places the machines closed. Naveen was one of the parts. His twelve years were real, and his skill was real, and the trade went on remaking itself around the machines while his place in it closed. The platform learns from every team’s failures at once; a person learns from her own, one at a time, and pays for each. That is why “the company became antifragile” can be true in the same year that many who left it are not. Nor did the cousins meet the same shock. Priya was given a heavier load: a new post, a platform to build, people to persuade. Naveen was cut off from his work. A muscle grows under a load it can bear, and it wastes when it is not used at all. A lay-off is not a heavier weight. It takes the weight away. A person’s success after a shock is not evidence that the shock was good for people. A winning bet is not evidence that the bet was wise.

## 30.2 The Redundancy You Build Before You Need It

Taleb’s word for what saves a system is redundancy. In a passage his publisher allowed a newspaper to reprint, he wrote: “Layers of redundancy are the central risk-management property of natural systems.”² People have two kidneys and can live with one. “Redundancy is ambiguous because it seems like a waste if nothing unusual happens,” he wrote. Then something unusual happens. He called redundancy “more like investment than insurance”, because spare strength can be used even when no disaster comes.

Toyota is famous for making cars with little spare stock, with parts arriving just in time. After the earthquake and tsunami of March 2011 cut its supply chains, it made one exception. Its suppliers were asked to hold “two to six months’ worth of chips”, the semiconductors that modern cars cannot run without. When chips ran short in early 2021, Toyota kept building cars while Volkswagen, General Motors and Ford cut production. The stock was one reason; another, Reuters reported, was Toyota’s own knowledge of how chips are made.¹¹ Its redundancy had a limit. In August 2021, with the shortage worse and Covid restrictions hitting suppliers in South-East Asia, Toyota cut its planned output for September by 40 per cent, “becoming the last major automaker to cut output”.¹² Redundancy bought it months, not immunity, and holding it cost money in every year when nothing went wrong.

For machines, spare capacity is cheap. A program can run in two places at once, and a copy costs almost nothing. Human time and attention are finite. Nobody can keep deep competence in five fields at once. A person’s redundancy is paid for in hours.

Here is the advice in a plain form. Spread what you can do, whom you know, what you earn and what you know, so that no single loss can break you. Each part has a price, and each can be tested on the two people in the café.

What you can do. Priya can test software, design how testing is done and explain it to people who do not want to hear it. Naveen can test software by hand, well, which is the part the machines took first. A second skill costs evenings, and evenings belong to whoever is not already spending them.

Whom you know. Nearly everyone Priya knows at work is inside one company. Naveen’s network was mostly the colleagues who left the firm with him, who were looking for the same jobs. People outside one’s employer who would take a call are not collected in a hurry. They come from years of small help given without being asked: an answer to a question, an introduction, a reference written honestly. That takes time in good years, which is the time a busy employee least has.

What you earn. The print edition asked every reader for “at least one income source independent of your primary employer”. Many employment contracts forbid it, or require written permission. In September 2022 Wipro, one of India’s largest IT firms, said it had dismissed 300 employees found working for competitors; its chairman said the company considered such moonlighting an “act of integrity violation”.¹³ A second income for a salaried person is often a question for the first employer.

What you know. Much of what Priya knows lives in her company’s systems: its code, its records, the history of its tests. If she left, a good part of her knowledge would stay behind. Naveen’s knowledge of the firm’s clients stayed with the firm the day he left.

Behind all four lies one word. **Slack** is what is left over when this month has been paid for: money, time, attention, and people who would help. Redundancy is built out of slack. Priya had some, much of it lent: the salary, the household, the standing. Naveen had a home loan, two children at school and his wife’s salary as a teacher, which covered most of each month. The advice to build redundancy before you need it is good advice. It is easiest to follow for the people who need it least, and its cost falls hardest on those with nothing to spare. Whether a society owes its workers a floor to fall on is a question for the law and for politics. It is not a separate question from this chapter’s, because a floor is slack that nobody can build alone. India’s state insurance scheme pays an insured employee who loses a job an allowance for up to 90 days, once in a lifetime. In 2026 it covered only employees paid up to ₹21,000 a month, below a newcomer’s usual pay at the largest IT firms. It would very probably not have covered a tester with twelve years’ experience.¹³

The word “before” carries most of the weight. Redundancy is cheapest in a good year, when it looks like waste, and dearest after the shock, when it is needed. Naveen could have learned a second trade in 2024 on a salary. In March 2026 he would have had to learn it on two months’ pay, while looking for work, with the loan still due. A shock takes away the very slack that redundancy is made of. That is why the advice says “before”, and why it is hardest for those whose good years were never good enough to spare anything. The machines add a twist. When a week of tests runs in eleven minutes, the days saved seldom come back to the tester as time to learn. More often they become more work, or fewer people. The technology that makes a second skill necessary is also used to take out of the working day the slack in which it could be learned.

### What would prove this wrong?

My claim here is about timing. Slack held before a shock can become a second skill or a wider circle of people. Slack handed over after it mostly buys time. If money given to people who had just lost work regularly led them to better jobs, the word “before” would carry less weight than I give it.

The evidence is **Contested.** In Austria, a severance payment worth two months’ earnings, what Naveen’s firm gave him, made job losers take longer to find work. David Card, Raj Chetty and Andrea Weber found that the longer search had “little or no effect” on the quality of the next job. Later, in the same country, Arash Nekoei and Andrea Weber found that longer unemployment benefits led people to better firms and higher wages. In Germany, Johannes Schmieder and two colleagues found that longer benefits lowered later wages, mainly because time out of work lowered the pay people were offered¹³.

These studies measured jobs and wages, not growth, and none was about machines. To test me, a reader would look for studies that follow two groups of similar workers through a lay-off. One group had savings or a second skill beforehand, and the other had neither. The test is where each group stands some years later.

An employer can lend slack or withhold it. It can give hours for learning, permission for outside work and notice before a cut. Notice is slack measured in time. Naveen’s firm gave him two months’ pay. That was more than nothing, and less than the time it takes to build a second skill. The papers he signed called his exit a resignation. Had they called it a retrenchment, Indian law would have owed a worker with twelve years’ service a month’s notice. It would also have owed fifteen days’ pay for each year of service, about six months’ pay in all. A labour court can look behind the word to what happened, but only if someone takes the case there. Since November 2025 the law has also asked the employer to pay fifteen days’ wages into a fund for the worker’s retraining. When Naveen left, the rules for that fund were still in draft.¹³ In 2025 groups of IT employees complained that one of India’s largest IT firms was pressing people to resign, without paying retrenchment compensation. The firm denied it. A worker may refuse to sign, but refusing, and the dispute that may follow, also takes slack.¹³ A floor written in law holds only those whose exit is written as the law describes. An employer that asks its people to be resilient, and gives them no notice, time or money to be so, is asking them to carry its decision.

Redundancy protects. It can also sedate. A company whose machine checks everything may stop practising the check. Its spare people are let go, and the knowledge of how to do the work by hand goes with them. A fallback that nobody has practised is not a fallback. The tools change with the habit. Screens built to show a reader only what the machine has flagged make a whole record harder to read by hand, as a plant near Lyon finds in Chapter 36. Aviation learned this early. In 2013 America’s aviation regulator warned airlines that constant use of the autopilot and similar systems could weaken pilots’ ability to recover an aircraft quickly by hand. It encouraged them to let pilots fly by hand when appropriate.¹³ Moving knowledge into a machine can also move a dependence rather than end it. A company that needs fewer of its own experts may come to need one provider’s model, cloud and terms. In July 2024 one faulty update from a single security company crashed about 8.5 million Windows computers, by Microsoft’s estimate, and flights were grounded around the world.¹³

## 30.3 Distributing Intelligence

There is a redundancy of knowing as well as of stock. A system whose knowledge sits in one head fails when that head is missing. Spread the knowing and the deciding across many people, and no single loss can stop the work. Meera Reddy tried this, and in January 2026 it failed.

Meera is vice-president of operations at the Hyderabad site of NovaChem Pharmaceuticals, in Genome Valley. NovaChem is a drug company with its head office in Basel. In January 2026 she had been at the company for nearly fifteen years. Twelve managers in quality, supply chain, manufacturing and procurement reported to her. From November 2025 she had changed how they worked. She called the change “from controller to cultivator”. The weekly status meetings went, and each manager was judged by how complex the decisions were that their people could make on their own. Decisions moved down, to the people nearest the work.

In the middle of January a supplier failed to deliver the aluminium foil that seals tablets into their blister packs. The tablets were promised to a customer in Germany, a distributor, on a date already agreed. The plant follows the European guidelines on good manufacturing practice. They require that suppliers of packaging materials be selected, qualified and approved with care similar to that given to the ingredients themselves. They also say that incoming materials should be “quarantined immediately after receipt”, “until they have been released for use or distribution”.¹⁴ Nothing packed in new foil could leave the plant before quality had released the foil. NovaChem had two approved suppliers of that foil. That second supplier was itself redundancy, and real redundancy, because it had been qualified before it was needed. A supplier on a list that nobody has checked is redundancy only on paper.

The site’s procurement agent, the program that compares suppliers, had the second supplier ranked first within the hour, with an expected delivery date. It knew prices and lead times. It did not know that new foil would wait in quarantine. One of Meera’s three supply chain managers moved the order to the second supplier, so that the date could be kept. A quality manager held the batch until the second supplier’s foil had passed its checks on arrival, as the plant’s procedure and the rules required. Each decision was lawful. Each was within the manager’s authority. And each manager wrote to the customer. In two days the distributor’s purchasing office received two messages about the same shipment: one said it would arrive on time, the other that it was held. The first was a promise that depended on a release nobody had yet given. A week later a long email reached Meera, formal and furious. The distributor had told its own customers, pharmacies and hospital buyers, the first date. Now it did not know which date to believe, and it said so.

Both decisions were sound within their local frames. Decentralisation does not eliminate the need for a centre. It changes what the centre does. What the print edition got wrong was the centre. It called the centre “a consciousness function, not an information function”, and said it needed “the quality of attention”. The centre Meera lacked was simpler. Since August 2025 every handoff in her teams’ redesigned workflows had one named person. Nobody had been named for the case where two well-held handoffs pull against each other. Here, one team gave a date that depended on a release the other had not yet given. The weekly meeting where such conflicts used to surface was gone. Meera had removed it without asking what else it did. The machine had made one team’s decision faster, and so the gap between the two teams opened faster too.

Computer engineers met this problem long before Meera, and they named it. In a cluster of machines that keep the same records, two machines that lose contact may each decide that it is in charge. Both go on writing, and the records split. The engineers call it *split brain*.¹⁵ Their remedy is a rule written before any failure. In one widely used design, a machine may lead only if a majority of the cluster has voted for it, and “The majority rule ensures that at most one candidate can win the election for a particular term”.¹⁶ The side that has lost the majority can no longer make any change final, so its work waits. The rule costs something: part of the system cannot finish its work, so that the whole does not contradict itself. Here the machines’ designers did better than Meera’s organisation did. They wrote the rule before the failure. She wrote hers after. The picture has a limit. An organisation is not a cluster of machines. A vote can settle which machine leads. It cannot settle whether a decision is right, or who answers when it harms someone.

She read the email twice and then went through the record of the two decisions. She found nothing either manager should have done differently inside the rules each was given. The fault was in what the rules left out, and she had written them. She answered for her part. She wrote to the distributor herself: what had happened, the new date, and who would now be the only person to write to them about dates. NovaChem paid for faster freight to recover part of the delay. Then she made a rule. When a date already promised to a customer depends on a quality release, nobody gives the customer a new date until quality has given one. One named person, with a named deputy for when she is away, tells the customer, within a day of any hold, what is known and what is not. A half-hour meeting returned each week, for conflicts between teams only. The price was slower days and a little less freedom for the managers she had just set free. The manager who had moved the order asked her, at the first of the new meetings, whether the decisions were theirs only until they mattered. She said the decisions were still theirs. What waited now was the promise to a customer, and she would rather have that written down than learn it again from a customer. He was not satisfied, and he was right to ask. The rest of the cost was not hers alone. The site’s quarterly report to the head office in Basel counted deliveries made on time. A measure that counts dates kept rewards the promise, not the truth about it. The office that set the measure carried none of the cost of meeting it; the people nearest the work carried that. Taleb’s rule applies to a measure as much as to a trade. The quality procedure required the hold, and the rules put it beyond argument. Nobody above her had said what a customer should be told while a hold lasted. At the end of that January another of her managers told her something harder. Meera had asked them to stop controlling their teams, the manager said, without stopping controlling them. Meera knew it was true of the new rule as well.

Distributing knowledge has a second face, and it was in Priya’s building that March. Arvind, an infrastructure engineer at her company since 2018, had worked nearly alone for years on the oldest and most fragile code. The company’s knowledge of that code lived mostly in his head. As the platform spread, Priya’s team gave a model the old code and its history and asked it to explain what each part did. It explained most of it correctly. Arvind was asked to check its explanations, in his own hours. The company became less fragile that month: if Arvind left, the knowledge would not leave with him. Arvind became more exposed, and Priya’s own work had made him so. One whole’s redundancy can be one person’s fragility. Whether the company would later count that against him, nobody had yet had to decide. What it owed him was not hard to name. It could tell him why the work was being done, and count his hours of checking as work rather than as a favour. It could also say in writing whether the model’s explanations would ever be used to judge him. In 2025 the company had written into its appraisal guidance that the platform’s predictions about code would not be used to judge individuals. Whether that covered a model’s account of a man’s own code, nobody had yet had to test.

The model had captured what the code does. It had not captured why it was written that way: which changes had broken things before, which parts nobody should touch in a hurry, what an old failure had taught. Information can be taken from a person without the judgement that made it useful. Taking knowledge out of a head, keeping it, and making someone else able to use it well are three different things, and the company had done only the first two. The rest was a matter of power. Nobody had asked Arvind before the work began, or said whether he could correct the model’s account of his own code. Once it was written down, that account could become the official one, even where it was wrong. Arvind paid for the company’s new redundancy in his evenings, and the company held it. What his contract allowed and what a fair arrangement owed him are different questions, as Chapter 20 argued. A company that learns from its people takes on duties to them, not only advantages over them.

The same is true at the scale of one life. Whatever a person knows that lives only in her own head is a risk to the people who depend on her. Whatever she knows that lives only in her employer’s systems is a risk to her.

## 30.4 The Practice of Growing from Disorder

“Growing”, in this title, has to mean what the muscle means: a gain from a dose, with time to recover, paid for by someone. With that meaning, four of Taleb’s ideas for living with disorder can be tested. Together they say one thing: protect what must not be lost, keep open the choices that may matter, remove false safeguards, and make those who decide carry what follows.

The first is the barbell, the bar a weightlifter raises, with its weight at both ends and nothing in the middle. Put most of what you have somewhere very safe, Taleb says. Put a small part in bets that could pay a great deal and can lose only what you put in. Avoid the middle, which feels safe and is not.¹ In his own example the safe end is money. It is tempting to say that for a worker the safe end is a disposition, such as adaptability or the habit of learning. Dispositions help a person meet change, and they promise nothing. They are not a safe end. For a person, the safe end of a barbell is a wage, a household or a public floor. Priya has a barbell of her own. Her post is the safe end. The small bet is a study of volunteers she asked to run for three years, which the chief product officer approved in April for one. Her employer funds both ends. Naveen’s barbell had only a thin safe end: his wife’s salary, which covered most of each month and not all of it. His savings were the bet, and they paid for the rest. Neither safe end is safe in a bond trader’s sense. Priya’s post depends on her company’s choices, and a teacher’s salary on a school’s. Two incomes that come from the same shaken trade can also fall together.

The second is the *via negativa*, a Latin phrase for the way of saying what a thing is not. Taleb uses it for improving by taking away: what to stop, what to avoid. It is a scalpel. Used well, it removes what is making you fragile. Used carelessly, it cuts what is keeping you alive. The weekly meeting Meera cut was such a case. It looked like waste, and it was where conflicts surfaced. A capacity is not waste because it sits idle. Before cutting it, one should know which failure it makes survivable.

The third is optionality: keeping choices open, so that when the world changes one of them becomes valuable. Options cost something to keep. An option held by a company is usually a job held by a person, someone kept on a project in case it is needed. And an option is not yet an act: the cost of keeping choices open falls on those who wait while the choosing is put off. A short contract shows both sides. For a firm, a six-month contract is an option, easy to end if the tools arrive. For the tester it is an act. He gives up the search for something better, and stakes six months on a team that may not need him by the end of them. For him the contract is Taleb’s barbell turned upside down. His gain is capped at six months of lower pay. His loss is not capped: when the tools arrive he is looking for work again, older and with less saved. The law does not count the end of such a contract as a retrenchment, so he is owed no retrenchment compensation.¹³ The firm holds the barbell the right way round, with a small commitment and the freedom to end it.

The fourth is what Taleb calls skin in the game: bear the downside of your own decisions. His ethical rule follows from it. Do not take the gains of disorder while others carry its losses. This book asks for more than exposure to loss. To answer for a decision is to explain it, to put right what can be put right, and to change the practice that caused the harm. A trader who loses his own money has skin in the game. He has not answered to the people his trade harmed. Naveen never learned who at his firm had decided to let the testers go. Answering, for that person, would have begun with a reason given in their own name. It would have gone on to what notice and money could put right. Whoever it was, the downside of a wrong decision would fall on others first. The clients’ software would be checked less carefully if the tools were weaker than promised. The testers who stayed would take on the work of those who left. The people who decide a shock and the people who carry it are often different people. Taleb’s rule is written for exactly that gap.

Bhartṛhari’s verse returns here with a question it did not ask, and the reading that follows is mine. Its “cowards”, in Tawney’s word, are those who do not begin, for fear of obstacles. Some who do not begin are cowards. Others have no floor to fall on if the beginning fails. A man with a home loan and two children at school who will not try a risky new trade may be judging his slack correctly. The verse ranks people by how they meet obstacles to a work they chose. It says nothing of a work someone else ended. Going on is not always wise, either. To leave a trade that is closing, or to refuse training that leads nowhere, can be judgement rather than fear. Taleb’s *via negativa* has room for that; the verse does not.

Time binds the four. The professional who restructures their entire career by Friday has not become anti-fragile. They have added fragility. Growth happens at its own pace, which is never the pace that anxiety demands. The print edition was wrong to add that patience “is not a property of systems. It is a property of consciousness.” Patience in a working life is mostly paid for. The person who can wait is the one whose month is covered. Priya’s salary pays for her waiting. Naveen, in March, had no salary to wait on. Every week he spent looking for the right job was a week of the two months’ pay.

## 30.5 What Breaks, and What People Make of It

The print edition said that when a machine meets a severe shock, “it breaks, recovers to baseline, or adjusts its parameters. What it does not do is grow.” As of 2026 that is not true in the sense that matters for this chapter. Machines are retrained from their failures, and the companies that run them can learn from many users at once. Whether anything is undergone inside a machine while it is retrained is an open question, and nothing in this chapter depends on the answer. What the platform is not is the one on whom its failures land. A wrong release does not fall on its month, its family or its name. It falls on a customer, a developer, a company and, through them, on people. That is no defect of the machine. It is a fact about whose lives are at stake, and it holds however well the machine learns.

What people do after a hard blow is better studied, and the studies are more careful than the slogan. The psychologists Richard Tedeschi and Lawrence Calhoun named the idea. “Posttraumatic growth is the experience of positive change that occurs as a result of the struggle with highly challenging life crises,” they wrote in 2004.¹⁷ They use “trauma” broadly, for crises of many kinds, and call their usage “a bit broader and less restrictive” than that of some literatures, citing the American Psychiatric Association’s diagnostic manual. In the manual’s current edition, trauma means exposure to actual or threatened death, serious injury or sexual violence.¹⁸ Losing a job can be a crisis in Tedeschi and Calhoun’s sense. It is not trauma in the manual’s. They also warned against expecting growth: “The widespread assumptions that traumas often result in disorder should not be replaced with expectations that growth is an inevitable result.” And they wrote: “continuing personal distress and growth often coexist.”

The print edition called the phenomenon “real and well-replicated”, and two paragraphs later admitted that it is measured mostly by what people report. The second is nearer the truth. In 2006 Vicki Helgeson, Kerry Reynolds and Patricia Tomich published a meta-analysis of 87 studies. It found that people who reported benefits from a crisis were less depressed and reported more well-being, “but also more intrusive and avoidant thoughts about the stressor”. Reported growth was unrelated to anxiety and to distress overall. All 87 studies were snapshots, taken at one time.¹⁹ In 2009 Patricia Frazier and colleagues measured students before and after hard events and asked them afterwards how much they had grown. Their reports of growth, the authors found, “generally were unrelated to actual growth” in the same areas.²⁰ That does not make all reported growth false. It means that what people say about their growth is a poor guide to what has changed. In 2010 Mark Seery, Alison Holman and Roxane Cohen Silver followed a large national sample for several years. People with “some but nonzero lifetime adversity” did better than people with none and people with a great deal. The authors’ conclusion carried its own limit: “in moderation, whatever does not kill us may indeed make us stronger”.²¹ The adversity they counted was what had happened to people, such as illness, bereavement, violence and disaster. Their data, they wrote, “cannot establish causality”, and their results “should not be construed to endorse intentional trauma exposure”. Nietzsche’s sentence survives, then, with a dose attached, and as a pattern in survey answers, not a law of the self.

There is a test that keeps all this honest. A job bundles four goods: income, a practice, standing among others, and a voice in how the work is done. Naveen lost all four in February. His income fell to two months’ pay and his wife’s salary. His practice, the careful reading of software by hand, was the part the machines had taken first. His standing, twelve years of being the one his colleagues asked, stayed with the firm. And he had no say in the decision that ended his job, and no one there to whom he could say so. Growth after harm, even where it is real, restores none of the first three. At most it changes how a person carries what he lost. That matters, and it is his. It is not a wage. A person can gain something from a loss and still be worse off overall.

The print edition’s last claim was the largest. “You are not the thing that breaks. You are the thing that watches the breaking and finds, in the watching, that it was never breakable.” That is not antifragility. It is a promise of a citadel inside, where no harm reaches, and it is offered to people in distress. Some find such a place, and many traditions have honoured it. The Roman emperor Marcus Aurelius called a mind free of passions a fortress.²² A fortress can also become a place to hide from what should be fought. But a book that tells the harmed that their real self was never touched has told them that the harm does not count. A harm someone decided is someone’s to answer for. The one who imposed it may not call it a gift, and nobody may require the one who suffered it to grow.

What a person may honestly make of a harm is her own. She may find that she values different things, or knows people better, or is less afraid of the next blow. She may find none of that, and be no less a person. Finding a benefit, making a meaning and growing are different things. A person can make meaning of a loss without becoming stronger, and can grow while still resenting what caused it. Bhartṛhari’s best are struck back again and again and do not give up. Tawney’s English promised that they would “prevail”; Aurobindo’s free rendering sends them on “full of wounds”. The Sanskrit promises neither, and it does not pretend that the blows are good for them. That is a truer picture of going on than a self that cannot break. It also leaves room for the one who cannot go on yet, and owes nobody an account of why.

## Ending: Two-Thirds of His Old Pay

In late April Naveen asked to meet again, at the same café. He had an offer. A smaller firm wanted a manual tester on a six-month contract, at about two-thirds of what he used to earn. He had until the end of the month to say yes.

Priya began to say that it might turn out to be a chance.

“Do not tell me it is a chance,” he said. “Everyone at home has told me. I lost twelve years of work. That is a real loss, and it is unfair, and no idea in any book makes it less unfair.”

She stopped. Then she told him what she knew, and said that it was only what she had seen. Teams like the one hiring him usually received tools like hers within a year or two. She did not tell him what to do. It was his month to cover, and his family’s, and his choice.

He thanked her and paid for the tea. When she went back across the road, he had not yet decided.

### Practice

*Before you need it.* Write four lines about what would carry you for six months if your main work ended. First, what you can do that someone outside your employer would pay for. Second, who outside your employer would take your call. Third, what money, or which other earner, would hold the household. Fourth, what you know that lives only in your employer’s systems or only in your own head. Beside each line, write who pays for it now: you, your household, your employer or the public. Then write one shock you impose on someone else, as a manager, a client, a contractor or a parent, and what would carry them. Choose one line to strengthen this year, and write its cost in hours and whose hours they are. Then write the other side. Who can make a decision that changes your livelihood? What notice do they owe you, and what happens to what you have taught their systems? Write one thing within your own power, and one condition that only an employer, a union or the state can change. If your contract limits outside work, read the clause before you begin. This takes an evening, which not everyone has. It will not keep your job, replace a wage or promise calm. It shows how much of what holds you is yours, and how much is lent.

## Words to keep

- **antifragile:** Taleb’s word (2012) for things that gain from shocks, within a dose and with time to recover; the fragile breaks, and the resilient resists and stays the same.
- **barbell:** a bond trader’s strategy that Taleb applies to life: most of what one has somewhere very safe, and a small part in bets that can lose only what was put in.
- **optionality:** keeping choices open so that one may become valuable when the world changes; the cost of keeping them is usually paid by someone.
- **post-traumatic growth:** Tedeschi and Calhoun’s name for positive change reported after a struggle with a hard life crisis; mostly measured by what people say about themselves.
- **redundancy:** spare capacity held before it is needed, such as a second supplier, a second skill or savings.
- **retrenchment:** in Indian labour law, an employer’s ending of a worker’s job for any reason other than discipline, with some exceptions, such as retirement or the end of a fixed-term contract; the law attaches notice and compensation to it, which a “resignation” does not carry.
- **skin in the game:** a phrase Taleb uses for bearing the downside of one’s own decisions.
- **slack:** what is left over when this month has been paid for: money, time, attention, and people who would help.
- **split brain:** in computer clusters, two machines that each believe they are in charge and write conflicting records.
- <strong><em>via negativa</em></strong><strong>:</strong> Latin, the way of saying what a thing is not; in Taleb, improving by taking away.
- <strong><em>vighna</em></strong><strong>:</strong> Sanskrit, an obstacle or hindrance; the word Bhartṛhari’s verse repeats.

## Sources

1. Nassim Nicholas Taleb, *Antifragile: Things That Gain from Disorder* (New York: Random House, 2012): Prologue (the definition and the coinage; the ethical rule, in the section “Upside at the Expense of Others”: “The chief ethical rule is the following: Thou shalt not have antifragility at the expense of the fragility of others.”); Book I, chapter 4, “What Kills Me Makes Others Stronger” (the restaurants; a system’s antifragility at the expense of its parts’ fragility); Book III, chapter 11, “Never Marry the Rock Star” (the barbell); Book VI, “Via Negativa”; Book VII, “The Ethics of Fragility and Antifragility” (ethics as transfers of fragility). Chapter titles from the Random House edition’s table of contents; the Prologue’s wording as in the Random House ebook (ISBN 9780679645276).
2. Nassim Nicholas Taleb, from *Antifragile* (chapter 2, “Overcompensation and Overreaction Everywhere”), excerpt “reprinted by arrangement with Random House” as “Why trauma may be just what you need,” *The Globe and Mail* (online, undated book excerpt).
3. Friedrich Nietzsche, *The Twilight of the Idols* (1888; published 1889), “Maxims and Missiles” 8, trans. Anthony M. Ludovici (Edinburgh and London: T. N. Foulis, 1911), Project Gutenberg eBook 52263. German: “Aus der Kriegsschule des Lebens. Was mich nicht umbringt, macht mich stärker.” (the printed dash after “Lebens.” omitted, here and in Ludovici).
4. Bhartṛhari, *Two Centuries of Bhartrihari*, trans. C. H. Tawney (Calcutta: Thacker, Spink, 1877), p. 11 (reprinted from the *Indian Antiquary*), Internet Archive; punctuation of line 3 normalised (the OCR shows a full stop after “daunted”).
5. Bhartṛhari, *Nītiśataka*, the verse *prārabhyate na khalu vighnabhayena nīcaiḥ*, as printed in “Neeti Shatakam of Bharthruhari” (stotram.lalitaalaalitah.com, 15 November 2011), verse 1.26, whose last quarter reads *uttamaguṇāḥ* (“those of the highest quality”), as does the form of the verse in the *Pañcatantra* (III.244, with a different third quarter). Many other texts, among them a CBSE school text, read *uttamajanāḥ* (“the best people”). Wortham numbers the verse 27. The plain sense given in the text is mine.
6. *The Śatakas of Bhartṛihari*, trans. B. Hale Wortham (London: Trübner, 1886), Nīti Śataka, stanza 27, p. 5: “when he has once begun the pursuit of wisdom he does not give it up”.
7. Sri Aurobindo, “The Century of Life: The Nitishataka of Bhartrihari, freely rendered into English verse”, “Great and Meaner Spirits”, in *Translations*, The Complete Works of Sri Aurobindo, vol. 5 (Pondicherry: Sri Aurobindo Ashram Publication Department, 1999), p. 325.
8. OpenSSL Security Advisory, 16 October 2024, CVE-2024-9143: “Severity: Low”; reported on 16 September 2024 by Google OSS-Fuzz-Gen.
9. Oliver Chang, Dongge Liu and Jonathan Metzman, “Leveling Up Fuzzing: Finding more vulnerabilities with AI,” Google Security Blog, 20 November 2024 (26 new vulnerabilities; CVE-2024-9143 in OpenSSL). Google’s fuzzing documentation, “Glossary” (google/fuzzing, GitHub): a fuzz target is “A function to which we apply fuzzing”; a fuzzing engine is “A tool that tries to find interesting inputs for a fuzz target by executing it”.
10. IANS, “Top IT firms’ hiring turns negative as headcount falls over 7,000 in FY26,” 25 April 2026: TCS, Infosys, HCLTech, Wipro and Tech Mahindra together down 7,389 in FY26 (April 2025 to March 2026) after a net addition of 12,718 in FY25, “amid an increasing focus on Artificial Intelligence (AI) and changing demand in the industry”, with layoffs of “around 12,000 employees” announced by TCS; see Chapter 4.
11. Norihiko Shirouzu, “How Toyota thrives when the chips are down,” Reuters, 9 March 2021 (as republished by Nippon.com): the business continuity plan after 11 March 2011; suppliers to stockpile “two to six months’ worth of chips”; rivals including Volkswagen, General Motors and Ford cutting production; sources said Toyota’s understanding of semiconductors “was a major reason why it has managed to avoid being hit by the shortages, in addition to its continuity contracts”.
12. Reuters, “Toyota slashes September output amid chip crunch,” 19 August 2021 (as republished by the *Dhaka Tribune*): global production plan for September 2021 cut by 40 per cent, about 360,000 vehicles; “becoming the last major automaker to cut output”.
13. Manish Singh, “IT services group Wipro fires 300 employees moonlighting for competitors,” TechCrunch, 21 September 2022 (Rishad Premji, reported: an “act of integrity violation”). India’s public floor: the Atal Beemit Vyakti Kalyan Yojana of the Employees’ State Insurance Corporation (ESIC) gives an insured person who becomes unemployed relief for “upto 90 days, once in a lifetime”, at “50% of wages” as ESIC’s page on the scheme states it (esic.gov.in/abvky, last reviewed 4 April 2025); the 2026 reply below does not restate the rate. The person must have been in insurable employment for at least 12 months immediately before, with contributions for at least 78 days in one completed contribution period, and unemployment from misconduct, lock-out or superannuation is excluded (Ministry of Labour and Employment, written reply in the Lok Sabha, PIB, 27 July 2026). The scheme ran from 1 July 2024 to 30 June 2026 (PIB, 8 October 2024) and was extended to 30 June 2027 (PIB, 27 July 2026). Coverage is for employees with wages up to ₹21,000 a month (₹25,000 for persons with disability), the ceiling since 1 January 2017 and still the ceiling in July 2026 (Lok Sabha Unstarred Question 1345, “Wage Ceiling Limit for ESI Coverage”, answered 27 July 2026; esic.gov.in/coverage). The median fresher salary at India’s top IT firms rose from ₹2.45 lakh to ₹3.55 lakh a year between FY12 and FY22, about ₹29,600 a month (a Moneycontrol analysis, reported by *Business Standard*, 26 December 2025). That Naveen was outside the scheme is the book’s inference from these figures; the chapter does not give his pay. Retrenchment: Industrial Disputes Act, 1947, section 25F (one month’s notice or wages in lieu; compensation of fifteen days’ average pay for every completed year of continuous service), carried into the Industrial Relations Code, 2020, section 70, with the worker re-skilling fund of section 83 (the employer pays in fifteen days’ wages last drawn, and the fund credits them to the retrenched worker within forty-five days); “retrenchment” excludes, among other things, retirement and the expiry or non-renewal of a contract of employment, including fixed-term employment (section 2(zh)); the four labour codes in force from 21 November 2025 (PIB, 21 November 2025); the Centre’s draft rules were published on 30 December 2025, and the Industrial Relations (Central) Rules, 2026 and the notification setting up the Centre’s fund (S.O. 2316(E)) followed on 8 May 2026 (PRS Legislative Research; Taxmann, 11 May 2026); Karnataka’s draft rules, which also provide for such a fund, were open for comment until 9 March 2026 (Lexplosion, 29 January 2026); a “worker” excludes those employed mainly in managerial or administrative work and supervisors paid above a set wage (section 2(zr)). Complaints of forced resignations in IT: the Nascent Information Technology Employees Senate (NITES) complained to the Ministry of Labour and Employment in July 2025 that under TCS’s new bench policy employees would “face termination or be pressured to resign” (Storyboard18, 17 July 2025), and wrote to the Chief Minister of Maharashtra that TCS had pressed about 2,500 employees in Pune into “voluntary resignations” without paying “any statutory retrenchment compensation” (*Telangana Today*, 1 October 2025); TCS called the claims “inaccurate and purposefully mischievous” and said those affected had been given “due care and severance”. The Karnataka State IT/ITeS Employees Union (KITU): “Every employee has the legal right to refuse to sign a forced resignation” (*Business Today*, 31 July 2025). Redundancy that sedates: Federal Aviation Administration, Safety Alert for Operators 13002, “Manual Flight Operations”, 4 January 2013 (“continuous use of autoflight systems could lead to degradation of the pilot’s ability to quickly recover the aircraft from an undesired state”; it “encourages operators to promote manual flight operations when appropriate”). Concentration: David Weston, Microsoft, “Helping our customers through the CrowdStrike outage”, 20 July 2024 (“We currently estimate that CrowdStrike’s update affected 8.5 million Windows devices”); flights grounded: Reuters, 19 July 2024, as republished by *The Kathmandu Post* (“Global cyber outage grounds flights and disrupts businesses”). Slack after a lay-off (added with “What would prove this wrong?” in 30.2): David Card, Raj Chetty and Andrea Weber, “Cash-on-Hand and Competing Models of Intertemporal Behavior: New Evidence from the Labor Market,” *Quarterly Journal of Economics* 122, no. 4 (2007): 1511–1560; NBER Working Paper 12639 (2006), abstract (Austria: severance pay equal to two months of earnings lowers job-finding rates by about 8 to 12 per cent, and the longer searches have “little or no effect on subsequent job match quality”). Arash Nekoei and Andrea Weber, “Does Extending Unemployment Benefits Improve Job Quality?,” *American Economic Review* 107, no. 2 (2017): 527–561, abstract (Austria: longer benefits raise re-employment wages, mainly through better firms). Johannes F. Schmieder, Till von Wachter and Stefan Bender, “The Effect of Unemployment Benefits and Nonemployment Durations on Wages,” *American Economic Review* 106, no. 3 (2016): 739–777, abstract (Germany: extensions lower re-employment wages, mainly through longer time out of work). None of the three measures growth.
14. European Commission, EudraLex Volume 4, *EU Guidelines for Good Manufacturing Practice*, Part 1, Chapter 5, “Production” (Brussels, 13 August 2014), paragraphs 5.5 (“Incoming materials and finished products should be physically or administratively quarantined immediately after receipt or processing, until they have been released for use or distribution”), 5.27 (on suppliers of starting materials) and 5.45 (suppliers of primary and printed packaging materials “shall be accorded attention similar to that given to starting materials”).
15. Jana Giceva, “Consensus”, lecture 5 of *Cloud-Based Data Processing*, Technical University of Munich, winter semester 2021/22, slide 4 (split brain; “Prevent two leaders at the same time”; “Require a quorum of nodes to elect a leader in a term”); Martin Kleppmann, *Designing Data-Intensive Applications* (Sebastopol, CA: O’Reilly, 2017), chapter 5, “Handling Node Outages” (split brain, when two nodes each believe they are the leader).
16. Diego Ongaro and John Ousterhout, “In Search of an Understandable Consensus Algorithm,” USENIX Annual Technical Conference, 2014 (extended version, 20 May 2014), section 5.2.
17. Richard G. Tedeschi and Lawrence G. Calhoun, “Posttraumatic Growth: Conceptual Foundations and Empirical Evidence,” *Psychological Inquiry* 15, no. 1 (2004): 1–18.
18. American Psychiatric Association, *DSM-5-TR*, Criterion A for posttraumatic stress disorder (“Exposure to actual or threatened death, serious injury, or sexual violence”), as reproduced in the VA/DoD *Clinical Practice Guideline for Management of Posttraumatic Stress Disorder and Acute Stress Disorder* (June 2023).
19. Vicki S. Helgeson, Kerry A. Reynolds and Patricia L. Tomich, “A Meta-Analytic Review of Benefit Finding and Growth,” *Journal of Consulting and Clinical Psychology* 74, no. 5 (2006): 797–816 (87 cross-sectional studies in 77 articles).
20. Patricia Frazier, Howard Tennen, Margaret Gavian, Crystal Park, Patricia Tomich and Ty Tashiro, “Does Self-Reported Posttraumatic Growth Reflect Genuine Positive Change?” *Psychological Science* 20, no. 7 (2009): 912–919.
21. Mark D. Seery, E. Alison Holman and Roxane Cohen Silver, “Whatever Does Not Kill Us: Cumulative Lifetime Adversity, Vulnerability, and Resilience,” *Journal of Personality and Social Psychology* 99, no. 6 (2010): 1025–1041; the seven categories of events (among them own illness or injury, bereavement, violence and disaster), “Although our data are correlational and therefore cannot establish causality”, and “Our results should not be construed to endorse intentional trauma exposure”, from the article as deposited in eScholarship (University of California).
22. Marcus Aurelius, *Meditations* 8.48, trans. Gregory Hays (New York: Modern Library, 2002); see Chapter 2.

*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 anti-fragility is “an inherent property of consciousness itself”, which “only consciousness makes available”, and that the chapter extended Taleb “into territory he did not explore”; that AI caught “the easy bugs” and missed “the bugs that had no pattern”, leaving Priya’s judgement exposed “like a rock formation revealed by a receding tide”; that the disruption “had forced the transformation” and that the response “is a design principle” that “can be learned”; that the meta-competences are the barbell’s safe end; that “For a person with options, disruption is not a threat. It is a selection event”; that the centre of a distributed organisation performs “a consciousness function, not an information function”; that patience “is a property of consciousness”; that a machine under shock does not “grow” and “does not experience” it; that post-traumatic growth is “real and well-replicated”; that “You are not the thing that breaks”, and that the self “was never breakable”; that a boy’s acceptance of an honest answer was a form of anti-fragility “that requires someone home”. It also told a cloth exporter closing his laptop on the technology for three months and being rebuked into action by a supplier called a “cotton supplier and old friend”; gave a Hamburg buyer a twelve per cent larger order; ran a manager’s redesign for “three months” before a failure in which a team lead accepted “a slight quality variance” from a rerouted supplier; gave a Michigan autoworker’s porch a second telling and put him “five thousand kilometres” from Bengaluru; made a Class 9 student twelve and set his scene in April; opened on a night’s sleep after weeks of broken rest; said Taleb named antifragility “across five books”; and ended on a bridge to the next chapter.
- *What I now hold:* antifragility, in Taleb’s sense, is a gain from shocks within a dose, with time to recover, and it often belongs to a whole while some of its parts break (his chapter 4). In an AI world the clearest antifragile things are machines and the companies that run them. As I read Taleb’s rule, a company may learn from failures its customers met, but may not test less because they will meet them. A person can come out of a shock stronger, under conditions that are mostly slack, and much slack is lent by an employer, a household or the public. Redundancy (spare skills, people, income and knowledge) is built before it is needed and paid for in hours, and it is easiest for those who need it least. Distributing knowledge and decisions needs a rule, written before the failure, for when two sound decisions conflict; and one whole’s redundancy can be one person’s exposure. Reports of growth after a crisis are real as reports and uncertain as change; they coexist with distress, and nobody may demand them. A harm someone decided is someone’s to answer for, and its author may not call it a gift. A gain needs a subject: the same shock can make a system antifragile and a person poorer, and a lay-off is the removal of a load, not a heavier one. A machine’s gain from failure is a loop of human work, and it learns only what is recorded. An employer’s plan for a change decides who is inside it; a floor written in law holds only those whose exit is written as the law describes; and the technology that makes redundancy necessary is often used to remove the slack it needs. Moving knowledge into a machine can move a dependence rather than end it. Taking knowledge from a person is not the same as keeping her judgement, and a company that learns from its people takes on duties to them. A short contract can be a barbell turned upside down. Going on is not always wise, and growth after a loss is not compensation for it.
- *What changed my mind:* Taleb’s own fourth chapter and his ethics; the Prologue’s rule against defining people by what machines cannot yet do; Google’s report of AI-written tests finding a flaw two decades old; Toyota’s stockpile and its limit; Wipro’s dismissals for moonlighting; the European rules on packaging suppliers, which made the print’s “slight quality variance” impossible; computer engineers’ rule against split brain; Tedeschi and Calhoun’s own warning, Helgeson and colleagues’ meta-analysis, Frazier and colleagues’ measurements and Seery and colleagues’ “in moderation”; the canon of this edition, which removes the cloth exporter’s closed laptop, puts the boy in another class in 2026 and gives the Michigan autoworker to the chapter before this one; three outside readings of an earlier version of this chapter, in October 2026; India’s retrenchment law and the labour codes in force from November 2025; the American aviation regulator’s warning on manual flying; the outage of July 2024.
- *What survives:* four of Taleb’s ideas tested (barbell, via negativa, optionality, skin in the game); “Not all suffering teaches”; redundancy as not waste; the finitude of human time and attention, in this edition’s own words; the scalpel of the via negativa, nearly word for word; “Redundancy protects. It can also sedate.”; the four kinds of redundancy, now named plainly; “Both decisions were sound within their local frames” and “decentralisation does not eliminate the need for a centre. It changes what the centre does.”, with the centre narrowed to a rule about what a customer is promised and a named person who tells it; “by Friday” and “never the pace that anxiety demands”; the honest paragraph that post-traumatic growth is measured by self-report and that meaning-making is not gaining density; the last bridge’s buried truth, that anti-fragility can be “the capacity to benefit from disorder while others are damaged by it”; and the sentence “That is a real loss, and it is unfair, and no idea in any book makes it less unfair.”, now Naveen’s.
- *Section titles in the print edition:* 30.5 was “The Consciousness That Cannot Break”. The other four keep their print titles.

*Next in the series:* [*Chapter 31: The Moral Responsibility of Builders*](https://nithinpalal.com/ai/ai-proof/chapter-31-the-moral-responsibility-of-builders)*.*