January 2025. Seoul, Yeonnam-dong, Mapo-gu. 2:14 a.m.
The translator had a callus on the tip of her left index finger. Right where the Shift key sits. For twelve years, she had pressed Shift more than eight hours a day, toggling between uppercase and lowercase. It was the physical trace of moving between English and Korean.
Three law dictionaries were stacked on her desk. One had its cover so worn the title was no longer visible. IT white papers, medical device manuals, legal contracts. Work from her anchor client — steady for six years, three to four jobs a month.
Each night, while she brewed a cup of coffee, the sound of water falling through the dripper was the only noise in the room. Her husband and daughter were asleep. On the left monitor: the source text. On the right: her translation. Her eyes shuttled between the two screens. For twelve years, that rhythm had been her breathing.
A KakaoTalk notification appeared. It was her anchor client. A sentence that began with Thank you. The company was switching to AI translation with native-speaker post-editing. Starting next year.
When she started, the rate was 80 won per word-unit. Five years ago, it dropped to 60. Last year's offer was 35. Now the 35 won was gone too.
She read the notification and took a sip of coffee. It was already cold.
The same day. San Francisco, the Mission District. 6:00 a.m.
A cramped apartment living room doubled as an office. Blue light from four monitors spread across the walls. Three crumpled energy-bar wrappers sat next to a keyboard. Four engineers, three years out of MIT, were at their desks.
They were building a code editor called Cursor. They had checked the server logs the night before. Monthly recurring revenue had crossed $8 million. One of them posted the number on Slack. No one reacted. They were already coding the next update. Just as Arkwright, seeing the first waterwheel turn at Cromford, was already scouting the next factory site.
Twelve months earlier, that number was zero. The team was roughly fifty people. What they had built was a tool that uses AI to write code for developers. Not the technology itself — they did not invent a new model. They took language models from OpenAI and Anthropic and redesigned the entire way developers work.
By the end of 2025, annual recurring revenue surpassed $1 billion. The fastest in SaaS history. Valuation: $29.3 billion. Headcount: 300. It had started with four monitors.
The starting point of an MIT dormitory was itself a privilege. Structural luck runs through this narrative from beginning to end. Crassus had his senatorial lineage. Arkwright had his partners, Strutt and Need.
Same technology. Same era. Opposite trajectories.
In the previous chapter, we examined where capital concentrates — the trinity of data, computing power, and AI models. That concentration produces two human trajectories. The Displaced and the Discerning. What separates them?
1. The Displaced — When Expertise Becomes the Target
Start with the translator.
Forty-nine percent of freelance translators reported a "significant decrease" in workload in 2024. Another 21 percent experienced a moderate decline. Seventy percent, in total, saw their work shrink. Since the launch of ChatGPT, freelance translator income has fallen 30 percent on average. In extreme cases, the drop reaches 60 to 80 percent.
One Irish-language translator had earned a stable income for years from EU institutional work. After the spread of AI tools, his income fell 70 percent. He refused AI post-editing jobs. It felt like training the software that would replace him. The more the AI learned from his work, the more obsolete he became.
South Korea is no exception. Korean freelance translation rates sit at 40 to 70 won per word-unit — roughly the same as 2009. Fifteen years of effective stagnation. The Korean AI translation software market is projected to grow from $100 million in 2024 to $400 million by 2033. AI translation tool adoption has already reached 67.6 percent. In one survey, 90.9 percent of respondents named translation as the profession most vulnerable to AI displacement.
Translation is not alone.
Employment growth in the legal services industry is 1.6 percent — less than half the economy-wide average of 4.0 percent. Growth for paralegals and legal assistants: zero. Effectively frozen. Harvey AI, a legal AI startup, grew its annual recurring revenue 3.8 times in a single year — from $50 million to $190 million. Legal research time has been cut by 60 to 80 percent since AI adoption began.
In accounting, the structural shift is even more visible. Graduate hiring postings at the UK's Big Four fell 44 percent year over year. PwC plans to cut U.S. entry-level hiring by one-third over the next three years. In South Korea, Big Four firms (Samil, Samjong, Anjin, Hanyoung) may reduce new accountant hiring to fewer than 700. Of the 1,200 candidates who passed the CPA exam, only 338 — 28 percent — registered with training firms. The remaining 72 percent passed but cannot become practicing accountants.
The organizational structure of the Big Four is changing. The old "pyramid model" is becoming a "diamond model." The pyramid had a wide base of juniors tapering to a narrow peak of seniors. AI is hollowing out the base. Data collection, ledger reconciliation, routine analysis — these tasks are migrating to machines.
The problem is that when the base disappears, the career ladder weakens with it.
Content writing follows a similar trajectory. Freelance writing job postings fell 33 percent after ChatGPT's release. Entry-level writing roles dropped 27 percent. Freelance gigs declined 35 percent. Polarization is underway. Top-tier writers who strategically edit AI output and design brand voice have seen their rates rise 11 percent. Even within the same profession, the Displaced and the Discerning are diverging.
Platform-wide freelancer data confirms the polarization. Freelancers in AI-exposed occupations experienced, on average, a 2 percent decline in contracts and a 5 percent drop in earnings. The striking finding: high-performing, highly skilled freelancers were hit harder than the rest. AI is leveling the competitive field.
A translator with ten years of experience loses work to one with three years and an AI toolkit. Brynjolfsson and colleagues quantified this: a new hire with two months of experience using AI tools matches the output of a veteran with six months of experience working without AI. The value of experience is being compressed.
2. Inside the Displaced — The Speed of Identity Collapse
Behind the numbers, there are people.
Look again at the translator in Seoul. For twelve years, translation was more than a job. It was a bridge between two languages. Catching the nuance of the source text. Restructuring it so a Korean reader could absorb it naturally. It required deep familiarity with technical documents. It required a feel for Korean sentence rhythm. It required tacit knowledge of industry conventions. That combination was her identity. The law dictionary with its title worn away was the evidence. Just as the smallholder had the field his grandfather cleared, just as the handloom weaver had his five-year apprenticeship, that dictionary was the physical trace of twelve years of accumulated expertise.
Six months earlier, she had given birth to her second child. The translation market did not pause during her maternity leave. When she returned, the landscape had changed. Rates had fallen 40 percent. In six months, DeepL Pro had updated. Two domestic AI translation engines had entered the market. Two of her three clients had already switched to AI-first translation.
In 138 BC, a smallholder returning from seven years of military service in Hispania found that the boundary stones of his field had vanished. The olive trees stood in the wrong places. The translator, returning after six months, found that the boundary lines on the rate sheet had disappeared. The duration of absence differed by 2,163 years, but the structure of the changed world they returned to was identical.
On her monitor was a translation the AI had produced in thirty minutes. It would have taken her two days. She read it line by line. Her left hand hovered over the Shift key and stopped. The context-dependent terminology was off. Legal nuance was missing. Industry-specific expressions were wrong.
She marked the errors in red. Corrected them. Then her fingers stopped again. Because it occurred to her that every correction she made trains the AI to be more accurate.
The AI had begun to imitate that combination. Not perfectly. The question was whether the client noticed the difference. Most did not. "Good enough" defeats "accurate."
The dripper had stopped. No more water falling. The room went silent. The only sound was her daughter turning over in the next room.
A 2025 study identifies six psychological stages of AI displacement: emotional shock, erosion of professional identity, chronic anxiety, social withdrawal, attempted adaptation and frustration, and organizational betrayal. Sixty-eight percent of white-collar workers worry that their role could be automated within five years.
The upskilling paradox is the core dilemma. Sixty-one percent of workers "consider" upskilling in response to the AI threat. The share actively "pursuing" AI-related training: 4 percent. A 57-percentage-point gap. The chasm between intention and action.
The structure of this gap is identical to the handloom weavers of the Industrial Revolution. As we saw in Chapter 9, the weavers knew the power loom was a threat. They could have learned to operate the machines. Most did not make the transition. Factory work collided with their identity. Independent artisans. Skilled craftsmen who worked on their own time, in their own homes. They could not surrender that pride.
The modern knowledge worker faces the same bind. She knows she should learn AI. But learning AI tools means mastering the technology that renders her expertise meaningless. For a translator, AI post-editing is learning "how to make my replacement more efficient." The psychological resistance is not irrational. It is a rational instinct of self-preservation.
In South Korea, this dilemma cuts deeper. Sixty-one point three percent of Korean workers are employed in occupations at high risk of AI and robot displacement. More than half of office workers report feeling their jobs threatened by AI. At the same time, workers who use AI daily score higher on salary and job security measures. Within the same organization, the same tool produces opposite outcomes.
3. The Discerning, Type One — Leaders Who Transform Organizations
Before we turn to AI-native startups, consider the leaders restructuring existing organizations with AI as leverage.
In April 2025, Shopify CEO Tobi Lutke posted a company-wide memo. The message was simple. AI proficiency is a baseline expectation. If you want to hire additional staff, first prove AI cannot do the job. AI usage is part of performance reviews. The CEO himself is no exception.
The same week, Fiverr CEO Micha Kaufman sent a similar memo. "AI is coming for your job. Including mine." Then he cut 250 positions.
IBM has been driving AI-powered productivity transformation since early 2023. By the end of 2025, it achieved $4.5 billion in annual savings. A 6,000-person software team adopted AI coding assistants. Productivity rose 45 percent in four months.
This is the twenty-first-century version of the factory owner telling the handloom weaver: "Use the power loom or leave." With one difference — this time, the factory owner admits he is also at risk of being replaced.
Their core competency is not the technology itself. It is the ability to design the integration of AI and humans. Arkwright was not an inventor; he was a systems designer. These leaders are not AI engineers. They are architects who restructure a 100-person organization to achieve 10-person efficiency using AI.
4. The Discerning, Type Two — Those Who Create Systems
Back to Cursor.
The story of this company, born in an MIT dorm, echoes one from 250 years earlier. In 1768, a Preston wigmaker named Richard Arkwright met the clockmaker John Kay. Arkwright did not invent the spinning frame. He took the technology of Kay and Highs and organized it into a system called the factory. The competitive advantage was not the machine. It was the system.
When his patent was revoked in 1785, Arkwright's business thrived. You could take the machine. You could not take the system.
The four at Cursor follow the same logic. They did not invent the large language model. They took models built by OpenAI and Anthropic and redesigned the entire developer workflow. Not a tool that writes code for you — a system that redefines the act of coding itself. Annual growth rate: 1,100 percent. Zero to $1 billion ARR in 24 months.
Arkwright needed roughly seven years to turn his first profit at Cromford Mill. Time itself has been compressed.
Cursor is not alone.
Midjourney employs 130 to 163 people and generates $500 million in annual revenue. External venture capital: zero. Profitable within one month of launch. Revenue per employee: $3.1 to $3.8 million — ten to thirteen times the traditional SaaS average of $300,000. Its Discord community doubles as its marketing channel. Marketing spend: zero. The product is the marketing.
Three twenty-year-olds from the Bellarmine high school debate team founded Mercor, an AI recruiting platform, in 2023. Two years later: a $10 billion valuation. They raised $100 million. They run a small economy, paying $1.5 million daily to 30,000 contractors. Revenue per employee: $4.5 million.
AI-native companies are emerging in South Korea as well. Wrtn has rapidly grown its user base as a generative AI platform. Upstage developed its own LLM, Solar, and is expanding into global markets. These companies are designing AI systems not in Silicon Valley, but in Seoul.
The speed at which AI-native startups reach $1 million ARR is rewriting history. The average: 12 months. That is 5.1 times faster than the SaaS-era average of 60 months. Salesforce took 72 months. Dropbox took 60. Slack took 36.
What do these companies have in common? Five things.
First, systems design. The ability to architect the integration of AI and humans. Cursor redesigned the developer workflow. Midjourney redesigned the creative workflow. Mercor redesigned the hiring workflow.
Second, speed-first iteration. Build fast, absorb market feedback, rebuild. Bolt.new was on the verge of shutting down when an AI pivot delivered $40 million ARR in five months.
Third, platform leverage. Combining LLM APIs, cloud infrastructure, and open-source tools. They do not own the technology. They compose it.
Fourth, community building. Midjourney's Discord. Cursor's developer ecosystem. They create self-reinforcing loops around the product.
Fifth, the judgment of "necessary and sufficient." Distinguishing between domains where AI is enough and domains where humans are essential. This is the most subtle capability, and the most important.
5. The Discerning Can Be Wrong Too — The Klarna Lesson
The fifth capability matters because Klarna shows what happens without it.
In February 2024, fintech company Klarna boasted that its AI assistant was performing the work of 700 employees. Customer satisfaction scores were equivalent. Repeat inquiries dropped 25 percent. An estimated $40 million in savings. CEO Sebastian Siemiatkowski was invited to speak as a model of AI transformation.
Total headcount was cut from 7,000 to 3,500.
A year later, he admitted that the AI transition had negatively affected service quality. Klarna began hiring humans again. An "Uber model" compromise emerged: AI handles simple inquiries, humans take complex cases. It was a case study in the limits of AI maximalism.
Arkwright's story has no such scene. The factory system was overwhelmingly more efficient than the handloom. The Discerning of the AI era must be more humble. Systems design includes drawing the right boundary between human and machine.
Duolingo had a similar experience. It declared an "AI-First" strategy and replaced contract content creators with AI. The CEO announced that "the same headcount can produce four to five times more content." Users pushed back. The feedback: the content felt repetitive and mechanical.
Vibe coding tells the same story. Coined by Andrej Karpathy in early 2025, the approach delegates code generation to AI and checks only the output. Twenty-five percent of Y Combinator's Winter 2025 batch had 95 percent of their code generated by AI. Collins Dictionary named it the 2025 Word of the Year. One report found security vulnerabilities in 170 of 1,645 web apps built with vibe coding — 10.3 percent — with exposure serious enough to allow access to personal data.
The trade-off between speed and quality.
Survivor bias must be acknowledged here. The success of Midjourney and Cursor does not represent AI startups as a whole. The failure rate for AI startups is 90 percent — significantly higher than traditional tech startups at 70 percent. Median survival time: about 18 months. According to MIT's NANDA research, 95 percent of enterprise generative AI pilots produce no measurable impact on the bottom line. Only 5 percent achieve revenue acceleration.
Behind every Midjourney, hundreds of AI startups vanished within 18 months. Behind every Shopify AI transformation, 95 failed AI pilots.
The founders of successful AI-native companies had structural luck on their side. Elite education (Cursor's MIT pedigree, Mercor's Silicon Valley network). Access to capital (Mercor raised $100 million). Geographic advantage (most are San Francisco-based). Timing (the overwhelming majority were founded in 2022-2023). Capability and fortune are always intertwined.
Arkwright had structural luck too. The natural waterpower of Cromford. A Derbyshire location safe from the Lancashire riots. The right partners — Strutt and Need. Crassus had the one-time opportunity of Sulla's proscriptions. The success of the Discerning always includes structural fortune.
6. The Crossover — Why the Same Technology Produces Opposite Outcomes
The same day. 9:00 a.m.
The translator in Seoul opened the post-editing screen. Her task: catching errors in a translation the AI had produced in thirty minutes. On the left monitor: the AI output. On the right: the source text. The rhythm she had known for twelve years was reversed. She used to start from the source and arrive at her own language. Now she started from the AI's language and hunted her way back to errors. The direction had flipped.
She knew that every line she corrected made the AI more accurate. She was building her own replacement with her own hands. Just as the smallholder's military service supplied the slaves who replaced him on the land. Just as the handloom weaver's demand for cheap thread proved the economics of the power loom. The Displaced were trapped inside a structure that accelerated their own displacement.
The same hour. San Francisco.
A Cursor engineer was studying the user logs. He was extracting training data from the patterns developers left when they corrected AI-generated code. Structurally identical to the translator's post-editing. User corrections made the product more accurate. The difference: this side was designing that structure.
One person was inside the structure. The other was building it.
This distinction cuts across all three eras. The smallholder plowed his own field. Crassus designed a buy-rebuild-rent system. The handloom weaver operated his own loom. Arkwright designed the entire factory system. The translator translates. The Cursor team designs the developer workflow. The Displaced occupy the execution layer. The Discerning occupy the design layer.
The structural asymmetries created this distance. Capital, information, scale, time horizon. The four asymmetries first identified in Chapter 5 are operating here for the third time. The translator funds her own retraining. Cursor accesses over $100 million in venture capital. The translator underestimates the pace of AI advancement. Cursor experiments on the front lines every day. The translator is confined to her own capacity. Cursor's team of 40 achieves the output of tens of thousands. The translator survives month to month. Cursor designs on a three-to-five-year vision.
But as we saw in Chapters 5 and 9, knowing the asymmetries exist is not enough. In Chapter 16, we will ask why these four and whether they constitute a law or a lens.
Here is the insight worth writing down. The "Arkwright Test" derived in Chapter 9 can be applied to the AI era. The original formulation: "If this company's patent were revoked tomorrow, would it still have a competitive advantage?"
The AI-era version: "If this organization's AI tools were swapped out tomorrow, would the competitive advantage survive?" Replace ChatGPT with Claude. Replace Claude with Gemini. If they still win, they are the Discerning. If not, they are merely dependent on a tool.
7. Across Three Eras — The Dematerialization of Leverage
Place this chapter's two figures alongside the two Romans of Chapter 5 and the two Britons of Chapter 9.
The speed of decline is accelerating. The Roman smallholder's fall took 150 years — a gradual descent spanning two to three generations. The handloom weaver's fall took 30 years — completed within a single generation. The AI-era translator's fall takes roughly three years — unfolding within her own career. An acceleration of approximately fifty times.
In annualized income decline, the pattern sharpens further. The smallholder: 0.7 percent per year. The handloom weaver: 5.7 percent. The translator: 30 percent. Each era accelerates five to eight times over the last. The final magnitude of decline is similar — income loss exceeding 80 percent. The difference is how quickly they get there.
Why has it accelerated so dramatically? Because the marginal cost of the replacement technology is converging on zero. A single slave cost 300 to 500 denarii. A power loom cost several hundred pounds. A single LLM API call costs $0.001.
When the marginal cost approaches zero, there are no physical constraints on diffusion. No roads to build. No factories to construct. ChatGPT reached one million users in five days.
The profile of the Discerning is changing too.
The dematerialization of leverage. Crassus's leverage was tangible — land, buildings, 500 construction slaves. Arkwright's leverage could be described but not touched — the factory system, shift rotations, a licensing network. The AI-native founder's leverage is difficult even to describe — API calls, prompt strategies, cognitive integration. The source of wealth is migrating from physical assets to cognitive composition.
The democratization of entry barriers. Becoming Crassus required senatorial lineage. A door open to less than 0.01 percent of the population. Becoming Arkwright required organizational ability and modest capital. Open to 1 percent. Becoming an AI-native founder requires domain knowledge, AI proficiency, and a laptop. Theoretically open to millions.
A ChatGPT Plus subscription at $20 per month is roughly one forty-thousandth of Arkwright's initial capital of 500 pounds.
"Anyone can do it" comes with a significant caveat. Access to tools has been democratized. The ability to use them has not. According to PwC, the AI-skills wage premium is 56 percent — double the previous year's 25 percent. The tools are nearly free. The proficiency gap is creating a new inequality.
The shifting complexity of moral responsibility. Crassus forced fire-sale purchases at burning buildings. Explicit exploitation. Moral responsibility could be assigned to an individual. Arkwright ran child labor and 13-hour shifts. Systemic exploitation. The weaver's ruin was a consequence, not an intention.
The AI-native founder did not intend to put the translator out of work. OpenAI did not build language models to replace translators. Duolingo cut contractors as a cost-optimization decision. Who bears responsibility?
The more moral attribution disperses, the harder institutional response becomes. It is paradoxical. The speed of decline accelerates while the capacity to respond grows weaker.
Technology and capital move before institutions do. In the gap before institutions catch up, the Discerning build wealth. The Displaced pay the price. That sentence first appeared in Chapter 5, about Rome. It was confirmed during the Industrial Revolution. In the AI era, it is being confirmed for the third time.
8. Are You the Displaced, or the Discerning?
Return to the opening of this chapter.
2:14 a.m. in Yeonnam-dong and 6:00 a.m. in the Mission District. Cold coffee and crumpled energy-bar wrappers. The distance between their trajectories cannot be explained by technical skill alone. Position in the value chain, access to capital, access to information, time horizon. These structural asymmetries created that distance.
There is a trap in the question itself. "The Displaced" and "the Discerning" are not fixed categories. Some Roman smallholders migrated to the city and found new opportunities. Some handloom weavers became factory supervisors. Some translators are using AI tools to multiply their productivity three to five times over.
An analysis by the Bank of Korea captures this duality. Twenty-seven percent of Korean workers fall into the category of "high AI exposure, low complementarity" — the at-risk group. Twenty-four percent fall into "high exposure, high complementarity" — the beneficiary group. Exposed to the same technology, the presence or absence of complementary skills produces opposite outcomes.
Two people fall into the same water. The one who can swim and the one who cannot meet very different fates.
Recall the central thesis of this book: "When productivity explodes, economic structures shift, and people's lives change at their foundations." It happened in Rome. It happened during the Industrial Revolution. It is happening in the age of AI.
If there is a difference this time, it is that we know the pattern. The handloom weaver did not have the information to predict that the power loom would destroy his world. The smallholder had no means of measuring the speed of latifundia expansion. We have that information. The fact that you are reading this right now is the evidence.
Does information convert into action? That is what separates one person's fate from another's.
Individual adaptation alone is not enough. The handloom weaver's problem was not personal laziness. It was institutional absence. The Factory Acts took 64 years to arrive. The smallholder's problem was not personal incompetence. It was the failure of the Gracchan reforms.
The translator's income is falling 60 percent in two years. Accounting firms are cutting junior hiring by 44 percent. Who should manage this transition? For the handloom weaver, factory legislation came a century late. How quickly must it come this time?
In the next chapter, we examine the institutions that have not yet arrived. Education, labor law, taxation. What must be redesigned for the age of AI?
End of Chapter 14. Next: Chapter 15 — Institutions Yet to Come: Education, Labor Law, Taxation — What Must Be Redesigned