Complete English translation.

Preface · Eleven Hundred Years

Chang’an, 587 CE.

Yang Jian is paging through a list.

The seventh year of Kaihuang. Officials need filling. Most of these names he doesn’t know; each carries the name of a recommender. The recommenders he knows. The Three Departments’ clerks, the prefectural Rectifiers — for decades they’ve cycled through, the same handful of clans.

Cui. Lu. Wang. Xie.

These names appear on the dispatches he signs, and on the dispatches his grandfather signed, and on dispatches sent by emperors well before that. Northern Zhou is gone. Northern Qi is gone. Liang is gone. Chen is about to fall. These names have not.

He has unified the north and is about to unify the south. From his generation forward, this land will no longer be the north or the south of one dynasty or another. It will be Sui. But every appointment he intends to issue must first pass under these names.

The Nine-Rank System has run for three centuries. Rectifiers grade local men into nine ranks; the central court appoints by grade. On the surface it picks talent; underneath it counts pedigree. An emperor seated in Chang’an, every appointment he sends out has already passed through the brushes of a few dozen clans in the provinces.

Yang Jian wants this country to obey him.

He wants a method. No Rectifiers. No recommendations. No clans passing things along. Anyone, if they can pass a public standard, gets identified directly.

In the seventh year of Kaihuang, the court issues an edict: examination by category.

A line begins to show its shape from here.

For thirteen hundred years afterward, the details would be revised over and over. Subjects changed. Content changed. The palace examination appeared. Anonymous transcripts appeared. The eight-legged essay appeared. Dynasties turned over, again and again. What stayed unchanged underneath was one thing:

Let a man seated at the centre pick, out of ten million households, the small handful that can be used; and let that picking pass through no pedigree, no recommendation, no intermediary.

The exam was not designed to give scholars a path. Giving scholars a path was a by-product.

Eleven hundred years pass.

Potsdam, 1763.

Frederick II, seated at Sanssouci, signs a decree.

The Seven Years’ War ended half a year ago. Prussia won. The peace is a winner’s peace, but this winner is nearly hollow. Half a million dead in battle and famine. Eastern villages emptied wholesale. Tax revenue close to dry. What carried him through the war was a standing army trained by the state and an industrial apparatus directed by the state.

With the war over, he knows the next will come. He needs a particular kind of person — one who can shuttle between factories and barracks, who can be slotted into a position inside standardised processes.

This kind of person does not appear on his own.

Children in the countryside cannot read, cannot do arithmetic, cannot keep time, cannot accept abstract orders. They answer to their fathers, to their clans, to the parish priest. The state is not on their list. To get the state inside them, the state has to intervene earlier.

On 12 August, Frederick signs the Generallandschulreglement — the General School Regulations.

Children aged 5 to 13 must attend. The state prescribes the curriculum.

It is the foundational text of Prussian state-mandated schooling, and a key reference for later national-education models across Europe.

Compulsory education trains four things.

Literacy. Arithmetic. Punctuality. Obedience to graded authority.

Literacy makes orders legible. Arithmetic makes basic tasks in factories and armies possible. Punctuality makes mass-scale coordination possible. Obedience makes a bureaucracy able to assign each man his slot.

The process of producing a child into what the state needs can now begin before the child knows what the state is.

Another hundred and fifty years pass.

Highland Park, Michigan. 1913.

Henry Ford is looking at a report.

It says that this year, to maintain thirteen thousand positions, the Ford plant hired fifty-two thousand people. Turnover: 380%.

Mass production of the Model T began in 1908. Earlier this year, he pushed the moving assembly line to full load for the first time. Chassis assembly time dropped from 12.5 hours to 93 minutes. A car got assembled in less time than it took a worker to finish lunch.

But the plant has a problem.

Workers don’t come. Those who come don’t stay. A machine sitting idle while it waits for someone who can turn screws costs more than the screw-turning itself.

The reason is clear. Most of the floor is recent immigrants. Poles, Italians, Greeks, Eastern European Jews, Mexicans. They cannot follow the foreman’s English. They cannot read the safety signs. They cannot remember which step their badge number maps to. They were used to one tempo back home; the factory needs another. The gap between a peasant arriving from the countryside and a man on the assembly line is something compulsory education does not bridge. Compulsory education gave him the four basics: literacy, arithmetic, punctuality, obedience. What the plant needs is for him to stand at his assigned station and, on a 93-minute beat, repeat one specific motion calibrated to the second. No existing system teaches that.

That year, Ford founded the Ford English School. Within a year, several thousand immigrant workers were attending classes inside the plant and after hours. The curriculum included English, American civics, personal hygiene, how to use a bank, how a household in America budgets. The next year it expanded into the Sociological Department: fifty investigators visited workers’ homes to assess whether they were “ready to be a Ford worker.” Those who passed earned twice the industry rate — five dollars a day.

It is one of the earliest large-scale corporate adult-training systems documented in the archives — paid for, designed, and run by the company itself.

The problem it tackles is not whether workers can read. Compulsory education has already answered that. The problem is the next one:

Once a person has been filtered through compulsory schooling and identified by the state as a usable object, how do you produce him further into an execution unit that can drop, immediately, into one specific position.

For a century afterward, the form would change repeatedly. Apprenticeships became on-the-job training, OJT became corporate universities, corporate universities became employee development systems, and those became today’s leadership programs, role certifications, internal business schools. The thing underneath did not change: take people who have passed the universal filter and produce them, further, into people a specific organisation can use directly.

The imperial exam answers a selection problem: given ten million households, how do you identify the usable objects.

Prussian compulsory schooling answers a production problem: how do you turn ten million households into identifiable objects in the first place.

Ford’s training apparatus answers a fitting problem: how do you take an already-identified object and machine him further into someone a specific organisation can use at a specific position right now.

Stack the three together and you get the full structure of modern teaching and training. First produce people into identifiable objects. Then filter the small subset that is usable. Then machine the usable into units that drop straight into a specific role.

After the twentieth century, every industrialised country runs all three machines at once. Compulsory schooling, exams and credentials, corporate training. The shapes vary. School-leaving ages vary, exam formats vary, the selection points vary, the language varies. Corporate training varies too — apprenticeships, onboarding, OJT, internal business schools, leadership programs, external MBAs. But literacy and arithmetic, mass-scale classrooms, age-graded cohorts, subject exams, credential ladders — plus role descriptions, SOPs, training manuals, performance reviews, promotion tracks — that whole understructure holds in every modern state and every modern organisation.

The point isn’t new in the academic literature. Bourdieu and Passeron’s Reproduction (1970), Bowles and Gintis’s Schooling in Capitalist America (1976), Andy Green’s Education and State Formation (1990) — half a century pointing at the same thing: the school’s actual function is to produce, for the existing state and economy, people who can be used.

Teaching and training together form a human-capital production system designed for the purpose of using people.

At its design moment, the system does not assume responsibility for whether a person lives well. The question it answers is:

Can this person be used.

Chapter 1 · The Current State

Yang Jian, Frederick, Ford — three points already in the past. Three problems, not in the past: who can be filtered out, who can be produced into a usable person, who can be fitted to a specific position. The system has run for another hundred and ten years since Highland Park in 1913. In April 2026 it is still running, but its shape is starting to deform.

Seven Cognitive Mechanisms

“Education trains people” sounds like one thing. Pulled apart, it is seven. Each one has different plasticity, a different fate after AI, and a different training method.

First is knowledge (crystallised intelligence, Gc). The facts, theorems, vocabulary, and case stock a person stores in his head. Reading, classes, problem sets, accumulated experience — most of the thickness lives here. This layer can accumulate without limit. After AI its scarcity has fallen close to zero; AI’s fact base outstrips any single person’s. But having read it, remembered it, understood it still has a use: when AI gives an answer, you can decide right away whether it’s true.

Second is fluid intelligence (Gf). The on-the-spot reasoning ability that, given a situation you’ve never seen, picks out structure, finds an entry point, and constructs a path. It does not lean on knowledge; it leans on whether a person can fire up thought in a new setting at all. This is hard to lift directly through training. Decades of psychology converge on the same conclusion: education and brain training do not reliably, far-transferably raise the Gf ceiling. What education can do is not raise the ceiling out of thin air; it is to make the next three layers solid.

Third is activation frequency. Even a person with a high Gf ceiling, if he doesn’t use it for a few years, leaves it idle. Gf, with no external pressure, does not start itself. The number of times a person is pushed in childhood and adolescence to the “have to use it” edge determines how much of his reasoning he can actually mobilise as an adult. This is where problem-bank training really earns its keep. Good problems manufacture structural unfamiliarity and force the student to model on the spot; bad problems train standard responses, and the more you grind them the less you think. After AI this layer does not depreciate, because AI has no continuous subject that can be trained — it cannot maintain your thinking-state for you.

Fourth is the strategy repertoire. Every solution painfully reasoned out the first time can be called up directly the next time the same shape appears. Chess has its formations, physics its tricks, engineering its feel. The gap in professional problem-solving is mostly not in Gf; it is in the depth of the repertoire. After AI this layer splits in two. The shallow, template-shaped strategies AI has all of, faster and more accurately than you. The deeper layer — judgments tied to specific domains (“this machine sounds wrong from the moment it starts up”) — AI cannot pick up, because it has no on-site sense.

Fifth is foundational automation. Reading, calculating, processing text, writing, all trained to the point where they no longer require dedicated thought. This layer determines how much cognitive capacity is left over when you do high-level work. Japanese basic education invests most heavily here; Japan and the Nordic countries top PIAAC adult assessments precisely because of this. After AI the role of this layer changes: it is no longer producing for you, it is letting you decide right away that “this looks wrong” when AI hands over an answer. A person without solid foundations using AI is outsourcing judgment to a system that cannot bear responsibility.

Sixth is cognitive decoupling. Stripping abstract structure out of concrete context and operating on it as a formal object. Proving a geometry theorem, pushing an algebraic step, debugging a program — they are doing the same thing. It is closely related to Gf but narrower, more specialised, more effortful. ISSUE 01 placed it at the centre of “the real divide in the age of AI.” After AI this layer is the scarcest. AI itself is a product of cognitive decoupling — language stripped into probabilities, reasoning stripped into parameters — but it cannot do the decoupling on a specific person’s behalf. AI can hand over the result; the act of decoupling has to be done by a specific person.

Seventh is metacognition. “Where do I not understand”, “under what conditions does this method hold”, “did I just skip a step too fast.” It is a habit ground in over years through real tasks and real feedback. The old education barely trains it. Exams don’t test it, teachers can’t grade it; it happens inside the student’s own head. After AI this layer is extremely scarce, because faced with AI’s output, “should I trust this right now” is the judgment AI cannot do for you.

Lay the seven side by side and a faint pattern shows up: the layers the old education invested most in (knowledge, the shallow part of activation frequency, the shallow part of the repertoire, foundational automation) are precisely the ones starting to depreciate after AI; the two layers it barely invested in at all (cognitive decoupling and metacognition) are precisely the ones AI has made scarcest.

Teaching and Training

The Preface introduced these two words.

Teaching is school education, from primary school to university. It does two things: sort people into ability tiers, and shape each tier into an identifiable object. Credentials, scores, certificates, school rankings — these are its outputs.

Training is what happens at work: apprenticeships, on-the-job mentoring, corporate training, master-apprentice pairs, industry certifications. It does one thing: take people who already have credentials and shape them into people who can drop, immediately, into a specific role. Role experience, process fluency, reporting skill, collaboration habits — these are its outputs.

Teaching produces sortable people. Training produces embeddable people. Both held in the past, because the world really did need large numbers of sortable, trainable, embeddable people.

Disqualified

Begin by separating “disqualified” from “failed”. A failure is something done badly. Disqualification is something whose qualification no longer holds. A thing can be done very well and still be disqualified.

The old education and the old training are not without use. They were once usable, because knowledge, experience, fluency, and formal compliance were genuinely scarce in the past. A person who could write a passable essay was, more often than not, someone who genuinely understood. A person who could write working code was, more often than not, someone with engineering chops. A person who could deliver consistently was, more often than not, fit for the role. The form indirectly testified to the substance.

What AI has changed is that proof relation. The form still works, still delivers, still gets people to feel “this is well done.” But it no longer reliably testifies that a corresponding subject stands behind it. ISSUE 04 called this “form being able to be produced independently of substance”.

Teaching and training are disqualified at exactly this point: the “qualified” they used to produce indirectly testified to actual qualification; today it doesn’t.

Four Reference Systems

Comparing national education systems is dangerous work. It falls too easily into national caricatures: East Asia drills, the Nordics free, America innovates, Japan disciplines. Each label is half right, and none of them can be used as an argument.

This essay picks China, the United States, Japan, and the Nordics as four reference systems, because between them they cover almost all of the possible shapes “teaching and training” took before AI. Each system is a complete, internally coherent design for producing people into “usable objects,” answering one specific historical problem. Plotted against the seven mechanisms above, each invests in a different direction.

China’s core is filtering by problem bank. From the Sui imperial exam to the contemporary gaokao (the make-or-break Chinese university entrance exam), to the recent wave of “top-tier innovative talent” programs in the Yangtze Delta and Beijing, the logic has not moved: through large-scale, tiered, feedback-rich standardised problems, push students repeatedly against the edge of their existing ability, trigger on-the-spot reasoning, and rank the results into a stable order. Its strength is high-density Gf activation and stable ranking. Of two people with the same ceiling, the one who has been pushed against the edge again and again will get more done than the one who never has. Its weakness is that once problem types have been exhaustively patternised, training degenerates from “triggering judgment” to “training standard responses.” Of the seven mechanisms, China’s thickness sits in knowledge, activation frequency, and the strategy repertoire.

Japan’s core is foundation and embedding. From the moment Meiji reform imported the Prussian model, the system trained literacy, calculation, reading, attention, and rule-awareness to extraordinary thickness — thick enough that by the time a student enters work, those motions are essentially automatic. Then, inside firms, on-the-job mentoring, rotation, and lifetime employment shape him into someone who can carry a specific function inside organisational processes for the long run.

The Japanese management scholar Ikujiro Nonaka summarised this kind of capacity as anmokuchi — tacit knowledge — that cannot be encoded in writing and only deposits over years on the post. Its strength is that foundational automation runs deepest: in PIAAC, Japan stays in the top tier because Japanese adults’ basic cognitive abilities don’t depreciate badly with age. Its weakness is that mass-scale social retraining is inflexible and cross-industry mobility is hard. Of the seven, Japanese thickness sits in foundational automation and knowledge.

The United States runs on a different logic: don’t try to push everyone, concentrate resources at the top. Top private schools, research universities, AP classes, honours classes, magnet schools, undergraduate research — that pipeline builds a chain of high-intensity cognitive environments, lets people who are already high-Gf get identified through open-ended tasks, and pours resources to amplify them. It is willing to accept that the median is unimpressive and the bottom is bleak, in exchange for the right-tail explosion at the top.

The intensity of cognitive-decoupling and metacognition training at its top end is unmatched anywhere else. To this day, the overwhelming majority of the world’s top AI researchers were trained in the American top tier. The cost is that almost nothing is invested below the median. By 2026 the picture is more extreme than that description suggests: the top and the bottom have already split into two different countries. The NAEP 2024 numbers below will lay this out.

The Nordics (with Finland as the case) take a fourth path: low pressure, average, long. They don’t push the student. They lean on small school-to-school variance, high teacher autonomy, and stable foundational training to bring the vast majority up to a respectable floor. Then the real weight goes onto post-school life: adult retraining, lifelong-learning entry points, and a learning-rights regime co-funded by unions, employers, and the state. A person at 30, 40, 50 is still learning new things. The PIAAC top tier is Finland, Sweden, Norway, the Netherlands, and Japan, all for this reason: Nordic adults don’t depreciate. The strength is long-term maintenance of the population-wide floor; the weakness is a lack of concentrated training in deep elite schemas. Of the seven mechanisms, the Nordic system has neither a particularly strong nor a particularly weak layer.

The four systems lean in completely different directions because the historical problems they answered are not the same. China had to filter usable engineers out of an enormous population. Japan, in postwar reconstruction, had to shape people into reliable members of industrial organisations. America, in an open society, had to keep producing entrepreneurs and innovators. The Nordics had to maintain the long-run productivity of a small high-welfare society.

Put together, these four systems cover almost every shape “teaching and training” could take before AI.

Several systems sit outside the main set and need a brief mention. South Korea is not in the main set but works better as a counter-example: students score very strongly on assessments, while adult skill profiles are noticeably weaker, and the hagwon ecosystem makes interpreting the assessments harder. It will appear later as a case in negative: the AI digital textbook withdrawn within four months in 2025, and three sub-OECD-average PIAAC 2023 adult scores — two ends of the same thread. Germany’s dual-track system is another vocational benchmark, with narrower coverage than China or the US, and will appear later for contrast. Singapore is a miniature hybrid, too small in scale to be a primary reference.

Together, these four systems cover the most distinct shapes “teaching and training” took.

Students: PISA 2022

PISA is the OECD’s triennial assessment of 15-year-olds in mathematics, reading, and science. The 2022 results are below (mainland China did not participate in 2022).

Mainland China last participated in PISA in 2018. The B-S-J-Z sample (Beijing, Shanghai, Jiangsu, Zhejiang) scored 591 in math, 555 in reading, 590 in science, covering 81% of the 15-year-old population in those four regions — not a national sample. Mainland China is also missing from PIAAC, and PISA only has those four regions, so strictly speaking there is no “China-wide” object available for international comparison.

PISA 2022 also includes a creative-thinking assessment. Singapore led with 41, Korea 38, Canada / Australia / New Zealand / Estonia / Finland between 36 and 38, OECD average 33. Korea’s high score has to be read carefully: it shows that Korean students perform very strongly under PISA’s test framework, but the private hagwon tutoring ecosystem, international competition entries, and a track record of cheating in tests can all affect the external validity between assessment performance and long-term adult ability. That is this essay’s reading, not a conclusion of the PISA data itself.

Adults: PIAAC 2023

PIAAC is the OECD’s adult-skills survey, assessing literacy, numeracy, and adaptive problem-solving in the 16-65 population. It assesses what people who left the education system years ago actually do with information at work and in life — closer to the ultimate output of teaching and training than PISA. 2023 was the second round.

Mainland China did not participate.

Nordics + Japan is the strongest adult-skills cluster, top tier across all three. The United States is below the OECD average on numeracy and adaptive problem-solving, and its 16-24 cohort also scores below average on all three. Korea is below average on all three; relative to the previous round, literacy is down 23 points, numeracy down 10.

Classrooms: The TIMSS Video Study

PISA and PIAAC measure outcomes. TIMSS, in 1995 and 1999, did something rare: it filmed real classrooms and analysed actual time allocation in 8th-grade math classes. 1995 covered Germany, Japan, and the United States; 1999 added Australia, the Czech Republic, the Netherlands, Switzerland, and Hong Kong.

Distribution of student work time across three categories of activity:

Share of 8th-grade math lessons earning the highest deductive-reasoning rating: Japan 39%, Germany 28%, United States 0%.

TIMSS 1999 has another set of numbers. Among teacher-designed problems requiring connections between different concepts (“making connections” problems), the share that actually got discussed and worked through in class: United States 8%, Australia extremely low, the other five systems (Japan, Czech Republic, Hong Kong, Netherlands, Switzerland) between 37% and 52%. Researchers gave Australia’s pattern a name of its own: “shallow teaching syndrome.”

These numbers are from 1995 and 1999. The exact percentages may have shifted; the structural difference largely has not.

America’s Bottom: NAEP 2024

NAEP is the U.S. National Assessment of Educational Progress — popularly, the nation’s report card. The 2024 round produced the clearest picture of the American education system in its current state:

In 4th-grade reading, the gap between the top 10% and the bottom 10% is 107 points, one of the largest in NAEP’s history. The 8th-grade math top-bottom gap is the largest ever recorded. In absolute terms: the bottom 25% of students score lower in reading and math than students at the same percentile in 1992. Over 32 years, the worst-off slice has not improved — it has gone backward relative to a generation ago. The 4th-grade lowest percentile is at a 20-year low; the 8th-grade lowest is the lowest in NAEP history.

The widening gap is not noise. From 2005 to 2024, the top 10% in 4th-grade math rose 4 points and the bottom 10% fell 3; in reading, top up 3 and bottom down 5. Harvard and Brown’s joint Annenberg Institute estimates that, over those two decades, internal achievement inequality in U.S. education widened by an amount equivalent to 1.3 academic years of learning. Add the pandemic, and by spring 2024 the national student average is still about half an academic year behind pre-pandemic levels, with no state having returned to its 2022 reading level.

At the same time, the American top end keeps producing the global AI industry’s core talent. MacroPolo’s Global AI Talent Tracker 3.0 traces the undergraduate origins of the world’s top AI researchers; the U.S. top-tier schools (MIT, Stanford, Berkeley, CMU) are, after China, the world’s most important supply source, and U.S. graduate schools concentrate the highest share of top AI talent.

The bottom 25% has slid back past 1992. The top end’s training-and-attraction machinery still leads the world. One country is folding two eras into itself.

Current Reforms

All four systems are reforming. The directions differ. None is finished. The earliest of their outputs reach the labour market sometime in the mid-2030s.

China timeline:

• July 2021, the “Double Reduction” policy lands; off-school subject tutoring is systematically suppressed.
• April 2022, the Ministry of Education issues a new compulsory-education curriculum framework, putting “core literacies” at the centre.
• December 2024, MoE issues guidance on strengthening AI education in primary and secondary schools.
• January 2025, the State Council issues the Education Strong Nation Plan (2024-2035), calling for “concentrated assessment of students’ key abilities, disciplinary literacies, and quality of thought,” and launching the “Fertile Soil Initiative” (science literacy for K-12) and the “Excellence Initiative” (top-tier innovative talent for high schoolers).
• From 2024, Shanghai introduces a mandatory AI Foundations curriculum for grades 4 and 7, one period a week, no fewer than 30 periods per grade per year, with the in-house “InnoSpark AI Education Model 1.0” available for any district school to plug into.
• Huangpu District’s “Climbing the Peak” program builds five subject-specific top-talent bases anchored in three flagship demonstration high schools (Datong, Gezhi, Xiangming), selecting no fewer than 300 students over three years; runs joint research-talent programs with Fudan, Jiao Tong, Tongji, and East China Normal that bridge high school and university.
• April 2026, five ministries (MoE, NDRC, MIIT, MoST, NDA) jointly issue the AI + Education Action Plan (Jiao Ke Xin [2026] No. 1), targeting by 2030 a vertically and horizontally connected AI curriculum across all education stages and a society-wide AI-literacy system — pushing AI education from local pilots to a national institutional deployment.

US timeline:

• From 2020, the “science of reading” reform is rolled out across multiple states, restoring systematic phonics instruction and early reading intervention.
• The 2024 NAEP shows zero state improvements in reading; multiple states pivot toward stronger foundational-skills assessment.
• April 2025, the White House issues an executive order on advancing AI education for American youth.
• From 2025, the U.S. Department of Education brings AI literacy and AI proficiency into federal-funding-eligible categories in multiple guidance documents.

Japan timeline:

• 2020, the new course of study takes effect; it introduces active learning, cross-disciplinary project-based learning, and mandatory primary-school programming.
• The GIGA School plan, 2020-2024, completes nationwide student devices and network deployment.
• December 2024, MEXT issues guidelines for using generative AI in primary and secondary schools.
• AI Education Accelerator Program launches; partnerships with industry expand teacher AI training.

Korea timeline:

• March 2025, AI digital textbooks are rolled out in math, English, information, and special-education Korean for grades 3-4 and the first year of middle and high school.
• July 2025, first-semester adoption rate is 37%.
• August 2025, the National Assembly downgrades AI digital textbooks from “core textbooks” to “supplementary materials”; school adoption becomes voluntary.
• Late 2025, newly elected President Lee Jae-myung rescinds the mandatory AI textbook policy; second-semester adoption rate falls to 19%.

Korea is among the most willing in the world to push AI education reform — and also the fastest to back away.

A Few Observations

Lay the data and the timelines side by side and a few things surface.

Student performance and adult performance don’t line up. Japan is strong at both layers. Finland is mostly strong at the adult layer. The U.S. is unimpressive at both. China’s developed regions are strong at the student layer; no comparable adult data exists. Korea is strong as a student, declining as an adult. You cannot read off adult ability from teenage scores.

Classroom-level differences are larger than assessment-level differences. German and U.S. 8th-grade math classes spend over 90% of the time practising procedures already learned; Japan spends 41%. PISA’s final scores do not reveal this, but it is the actual shape of basic-education delivery.

The U.S. internal split is the most severe of the four. The bottom 25% has fallen below its 1992 floor while the top end keeps shipping top-tier talent to the rest of the world. One country.

All four systems are reforming. The earliest reform outputs reach the labour market in the mid-2030s. The current generation of workers was produced by the old system.

The Core Problem

Four systems cover almost every shape “teaching and training” took before AI. Their measured profiles in PISA, PIAAC, the TIMSS Video Study, and NAEP are richer and stranger than the caricatures suggest. They are all reforming. The reforms point in different directions. The earliest of those reforms reach the labour market in the mid-2030s.

The current scene is here.

Education and training drill seven different cognitive mechanisms. Each has its own plasticity, its own fate after AI, and its own market scarcity. The four layers the old system poured its resources into — knowledge, the shallow part of activation frequency, the shallow part of the strategy repertoire, foundational automation — are precisely the ones depreciating after AI. The two layers the old system barely funded — cognitive decoupling and metacognition — are precisely the ones AI now prices on their own.

This mismatch is not the fault of the current workforce. The system produced them on its own logic, evaluated them by its own standards, sent them into the labour market — and then began to deform. Reform has started, and it is aimed at the children who will start school five years from now. The several generations who still have 20 to 40 years left to work in the labour market are not in scope.

Teaching and training have not lost meaning. The credentials they used to confer are coming undone.

Chapter 2 · Teaching, Disqualified

The four layers the old system invested most heavily in are precisely the ones depreciating after AI; the two layers it barely invested in are precisely the ones AI now prices on their own. The mismatch is structural. The premise behind the education system and the capability frontier of AI cover the same patch of ground.

The Source of the Mismatch

The premise of the old education was “produce identifiable people at scale.” A single national school faces hundreds; a country’s basic education system faces tens of millions to hundreds of millions. That order of magnitude forces any training inside the system to satisfy three things at once.

Standardisable. What gets taught has to be defined precisely in textbooks and curricula. Otherwise different regions and different teachers cannot deliver on a shared cadence. A gaokao math problem that students from Urumqi to Shanghai to a county-town high school can all attempt and have graded by the same standard is what this layer makes possible.

Examinable. How well it was taught has to be evaluated against an external standard. Otherwise tiers can’t stably rank, and credentials lose their use to the labour market. If “ability” amounts to “I know I get it” with no external signal anywhere, the entire credential system unravels.

Cheap to mass-replicate. Teacher costs cannot run so high that you cannot hire them. An ordinary middle-school teacher, with four years of teacher training, has to be able to deliver curriculum knowledge in a classroom on a shared cadence. That capacity has to be reproducible at scale. If only top-tier scholars could teach a given subject, that subject cannot be part of basic education.

Of the seven mechanisms, the layers that satisfy all three conditions are knowledge, the shallow part of activation frequency, the shallow part of the strategy repertoire, and foundational automation. They can be written into textbooks, posed as standardised exam questions, and delivered on a shared cadence by teachers of varying quality. That is why nearly all of the old system’s resources end up here.

The ones that do not satisfy the three are deep cognitive-decoupling training and metacognition.

Deep cognitive decoupling demands sustained engagement with structurally unfamiliar situations. Each “unfamiliar” is single-use. Once it gets standardised into a textbook, the unfamiliarity is gone, and training degenerates into standard responses. Institutionally, this is something you cannot mass-produce. The most effective parts of the Chinese problem-bank tradition — top schools’ Math Olympiad training, the math-major core at research universities — sit precisely outside the standardised-assessment frame, in small classes and supervisor-driven mentoring.

Metacognition demands the loop of real tasks plus real feedback, repeated. It can’t be tested, because “the extent to which you can monitor your own thinking” is not something a single test sheet can measure. It can’t be cheaply replicated either: providing real tasks and real feedback to one student requires someone above the student’s level to participate one-on-one for a long time. That cost is unsupportable inside scale education. K-12 has almost no slot for metacognition training.

This is the system’s internal boundary. The layers that scale, the system mass-produces. The layers that don’t scale fall outside the system’s coverage and get left to a small set of elite programs, master-apprentice arrangements, and individual self-study.

Up to here, the mismatch is just an interesting fact. After AI, it gets serious.

The essence of AI training is taking the standardisable, structurable corpus that human society laid down over centuries and turning it into output capacity that mass-replicates cheaply.

The standardisable parts — AI has all of them.

The examinable parts, i.e. the parts with model answers — AI gives them faster and more accurately than people.

The mass-replicable parts (knowledge, templates, formulas) — AI’s per-unit cost is essentially zero.

What AI covers and what the old education covers is the same class of “scalable cognitive output.” The relation between them is substitution, not complement.

The old education did not do anything wrong. What it did was mass-produce identifiable objects. AI is doing the same thing, cheaper, faster, with broader coverage.

This mismatch comes from the scale mechanism being used twice. Education used it for one to two centuries. AI used it for a few years. The two cover the same patch of ground, and AI’s coverage is more complete.

Question Banks, Disqualified

Problem-bank training works by pushing the student repeatedly to the working edge of his Gf: a situation he hasn’t seen before, where he has to identify structure, find an entry point, and construct a path with no ready-made solution. That is real on-the-spot reasoning. The harder he is pushed against that edge, the closer he gets to his usable Gf ceiling.

The mechanism only works if the problem is sufficiently “familiar” and sufficiently “unfamiliar”. A problem whose structure the student has never seen triggers on-the-spot reasoning, but the reasoning tools required have to be tools he has already learned. If the problem leans on totally foreign knowledge, he can’t solve it. If the same structure has been drilled for forty rounds, the unfamiliarity is gone. He stops firing up Gf and instead reaches into the strategy repertoire: identify the type, plug in the template, expand from memory.

After decades of grinding, the bulk of problem types have been exhaustively patternised. The high-intensity “problem-bank ability” of Chinese basic education is, in essence, this degraded form of training: students no longer doing on-site reasoning, but doing high-speed pattern recognition plus template recall.

This is something AI takes over without resistance. AI carries an enormous library of problem types and solution strategies, faster and more accurately than any drilling-prizewinning gaokao top scorer. A student who can get 95% of the gaokao math right is doing what AI does in one second.

This is not speculation. The author analysed the cognitive mechanisms of 2021-2025 gaokao math (567 problems across 7 distinct paper variants). Sorting by “extent to which Gf is invoked”:

“Drilling premium” is how much faster a student who has drilled 30 problems of that chapter is than one who hasn’t. 0 means parity; 1 means the drilled student finishes instantly while the un-drilled student basically can’t solve it. The category captures whether the problem actually pushes the student to the working edge of Gf.

By mechanism, 37.9% of problems test “foundational automation” (taking derivatives, plugging formulas), 33% test “deep schemas in the strategy repertoire” (3-D coordinate setup, conic-section + Vieta), totalling 71%. Genuine Gf problems are 18%.

A few counter-intuitive findings.

The Quan-Guo-Jia paper science section, four years (2021-2024), 90 problems total: zero GOOD. Including the so-called “anchor problems” — every step matches a standard weapon.

The drilling premium is highest on MIXED (0.50), lower on BAD (0.47), and lowest on GOOD (0.32). The places where “the drilled student finishes instantly while the un-drilled gets stuck” are the medium-difficulty conic-section, derivative, and probability problems — not the basic ones.

From 2024, the GOOD share jumps from 6-8% in earlier years to 14%. This reflects examiners adding “newly defined” problem types in the anchor slot — Beijing’s q21 is the marker. But 14% means 86 of every 100 problems can still be drilled into one-second answers.

This is the size of the question-bank disqualification at this exact moment in 2026.

The shape of question-bank disqualification: the mechanism itself is not wrong, but once problem types have been exhaustively patternised, training has shifted from “triggering judgment” to “training standard responses” — exactly the layer AI takes over.

Form-Production, Disqualified

The huge volume of “formal compliance” training the education system runs (writing, presenting, slide-making, report-writing, résumé-writing) used to serve as “indirect evidence of substance.” A student who could turn out a passable argumentative essay was, more often than not, someone who genuinely understood. A student who could write working code was, more often than not, someone with engineering chops. A student who could clearly articulate a project in interview was, more often than not, someone who had actually done it. Formal compliance was a low-cost signal for filtering trustworthy candidates.

After AI the threshold of formal compliance collapsed. A proposal that looks complete, code that runs, an article that reads structured, a résumé that looks neat — none of these reliably testifies to the corresponding subject behind it. ISSUE 04 named this “form being able to be produced independently of substance”.

Which means the “production capacity” that the education system has been training for decades has, as a filtering signal, sharply lost validity. A student who can write something passable and a student who can’t but knows how to use AI look the same from the outside.

Employers, graduate programs, scholarship committees all face the same problem: application essays, recommendation letters, research proposals, personal statements — the things they used to filter for “writes well, has ideas, has accumulation” — may today filter for “knows how to use ChatGPT.” The “essay craft” basic education trains depreciates faster against AI tools than even the question bank does. The question bank still requires that the student remember problem types; AI-writing requires no prerequisites.

A deeper layer: form-production disqualification doesn’t just affect filtering, it affects learning itself. A student trained from childhood to “build the essay structure first, then fill in content” is being trained on the implicit assumption “form carries substance.” In the act of training the form, he naturally deposits an understanding of the substance. After AI, that assumption breaks: the student can have AI produce the form directly, skipping the substance-deposition step. The internal feedback loop of the whole training mechanism is severed.

This is one of the education system’s most awkward positions in the age of AI: not only does production capacity depreciate, but “training substance through producing form” — a path that worked for centuries — also begins to fail.

The Standard Answer, Disqualified

The core move of an exam system is to standardise the answer. The point is to lower assessment cost so that scale assessment becomes possible. A model answer lets any reasonably trained grader produce nearly the same score.

This is the basis on which the education system has run for centuries. From the Sui imperial exam to the contemporary gaokao, an examination system that can be evaluated against a standard is the precondition for a credential to function as a social signal.

But what AI is best at producing is precisely the standard answer.

The closer the abilities the education system trains are to “can give the standard answer,” the more they testify that this is a thing AI already does. A gaokao top scorer can give an essentially perfect standard answer in the time allotted, but in the age of AI, that ability is no longer scarce.

A deeper layer: when an exam can score above 90 with AI, that exam is no longer measuring what it set out to measure. It is measuring “how good you are at using AI” or “how good you are at cheating.” That layer has come fully unmoored from the exam’s design goal.

In the past two years, university exam systems in both China and the U.S. have been confronting this. American universities widely report that ChatGPT can independently complete most undergraduate coursework. Chinese universities increasingly find “ghostwriting + AI tools” mixed-mode cheating. National responses go from banning AI, to AI detection, to switching exams to live oral defence or closed-book longhand — all of it, at root, an attempt to re-anchor the exam to “subject ability.”

But this only delays the problem; it doesn’t solve it. Standard-answer evidentiary power has failed in uncontrolled environments. Inside controlled exams it can still test foundational automation and parts of the strategy repertoire, but it can no longer independently certify that a person has full judgment.

Either the exam abandons the standard answer (and becomes open-ended, but expensive and weakly comparable), or it keeps the standard answer (but admits the exam now only tests what AI already excels at). Both roads end at the same fact: the credibility of examination as a filtering mechanism is breaking down.

This is deeper than question-bank disqualification or form-production disqualification, because it rocks the entire evaluation basis on which the education system runs.

Disqualification Squared

Exam disqualification is a first-order problem: the credibility of exams as filtering mechanisms is falling in the age of AI. The tutoring industry built around exams is a second-order problem: spending massive time and money to train students to perform on a filtering mechanism that has already been disqualified. The nature of that thing is disqualification squared.

The shape is plain in three of the four systems. After Double Reduction suppressed off-school subject tutoring in China, demand went underground, into “literacy” courses, and into AI-themed courses; total parental spend on tutoring is still enormous. A great deal of that tutoring exists to win the child a few more points on a gaokao mechanism that has already been disqualified, with no expectation that the child will actually master anything. American SAT prep, AP prep, and Common App personal-statement editing are a multi-billion-dollar industry; after AI, the evidentiary value of what they produce has fallen sharply, and the industry keeps growing. Korean hagwon are the extreme case of disqualification squared: send the child to cram school from age six and find that as adults, his foundational abilities have actually gone backwards (Korean adult literacy in PIAAC fell 23 points between rounds).

But disqualification squared will not be the steady state at the top.

Top families and top students are mostly already doing something different. American elite families send their children into programs that genuinely train cognitive decoupling: undergraduate research in research labs, competition math, debate, independent research projects — not SAT cramming. The output of those programs remains valid in the age of AI. High-knowledge families in Chinese first-tier cities are putting more time into top-tier-talent bases, Olympiad programs, and overseas summer-research projects, sidestepping scale-exam training. Nordic families never made exam tutoring the dominant mode anyway.

Two reasons the top withdraws. First, information advantage: the top sees earlier the difference between “playing along with a disqualified mechanism” and “training real ability.” Second, return on investment: for a high-Gf-ceiling child, putting training where it can’t be standardised (cognitive decoupling, metacognition) has the highest ROI, because that child has the latent capacity to make those layers thick.

But middle and long-tail families struggle to do this. The reason is that there is no substitute. A family in a third-tier city cannot find a local genuine research project, cannot find a mentor capable of high-intensity real feedback. The child’s only “ladder up” is still the gaokao. Parents put money into tutoring not because they trust it, but because there is nothing else to put it into. They know the value is going down.

This is what disqualification squared really costs. It doesn’t just worsen the situation of the tutored students in the age of AI; it widens the ability gap between top and non-top families further. The top can recognise the disqualified mechanism and withdraw, redirecting resources to training that still works. The non-top can recognise it but cannot withdraw.

At this layer, the disqualification of teaching has gone past education itself. It is redistributing the ability distribution of the next generation.

Four Failure Points

China. The strength of the Chinese system is high-density Gf activation and stable ranking. But the working edge of that strength is: the problem must be sufficiently “unfamiliar.” After decades of grinding, the unfamiliarity is gone, and training has degenerated into pattern recognition plus template recall. In the age of AI this amounts to “I’ll have AI do the drilling for me.”

The author’s 567-problem study, broken out by paper, shows the difference more sharply. The autonomous-paper provinces (Shanghai 14.4%, Beijing 13.6%, Zhejiang 6.8%) have a notably higher GOOD share than the unified new-gaokao papers (Paper I 8.7%, Paper II 5.3%). The Quan-Guo-Jia science paper, four years, 90 problems, has zero GOOD. The gaokao question bank has failed at “lifting Gf assessment”; the unified paper has pushed routinisation further. “New-context dressing” in the Chinese gaokao is mostly decoration; problems wrapped in Chang’e satellite signal transmission, basketball games, etc. are standard templates with the wrapper stripped away.

Recent reforms — Double Reduction, top-tier-talent programs, Shanghai’s mandatory AI curriculum, Huangpu’s Climbing the Peak — are all attempts to shift basic education from “standard-response training” to “unfamiliar-context reasoning.” The direction is right; coverage is limited. Climbing the Peak selects no more than 300 students over three years, an elite program inside one district, with no direct effect on the vast majority of national students.

The deeper problem: training for unfamiliar-context reasoning cannot itself be scaled. The Chinese system trying to keep its scale-filtering advantage while introducing un-scalable training is, structurally, a contradiction. There is no spreadable solution at present.

Japan. The strength of the Japanese system is the thickest foundational automation. The PIAAC numbers above tell the story: Japanese adults’ basic cognitive abilities don’t depreciate badly with age, because K-12 trains literacy, calculation, reading, attention, and rule-awareness to automation.

But Japan’s failure point is not in basic education; it is in the “embedding” assumption of corporate training: shape a person into a long-term carrier of a specific function. After AI, “specific function” is itself being rewritten. A person trained to carry a specific function for thirty years finds, when AI rewrites that function, that decades of “embedding capacity” have all gone to nothing.

At the basic-education layer, Japan’s failure point is the absence of an AI-use paradigm. MEXT’s 2024 generative-AI guidelines are a start, but they are far from forming an “AI-use paradigm in foundational automation.” Japanese basic education has not yet taken AI use into “trained to the point you don’t think about it.”

United States. American education has already split into three layers, each with a different fate in the age of AI.

The top end (AP, honours, research programs, top private schools, research universities) trains cognitive decoupling and metacognition at world-leading intensity. The overwhelming majority of top global AI researchers come from here. After AI this layer will be amplified further. The abilities they already have are precisely the ones AI prices on their own; with AI tools accelerating efficiency, the gap between the top end and everyone else widens.

The bell-curve interior. NAEP 2024 shows this layer in continued decline; the lowest percentile is at a 20-year low. States are trying to swing back to foundational basics — the “science of reading” reform is exactly that swing back. But returning to basics doesn’t help the current workforce; it helps children who will start school five years from now. The bell-interior workforce in the labour market today gets no relief.

The long-tail bottom. The bottom 25% has slid back past 1992. This layer will be further compressed in the age of AI. People without solid foundations using AI are outsourcing judgment to a system that cannot bear responsibility. The result is being held down by AI: they don’t have the ability to judge AI’s output, so they trust whatever AI gives them.

The most likely picture for the U.S. over the next ten years is the middle splitting in two: those who can complete a self-rebuild rise into the top end and accelerate; those who can’t slide into the bottom and get held down by AI. The bell-curve median, originally so populous, gradually disappears, and the social ability distribution moves from a bell curve to bimodal. This is the most dramatic disqualification picture among the four systems after AI.

The Nordics. The Nordic system’s strength is long-run maintenance of the population-wide floor. But the Nordics don’t train concentrated deep elite schemas, so at the student layer the right-tail explosive power isn’t there — and they don’t dominate the few extreme-output positions in the age of AI. That is the Nordic failure point: they can guarantee the floor doesn’t depreciate, but cannot guarantee a right-tail output. But the real opportunity for the Nordic system shows up in the age of AI. The mature lifelong-learning entry points keep Nordic adults from depreciating; if that mechanism extends to AI use, it amounts to extending the boundary of foundational automation into the age of AI.

Counter-examples: What Has Not Been Disqualified

The shape of training that remains fully valid after AI.

Real mathematics training: Math Olympiad, math undergraduate, math research. It trains the extreme form of cognitive decoupling: stripping abstract structure entirely out of context and operating on it as a formal object. Proving a theorem requires step-by-step derivation in a strict formal system. AI can verify; the act of training the human’s decoupling capacity itself can only be done by a human.

Top-tier research training: doctoral, postdoctoral, internships in research labs. It trains metacognition (knowing the boundaries of your own research direction, judging which problems matter, knowing when you’re deceiving yourself) and deep domain schemas (the “this solution looks wrong” judgment in the head of someone who actually understands physics). It depends entirely on real tasks plus real feedback; AI cannot do an advisor’s judgment.

Professional debate, moot court, debate club. Identifying argument structure under adversarial conditions, identifying hidden assumptions, constructing rebuttals. AI can simulate debate, but training people to call up judgment under pressure in live human-on-human exchange has to be done by real subjects.

Real engineering training. Engineering that requires debugging in complex systems, inferring anomalies, identifying hidden dependencies, judging when to rewrite an architecture rather than patch a bug. Engineering that is more than “code that runs”. This kind of training drills both deep domain schemas and metacognition.

Long-form mentorship: medical residency, traditional craft, clinical psychology. Training the kind of judgment that can only form through years of real feedback. A resident’s bedside judgment about a patient takes several years and several hundred real cases to form.

High-intensity writing training plus real feedback. The kind where you rewrite a draft thirty times a week — not the formality of a college writing class. The core of this kind of training is letting the student form, through repeated revision, his own judgment and his own voice — the channel through which “authorship,” in the sense ISSUE 04 discussed in its later half, is trained.

The shared characteristics of these forms:

Cannot be scale-taught (student counts have to be small).

Real tasks plus real feedback are required (assessment cannot be a stand-in).

The economics don’t add up (high teacher cost, low student conversion).

Coverage will inevitably be low (less than 0.1% of China’s 1.4 billion can plausibly be in such training).

This is the mirror of the earlier argument. The old education didn’t invest in these layers because they’re economically unattractive and can’t be done at scale. After AI, these are precisely the layers AI prices on their own.

The shapes of training that have not been disqualified share one property: none of them scale. The abilities the market prices on their own in the age of AI are exactly the abilities the education system has never covered.

Teaching’s Disqualification Is Not Inside Education

Teaching’s disqualification is, in essence, a structural mismatch between the education system’s premise (mass-produce identifiable objects) and the age-of-AI ability distribution (un-scalable judgment and decoupling). The disqualification is not in the teachers, not in the curricula, not in the exams.

This mismatch is the basic situation of education in the age of AI. It is not a one-shot problem to be corrected. Reform is aimed at the children starting school five years from now. The several generations on the labour market who were trained by the old system have to face the mismatch on their own.

Chapter 3 · Training, Disqualified

Teaching’s disqualification is in the content being taken over by AI. Training’s disqualification is somewhere else: training rests on three premises — the role is stable, the process is stable, the decision chain is stable — and after AI all three fail at once.

Training was not wrong in what it trained. It was disqualified because what it was training for is disappearing. This is a deeper disqualification than teaching’s. Teaching’s output has merely depreciated; training has lost its very premise.

Training’s Premises

The Preface told the Henry Ford story. The Highland Park moving line in 1913 compressed chassis assembly time from 12.5 hours to 93 minutes. What Ford solved was a specific problem: how do you take a group of farm-immigrants with no automobile-manufacturing background and shape them, in two weeks, into people who can stand at a single station and repeat one motion.

That is the origin of “training” as a modern institution. It is industrialisation’s particular way of shaping a person, distinct from education’s extension.

Training has three problems to solve.

First, turn a generic talent into a unit that can drop, immediately, into a specific role. A college graduate joining a company can’t produce value right away. He has to be trained into “this company, this role” — a part that runs steadily inside the operation. Role descriptions, SOPs, onboarding, shadowing, master-apprentice pairings, rotation — the core motion of all of this is that one thing.

Second, transmit experience between roles. The “what to do in what situation” judgment in a senior employee’s head cannot be fully written down. The organisation moves that tacit knowledge from older to newer employees through OJT, master-apprentice arrangements, long-term co-presence. This is the strong point of Japanese corporate training, and it is where Nonaka’s “tacit knowledge” actually applies.

Third, build stable expectations for a person inside the organisation’s process. “When I get this kind of request, this is how I handle it”; “when I report up, this is the format”; “when something breaks, this is who I find”; “this client gets handled this way.” These expectations carry the organisation’s culture, not the individual’s ability. The longer an employee is at a firm, the more organisational expectations he carries, and the more value he holds for the organisation.

Training’s outputs are role experience, process fluency, reporting skill, collaboration habits, organisational-culture carriage.

This system worked over the past several decades because three premises held simultaneously.

The role is stable. The shape of the role I am being trained into today will largely still be that shape next year.

The process is stable. The organisation’s way of working doesn’t get rewritten every year.

The decision chain is stable. Upstream-downstream relationships, who I report to, the boundary lines of division of labour — these are predictable.

Only when all three hold does training, as a long-term investment, make sense. An employee trained for five years to a certain capability is, by both the organisation and the employee, assumed to still hold that capability in year six, year ten, year twenty. If by year six the role is redefined and the capability fails, the five years of training become sunk cost. From the company’s view it’s money; from the employee’s view it’s five years of cognitive investment in a life.

Training rests on a stable world: stable roles, stable processes, stable decision chains. What AI changes is the possibility of all three holding at once.

Three Premises Failing at Once

Roles are no longer stable. A role is, at root, “a stable bundle of tasks.” A product manager’s role is the bundle of writing requirement docs, doing user research, drafting PRDs, tracking development, reporting up and down. An operations role is the bundle of data analysis, content production, process execution, KPI breakdown, daily ops. A middle manager’s role is the bundle of intake-requirements, break-down-tasks, hand-down-to-reports, track-progress, collect-results, report-up. Each role is a specific bundle that emerged from decades of organisational evolution. The bundle is stable on the precondition that the tasks inside collaborate — that pulling them out individually doesn’t get easy.

What AI changes is exactly that collaboration assumption. After AI, a large fraction of the tasks in the bundle can be done independently, and the cost of doing them independently approaches zero. Writing the requirement doc, synthesising user research, writing the PRD, doing baseline data analysis, generating templated content, clarifying requirements and producing task trees, tracking progress — every one of these AI can do. When 60-80% of the tasks in the bundle can be done independently by AI, the existence rationale of the bundle as a “role” evaporates.

The role is being deconstructed. It is now a collection of tasks that can be split apart, redistributed, and outsourced to AI at any time — not a stable cluster requiring long training to carry. People who were trained for ten, fifteen, twenty years on this kind of role carry the “carrying capacity” of a now-deconstructed bundle, not a stable capacity. That capacity has nowhere to land once the role is redefined.

Processes are no longer stable. An organisation’s processes are the specific paths that string those roles together. A request goes from PM to designer to engineer to QA to release. A client goes from sales to pre-sales to proposal to contract to delivery. A reimbursement goes from finance to audit to compliance to approval. Every process is a specific path ground out by several generations inside the organisation.

After AI, those paths can be rewritten. Three modes. The first is process compression: a path that needed five people now needs one with AI. The second is process recomposition: stages that once had to run sequentially can now run in parallel or merge with AI. The third is process disappearance: with AI changing the shape of the output, the process itself is no longer needed.

People trained on the old processes know every step’s specifics, who to find, how to interface. What they own is familiarity with one specific process, not a generic ability. Once the process is rewritten, that familiarity itself depreciates.

This is not separate from “the role being deconstructed.” As the role gets deconstructed, the process linking the roles also gets deconstructed. They happen together.

Decision chains are no longer stable. Every person in an organisation is a node on a decision chain: upstream hands me a task, I understand it, I break it down, I hand to downstream, I review the downstream’s output, I deliver to upstream. The default assumption of the old node is that upstream and downstream are people. My core value, as a middle node, is translation, coordination, and review between people.

After AI, that assumption fails. Downstream is no longer all people; it includes AI, processes, templates, agents. My value as a node is no longer “coordinate the people upstream and downstream” — it is “judge what work goes to AI, what goes to people, what I keep.”

This is a completely different value definition. The capability the old middle manager carried was “coordinate between people.” The capability the age-of-AI node carries is “redraw your own downstream boundary.” The training method, the judgment criteria, and the capability composition of the two have almost no overlap.

This sentence is the inverse of Henry Ford in the Preface. Ford turned generic people into useable members of an industrial organisation. The disqualified manager in the age of AI is one who is still using people to do what machines can already do. The age-of-AI manager has to redistribute work that used to go to people: let AI carry what machines can do, and let people carry what only people can do.

The role gets deconstructed, the process gets rewritten, the decision chain has to redraw its own boundary. Three things at once. Old training systems trained people in the capability of “running fluently on stable roles, stable processes, stable decision chains.” When the three premises fail at once, the capability itself does not fail, but the object the capability was carrying (the specific role) disappears.

This is what training-disqualification really is. It does not disqualify on content the way teaching does. It disqualifies on premise. Training was not wrong; what it trained for is disappearing.

Specific Forms of Disqualification

Fluency-training disqualified. The fastest-taken-over layer. Writing emails, summarising material, building tables, generating code, writing reports, making slides, taking minutes, organising research notes. This “fluency” used to be one of the core values of an employee. A person five years in could write reports faster, more cleanly, more uniformly than a person one year in.

After AI, that gap is gone. A one-year-in person who knows how to use AI writes reports faster and at higher quality than a five-year-in person who doesn’t. The five-year-in person still has value. He may still be thicker in judgment, in client relationships, in domain accumulation. But “writes reports fast and well-formatted” is no longer where his value comes from.

Mid-career employees thoroughly trained on fluency are in the most awkward position. They’ve worked ten, fifteen, twenty years; the layer they’re thickest at is precisely fluency. When fluency stops being priced, what they have to bring to the table is judgment, relationships, clients, domain understanding. None of which the organisation systematically trained — they accumulated those in fragments on their own.

Process-fitting training, disqualified. Old training shaped people into “stable carriers of a specific process.” That employee knew this SOP, that client, internal approval flow, the implicit rules of cross-team collaboration. After AI the process itself gets rewritten. The capability trained on a process — as “familiarity with one specific process” — depreciates with the process.

A deeper layer: when the process is unstable, “follow the process” itself is worth less. Organisations start to value “judging when not to follow the process” more — i.e., metacognition. Metacognition is precisely not what process training covers.

Experience-accumulation training, disqualified. The core reason “old employees are valuable” inside the old organisation was experience: more cases seen, knows what to do when something breaks, remembers the specifics of that client three years ago, anticipates what will go wrong in a given context.

Experience as a capability is layered. Shallow experience is “I’ve seen this kind of case.” That layer AI carries an enormous case library for, broader than any single person. Deep experience is “the judgment formed by feedback through huge numbers of real failures and successes.” That layer, AI cannot replicate, because it has no continuous subject.

The old training does not separate the two. An employee is often valued on “X years of work experience,” with the implicit assumption “time = experience = judgment.” After AI that equation breaks: shallow experience depreciates; deep experience stays scarce — but shallow is what 90% of employees are carrying.

A 15-year middle manager often says, “I’ve done this kind of thing 100 times.” Before AI, that was a proof of capability. After AI, it becomes awkward: AI has seen ten thousand times more cases than 100.

Standardised-delivery training, disqualified. Old training emphasised delivery standardisation: a complete plan, polished slides, professional report, formal email, structured presentation. Before AI this layer was the lubricant of the organisation. Formally compliant deliverables raised the efficiency of information flow. After AI the threshold of formal compliance is gone. This corresponds to “form-production disqualification” in teaching, but the scene is at work, not at school. A slick-looking deck no longer testifies that the person behind it understands the problem. A structurally complete report no longer testifies that the person did the analysis. The “professionalism” old training thoroughly shaped (writing emails, presenting, packaging plans) — the specific abilities that used to define “office worker” — has shifted, after AI, from competitive value to “stuff AI can produce.”

The Middle Has It Worst

Those four forms cover most employees. But one specific group has the harshest position in the age of AI: middle managers.

Middle managers are the largest “thoroughly trained by the old training” group on the labour market. Their disqualification mechanism, however, is not the same as front-line employees’. AI does not directly replace middle managers. It raises the standard of the middle manager’s core motion to a level the old-trained middle manager can’t reach.

The middle manager’s core motion is a loop. Any middle manager’s core motion can be broken into six steps: intake-requirements, break-down-tasks, hand-down-to-reports, track-progress, collect-results, report-up. Those six form a judgment-and-execution loop. The middle manager’s core value is not in the steps themselves, but in how fast, stable, and high-quality the loop runs. A good middle manager runs the loop fast, decides decisively, ships consistently; a bad one chokes the loop, decides indecisively, ships unevenly.

AI doesn’t replace the loop; it rewrites the standard of the loop. AI doesn’t replace the six steps. With AI, every step is faster, more accurate. At intake, AI helps clarify and rewrite the requirement. At breakdown, AI generates a structured task tree. At delegation, parts can go directly to AI; the rest goes to reports. At progress-tracking, automation does most of it. At results, AI does an initial quality check. At reporting, AI generates the various forms of reporting material. The whole loop runs noticeably faster and at higher quality. ISSUE 02 made the underlying point: AI lifts efficiency dramatically while keeping quality at or above the median. No company’s all-human team, no matter how it’s trained, can do that.

The all-human team’s quality ceiling. Any team made of people has a normally distributed quality, no matter how trained. The best output very high, the worst very low. The team’s overall output is held back by the weakest link. A 10-person team’s overall output quality is forever below that of the strongest person on the team.

AI raises every person’s work-quality baseline above the median. A one-year-in employee writing a report with AI can hit the median of five-year-in quality. A five-year-in employee with AI can hit the median of ten-year-in quality. The weakest link’s quality goes up sharply.

A middle manager who knows how to use AI runs a team whose total output quality stays above median. A middle manager who doesn’t runs a team that is still a normal distribution, dragged by its weakest link. Same resources, same headcount, same goal — the output gap between the two teams can be 2-3×.

Old training doesn’t include AI use. What middle managers were trained on for the past 10-20 years is router capability: intake, break-down, track, report-up. There is no “AI use” anywhere in that training, because the old training’s object was stable roles, stable processes, stable decision chains. AI use is a paradigm that emerged after 2023. Middle managers trained on the old path have good router capability but cannot use AI to rewrite the speed and quality of their own judgment-and-execution loop.

What “the middle has it worst” really means: the core motion he carries (the speed and quality of the judgment-and-execution loop) has had its age-of-AI standard raised dramatically, and the old training never taught him how to hit that new standard. Old-path middle managers, compared with AI-using middle managers, are sharply behind on both efficiency and quality. Same resources, same team, same goal, the AI-using middle is done while the old-path middle is still going. Or both are done, and the AI-using middle’s output quality is steady above median while the old-path middle’s is dragged by the weakest link. This is the real disqualification of the middle manager in the age of AI: old-trained middles are sharply behind AI-using middles in both efficiency and quality.

Four Reference Systems for Vocational Training

Japan. The strength of Japanese corporate training is foundational automation plus long-term embedding, shaping a person, through OJT, rotation, and master-apprentice arrangements, into someone who can carry a specific function for the long term. The failure point is that “specific function” itself is disappearing. A person trained to “carry a specific function for thirty years” finds, when AI rewrites that function, that the entire training has failed. Both ends of the long-term-employment assumption fail in Japanese organisations. One end: the firm no longer needs lifetime specialists. The other end: the employee’s specialism no longer fits new roles.

Japan’s failure is not in training poorly. It is that the direction Japanese firms train (deep embedding into a specific role) and the age-of-AI ability distribution (the role itself can be rewritten at any time) point opposite ways. The better Japan is at corporate training, the more brittle it is in the age of AI.

China. Chinese vocational training has a peculiar feature. Over the past 20 years it has concentrated on outputting “execution-type engineers”: people who can execute specific technical tasks, work 996, ship reliably. This layer is more brittle in the age of AI than Japan’s or America’s. The reason is that AI takes over precisely this kind of “standardised execution-type engineering”: writing code, debugging, integration, documentation. An engineer with AI can carry the workload of two or three engineers without it.

The job-demand for median execution-type engineers is likely to compress noticeably in the next few years. The impact of this in China is greater than in Japan or the U.S. because that layer’s share of the labour market in China is bigger. China’s other failure point is similar to Japan’s: the lifetime-employment assumption is also failing. But China’s failure happens faster, because there is no Japanese-style lifetime-employment culture as a buffer.

Other reference systems. American vocational training has always been fragmented, leaning on lateral mobility (job-hopping is normal), external certifications (PMP, CFA, AWS), and an apprenticeship revival. Fragmentation has an unexpected advantage in the age of AI: because Americans are not deeply embedded in specific firms the way the Japanese are, U.S. labour tolerates role change better. But the U.S. failure point is in the floor: people without solid foundations using AI in the age of AI are led around by AI rather than amplified by it.

The Nordics’ strength is mature lifelong-learning entry points. The failure point is small scale and small population covered, so it cannot be a transferable solution for other countries. But the mechanism itself still holds. If it extends to AI use (turn “working with AI” into a population-wide lifelong-learnable capability), it amounts to extending the boundary of foundational automation into the age of AI.

Korea is the extreme form of training disqualification. From hagwon filling all of a child’s time with exam training, to high-intensity employment preparation in university, to “presence-as-value” workplace culture — the entire chain treats high-intensity training as value itself. The result: Korean adult literacy in PIAAC 2023 is 23 points below the previous round. The most heavily trained generation of adults shows the largest decline. The whole chain is failing simultaneously, and the culture self-reinforces it.

Training Wasn’t Wrong

Training’s disqualification is not, in essence, a disqualification of training content. It is the simultaneous failure of training’s premises (stable roles, stable processes, stable decision chains) in the age of AI.

The several generations thoroughly shaped by old training carry the capability of “running fluently inside stable structures.” That capability is not failing; the object the capability was carrying is disappearing. When 20 years of work experience all corresponds to deconstructed roles, rewritten processes, and AI-taken-over decision chains, the person has to redraw his ability boundary. Old training did not teach him how to redraw.

Teaching’s disqualification means the children five years from now have to be re-taught. Training’s disqualification means the several generations on the labour market today have to be re-trained.

Teaching’s disqualification still has a reform path. Curricula can be redesigned, new exam systems built, new teaching methods spread. Results in 5-10 years.

Training’s disqualification has no such reform path. The reason: there is no “age-of-AI training” sitting around as a ready reference. Every organisation, every industry, every specific role has to find on its own what the new training is. The pace depends on AI’s capability frontier, the organisation’s reaction speed, and the individual’s self-rebuild capacity. Many people will be stuck in between old training and new demand.

The cognitive structure inside their heads was built for a world that no longer exists.

Chapter 4 · The Sandwich Generation

The diagnosis has to land on a specific group: the several generations who are currently 25 to 50, already out of K-12, with 20 to 40 years left on the labour market.

Teaching’s disqualification means they were already produced with a mismatch when the old system shipped them. Training’s disqualification means work then shaped them further into a form that is no longer needed. Reform has started, but reform is not for them.

Who They Are

By current age, sub-cohorts:

Each sub-cohort entered work at a different “pre-AI moment”. The 25-30s entered around 2020 and met ChatGPT essentially as soon as they started working. The 47-55s entered in the 1990s; their entire career’s cognitive mode was shaped for a different world.

Rough estimates by age structure, education, and labour-force participation: China has about 450 million people aged 25-50, of whom roughly 300 million have completed at least high school and are currently in the workforce. Globally there are about 2.7 billion people aged 25-50; of those in industrialised and semi-industrialised economies with formal education and formal jobs, about 1.2 to 1.5 billion.

By the seven mechanisms, what this group was strongly trained on is knowledge, the shallow part of activation frequency, the shallow part of the strategy repertoire, foundational automation, and formal compliance. What they were weakly trained on or not trained on at all is the depth of cognitive decoupling, metacognition, and the AI-use paradigm (which did not exist during their schooling and early career).

This ability structure was designed for a world of stable roles, stable processes, stable decision chains. Before AI it could fully sustain a career. After AI the support starts to fail.

They did not choose, back then, to be trained into who they are now.

A Temporal Mismatch

What the sandwich generation faces is not just an ability mismatch. It is a temporal mismatch.

Pace at which education reform reaches the labour market:

Pace at which AI is rewriting work:

Education reform runs on a 10-year clock. AI rewriting work runs on a monthly clock. The two timescales differ by two orders of magnitude.

The sandwich generation sits between the two timetables. They are not the beneficiaries of reform; the reforms reach them only after they retire. They are the absorbers of AI’s rewrite of work; that is happening every month.

Education reforms in all four systems share one feature: they target “the next generation of education-system output.” None of them target adults already on the labour market.

Adult retraining is not in the traditional ministry-of-education’s remit. It belongs to “vocational training” or “lifelong learning,” carried by employers, industry associations, individuals themselves. But corporate AI training stops at tool-skills, industry associations move slower than national education systems, and “individuals carry it themselves” means most people have no good resources.

The sandwich generation gets pushed off by three layers at once.

Three Forms of Deflection

Government punts to education reform. Education reform offers no direct help to the current workforce; it targets children of ten years from now. Government uses “we are reforming” as a response to the disqualification of the labour market, but that response is fully off-by-time. All four systems do this. China’s 2024-2035 Education Strong Nation Plan, the U.S.’s 2025 federal AI-education executive order, Japan’s MEXT AI guidelines, Korea’s 2025 AI digital textbooks. All target students at the basic-education stage. None targets adult-cognitive-rebuild for the working population. The sandwich generation’s relationship to education reform is, in essence, a non-coverage relationship. However well reform goes, it has nothing to do with them.

Companies punt to AI training. Most corporate AI training stops at “teach everyone how to write a prompt, how to use this tool, how to summarise documents.” That treats tool capacity as cognitive capacity.

What is actually needed is not tool capacity. It is task-redesign capacity: redraw the downstream boundary of your own work, judge what tasks go to AI, what go to people, what you keep. That demands deep metacognition; no tool training can reach it.

Companies don’t teach this, because it requires redesigning the organisation’s processes. Companies don’t know how to redesign them either. A company will not train its employees to “reinvent your job,” because the company itself doesn’t know what the job should be reinvented into.

The real corporate dilemma: AI requires every role boundary to be redrawn, but the company’s structure, KPIs, and reporting lines are designed around the old roles. Redrawing role boundaries equals redesigning the organisation. The cost and risk of that are too high. So companies pick the path of least resistance: teach employees to use AI tools, pretend that’s enough.

Individuals are punted to “lifelong learning.” A large part of the lifelong-learning market commodifies age-of-AI anxiety into paid courses. What those courses teach is still the knowledge layer and the tool layer — exactly the parts AI has already taken over. People who finish them feel anxiety relief temporarily; their way of working doesn’t change. That is the educational-market form of “rumination consumption” from ISSUE 01: pay, consume, psychological satisfaction, zero transfer. The deeper problem: real cognitive rebuild can’t be done through paid courses. It needs the loop of real tasks plus real feedback, repeated; it needs sustained one-on-one mentorship; it needs continuous metacognitive calibration. The cost is extreme, the coverage will inevitably be low, and it cannot be productised. The market looks like it’s solving the sandwich generation’s cognitive-rebuild need; what it’s actually doing is something else: monetising anxiety.

Anxiety and Despair

The sandwich generation is not uniform. Different ages, different roles, different accumulations — completely different positions. The three positions below cover the most representative sub-cohorts.

Entry-level text-work workers. 22-28, 0-5 years out of school. Work is text-shaped: emails, slide decks, reports, tables, research notes, baseline data analysis, standardised client communication.

Their entire value sits in “baseline execution capacity.” But they’re new on the job and have accumulated nothing yet: no domain repertoire, no work judgment, no internal familiarity with the process, no automation of any skill. “Baseline execution,” in essence, is “being walked through specific tasks by the organisation.” That is the standard onboarding entry point for a new hire.

After AI, that entry is disappearing. This isn’t in conflict with Chapter 3.3: there I said “a one-year-in person who knows AI writes faster and better than a five-year-in person who doesn’t.” That is a positive for the new hire individually — with AI, they can do an old hand’s work. But across the labour market, the picture flips: when new-hire-with-AI output equals old-hand-with-AI output equals AI-direct output, “the new hire can do the old hand’s work with AI” stops being scarcity. The labour market already has plenty of AI-using people, and AI-direct is also fine. The new hire has no comparative advantage.

Their entire output, AI can do. AI writes emails, AI makes slides, AI writes reports, AI organises research notes, AI does baseline analysis, AI drafts client emails — every one faster, more reliable, and more uniform than the new hire. AI’s marginal cost is in tokens, near zero. The new hire’s cost is salary, social insurance, training time, error tolerance. The company’s calculation is simple: rather than spend a year training a new hire to consistent delivery, just have AI do that year’s output directly.

The real consequence is not immediate unemployment. It is that the “accumulation entry” the new hire was supposed to earn in the first few years is gone. Earlier generations accumulated, through those early baseline-execution years, a domain repertoire, process familiarity, work judgment — and so moved up to middle and senior. This generation has no such entry. The baseline-execution work has gone to AI; the judgment work, they are not yet entitled to. They cannot “do the grunt work to earn their stripes” the way previous generations did, and they have not earned the cognitive-rebuild ability that the age of AI demands.

Their position is the hardest to resolve in the sandwich generation. They were not thoroughly enough shaped by old training (they hadn’t yet accumulated on the job), and they missed the accumulation entry (AI erased it).

The lived experience: a job won after sending out hundreds of résumés, a few months in, you realise the reports you write AI can write too. You ask a senior; the senior says “you need accumulation,” but cannot tell you exactly what “accumulation” is. AI-tool tutorials roll out in waves; finishing them eases anxiety briefly, but does not produce, in any real sense, accumulation.

Middle managers, 37-47. 15-25 years on the job, already moved from specialist to middle management. The core capabilities they were trained on are the running of the judgment-and-execution loop, cross-team coordination, reporting up, managing down.

After AI, the core motion they carry (the judgment-and-execution loop) has had its age-of-AI standard sharply raised, and old training contains no AI use. Old-path middle managers, compared to AI-using middle managers, lag noticeably on both efficiency and quality.

Lived experience: they realise the reports their younger reports can produce with AI are several times their own throughput. They start trying to use AI, but the use is awkward, unergonomic, slower than the younger reports. Some refuse, insisting “judgment can’t be handed to AI.” Some force themselves to learn but learn slowly. Some hang back and wait for the company to give a clear direction.

Their position has been the most stable in the organisation for the last decade and may be the least stable for the next.

Senior specialists, 47-55. 25-33 years on the job, with relatively thick domain-deep schemas. May be a sub-domain expert, a technical anchor, an industry veteran.

Compared to the previous two groups, they hold one real piece of capital: domain-deep schemas. That layer AI cannot take over. They also hold one real disadvantage: the AI-use gap relative to younger people. A 50-year-old specialist who doesn’t use AI still has valuable domain-deep schemas, but his output speed and breadth are sharply limited. A 30-year-old comfortable with AI may close the gap to within an uncomfortable distance in three years.

They have 10-18 years to retirement. Enough time to complete a self-rebuild, but it requires actively choosing to rebuild. Most will choose “work the last few years in the way I’m used to.” That choice isn’t itself wrong, but it means this generation’s domain-deep schemas leave with them at retirement, in bulk, without being passed down. That is the hidden cost of organisations in the age of AI.

Where the Responsibility Lands

The sandwich generation is not a forgotten group. They are too large, too loud, too consequential to the labour market and consumer market for anyone to actually forget them.

But they are a deflected group.

The education system isn’t designed for them. They’ve already left the education system; reform targets the system’s next-generation output.

Companies aren’t designed for them. Corporate AI training stops at the tool layer; it does not solve cognitive rebuild. Companies cannot bear the cost and risk of redesigning their processes.

The lifelong-learning market isn’t designed for them. It commodifies anxiety and treats consumption as the solution.

The cognitive rebuild of the sandwich generation lands, at the end, on themselves.

The sandwich generation’s cognitive structure was built for a world that no longer exists. That is not their fault, but the responsibility for repairing the mismatch lands on them.

Chapter 5 · Evolution

The seven mechanisms themselves do not disappear. What gets disqualified is the way of training them, not the mechanisms.

How the Seven Mechanisms Evolve

Sort training into three classes.

Class A: standardisable training. Foundational knowledge, rules, ethics, corporate culture, the AI-use paradigm, AI knowledge retrieval, general metacognition. Mass-produced; covers everyone.

Class B: standardisable tools that activate non-standardisable abilities. Problem banks, basic logic. Question authoring, scoring, and coverage are standardisable; the abilities activated (Gf invocation, deep cognitive decoupling) are not.

Class C: not standardisable at all. Domain-deep schemas, in-domain metacognition. Has to be done by the individual; AI is a tool, not a substitute.

Class A and B’s training objects in the age of AI take new shapes. The education system and corporate training both need to rebuild around the new shapes. Class C training is forever individual responsibility. AI changes the tools, not the responsible party.

Foundation, AI Use, AI Knowledge Retrieval, General Metacognition

Three things in knowledge training.

First, foundational knowledge is still required. Someone who doesn’t know Newton’s three laws will believe whatever AI tells him; someone without basic economics cannot judge AI’s policy analysis. This layer can’t be abandoned just because AI carries the knowledge.

Second, the boundary of knowledge. Under what conditions does this knowledge hold; under what conditions does it not. This is the real upgrade for knowledge training in the age of AI. Old training taught knowledge itself; new training simultaneously teaches the conditions of applicability. Someone who knows “X holds under Y” makes better judgments with AI than someone who only remembers “X.”

Third, how to retrieve knowledge through AI. Before AI, knowledge retrieval was a personal skill (search engines, literature databases, professional databases). After AI it becomes a structurable, trainable capability. How to ask so AI returns useful information, how to identify suspicious parts of AI’s output, how to have AI locate the actual primary sources, how to handle cross-domain knowledge connections with AI’s help. In 2026 this is still new; there are no mature curricula. But it will be standard in basic education and vocational training of the future.

Rules, ethics, law, corporate culture. This class of training was needed in the past and is needed now. Employee handbooks, compliance training, corporate culture, social rules. They decide whether AI tools can be used legally and ethically. Not unpacked here.

The AI-use paradigm as new foundational automation. Extend the old “literacy, calculation, reading until you don’t think about it” to “AI calls, AI workflows until you don’t think about them.” Borrow from the Japanese paradigm: high-frequency, broadly applicable, standardisable, memorisation plus practice. What needs to be drilled to automation: knowing what tasks go to AI and what don’t; knowing how to ask AI (structured, context-clear, demand-clear); knowing how to read AI’s output and decide right away whether it’s true; knowing how to use AI to calibrate judgment rather than substitute it; a set of AI-collaboration templates for common tasks; a set of counter-examples (when AI is wrong). The training form: a lot of repetition, standard templates, practice feedback, cross-task transfer. This doesn’t require a high Gf; it requires high repetition plus standardised practice — closely parallel to Japanese vocational on-the-job learning.

General metacognition can be trained. Metacognition isn’t unteachable. It is layered.

General metacognition can be mass-trained: knowledge about how to retrieve knowledge, knowledge about how to ask questions, knowledge about how to judge structure, knowledge about how to do post-mortems. All of these are teachable, drillable, examinable. Their training shape is clear, their evaluation criteria are clear, and there’s a defined path to improvement.

Some people have been teaching this layer before AI: information-literacy classes, critical-thinking classes, research-methods classes. But coverage was narrow, because it relied on high-quality teachers, and high-quality teachers are forever a scarce resource.

After AI this changes. AI is both a training tool and a feedback tool. After learning general metacognition, you can immediately calibrate against AI: is the question you asked good, is the follow-up direction right, is the judgment structure clear — AI can give immediate feedback. This is a training loop that didn’t exist before AI.

Domain metacognition remains individual responsibility. “When could I be wrong in this specific domain”; “which step of reasoning needs another check”; “where am I deceiving myself in this task.” This layer depends on real tasks and real failures in a specific domain; you cannot train it independent of the domain.

Curricula come from AI. Traditional curricula update on multi-year cycles. AI tools and paradigms update on multi-month cycles. Any frozen curriculum will lag immediately. The solution: curricula come from AI. Use AI to generate practice content, cases, counter-examples, and evaluation rubrics for the current AI tools. Let curricula update in real time alongside AI’s own evolution. If AI knowledge progress slows in the future (a maturity phase), one can return to frozen curricula. If it keeps accelerating, real-time curriculum updates are mandatory.

AI lets the high-cognition teacher’s coverage scale without limit. The premise of the education system is scale. Scale demands three things: standardisable, examinable, cheap to mass-replicate.

Training people with people has a fundamental problem: people are normally distributed in quality. No teacher-training will guarantee every teacher is above the median. A student in a third-tier city may meet a Chinese-language teacher whose level is below the 50th percentile nationwide. That student’s training quality is bounded by that teacher’s level.

What AI changes is that “AI helps one high-cognition teacher scale his teaching forms without limit.” A genuinely high-level teacher used to be able to reach the few classes and few hundred students he could teach in person. His teaching method, judgment, and understanding of the discipline could not be copied to other teachers. His cognition was a scarce resource, geometrically bounded.

After AI this changes. A high-cognition teacher can have AI learn his teaching methods, judgment criteria, and problem-design philosophy, and then have AI deliver that method, in the form of curricula, problem sets, feedback, and Q&A, to any student. AI here is not the default-level output (default AI is median or slightly above) — it is an amplifier of that teacher’s cognition. The training a student touches via AI can run far above AI’s default level.

This is the real opportunity for education in the age of AI: for the first time, the best teacher’s reach extends from a few hundred to several tens of millions. What is needed: high-cognition teachers willing to structure their methods so AI can learn them, a tool chain that supports it, and a channel for students to access “the AI shaped by the high-cognition teacher.” This is still early in 2026. But the direction is clear.

Question Banks Are Still Necessary

The real value of the problem bank. The argument earlier: good problems manufacture structural unfamiliarity and force students into the working state of Gf. The argument under “question-bank disqualification” was: once problem types are exhaustively patternised, training degenerates into standard responses.

The problem bank itself is not disqualified; what is disqualified is patternised problems. A structurally unfamiliar problem still activates Gf invocation, cognitive decoupling, and the deep layers of the strategy repertoire even in the age of AI. It is a standardisable trigger of non-standardisable abilities. The problem bank is “a standardisable tool that activates a non-standardisable ability.” Authoring, scoring, and coverage are standardisable; the abilities triggered (on-site reasoning, deep schemas) are not. That is why the problem bank is still necessary in the age of AI.

The strength of the Chinese system. Chinese basic education has, over 30 years, built the world’s thickest problem-bank training apparatus. This system shouldn’t be discarded in the age of AI; it should be reorganised. The problem with the old system: after exhaustive patternisation, training degenerates into standard responses. The new direction: keep the problem-bank training mechanism, refresh the contents. Keep large numbers of problems structurally unfamiliar.

Use AI to author problems. The method: generate new problems with AI. Update the problem bank yearly or even monthly. Let the bank’s iteration speed match the student’s drilling speed. Let “drilled-it” become an unreachable state. Before AI this was impossible. Authoring speed was constrained by human teachers; producing one good set took years. After AI it is possible, because AI can mass-generate new problems at low cost.

But there is a guardrail. The key isn’t surface novelty; it is structural novelty. AI can easily generate “surface-novel, structure-old” pseudo-unfamiliar problems. AI authoring has to layer in structural deduplication, difficulty calibration, and human spot-checks; otherwise it’s just old problems in new wrappers.

Problem-bank training plus AI auto-authoring equals the real upgrade path for the Chinese system. The upgrade preserves China’s strength in problem-bank training and removes the failure point.

A fundamental change in how to assess. The bank’s upgrade is only half. The other half is the assessment.

Old assessment: test whether the student can independently produce the standard answer. That assessment fails after AI (standard answers AI gives all of).

New assessment: assess how the student uses AI prompting to solve a problem at a cognitive boundary. Audit the student’s (or in-job trainee’s) full conversation with AI. The audit dimensions: what was the initial question; what judgment did he make after the first AI reply; what was the direction of his follow-up; how did he verify AI’s output; where did he identify AI’s error or limit; how did the final conclusion form.

What this assessment measures is judgment, decision, depth — the application of cognitive decoupling and metacognition. It is harder to substitute with AI than answers, because what it scores is process and judgment chain, not answer and result.

The assessment has to live in a controlled environment: complete logs, random follow-ups, on-site oral defence, version records, restricted external tools, required explanation of every judgment-pivot. Otherwise the student can have AI generate a plausible prompt trace; the process itself can be faked.

This isn’t a from-scratch imagination. AI competency frameworks, AI Assessment Scales, AI literacy rubrics, and the return of oral defence have all surfaced. They all point at the same thing: assessment in the age of AI cannot only look at the final answer; it has to look at how the person uses, constrains, questions, and verifies AI.

What’s missing is the synthesis of those scattered practices into a coherent process audit: audit the student’s initial question, the judgment after the first reply, the follow-up direction, the verification, the error identification, the conclusion formation, and the ability to explain every judgment-pivot live.

If Chinese basic education and vocational training are to do a real upgrade, the direction is: refresh problem-bank content via AI in real time; shift assessment from “independently give the standard answer” to “audit AI-collaboration process.”

Basic logic training. Mathematics, formal logic, basic argumentation. This layer is the foundation of cognitive decoupling. Chinese basic education has invested heavily here. The new direction is to keep the training shape “unfamiliar.” Keep the student in the state of “must reason on the spot” each time he faces a problem. Math Olympiad, formal logic, geometry proofs — programs once considered elite — should return to basic education in the age of AI. They don’t lift the Gf ceiling, but they activate Gf frequency and cognitive decoupling, which are the scarcest layers after AI.

Possibility for global borrowing. If the Chinese problem-bank apparatus completes the upgrade (bank iteration plus assessment change), it can become a real reference for other countries. Japanese and Nordic basic education don’t train Gf invocation and cognitive-decoupling activation as intensely as China. The American top end does; everything below the median doesn’t. If this upgrade succeeds, Chinese basic education for the first time has the chance to be a globally borrowable model. The premise is doing the upgrade actively, not waiting for reform to arrive on its own.

Corporate Know-how Distillation

Training disqualification has no ready-made reform path because there is no “age-of-AI training” sitting around as a reference.

What companies are doing now. Many leading firms are already distilling workflow and role know-how. They use AI to document, structure, and make callable: workflows, role responsibilities, internal decision paths, expert experience, domain know-how. Concrete forms include: capturing real workflows via process mining / task mining; piping internal wikis, emails, meeting notes, project docs, and decision records into corporate AI; having AI synthesise from case libraries “in this situation, use this approach”; distilling reusable rules from high-frequency situations in customer service, sales, IT, audit, operations; in some cases, recording experts’ or senior employees’ actual operational steps to make tacit processes and judgment explicit.

This work, between 2024 and 2026, has moved from experiments to platform deployments at leading firms. It is not yet mature everywhere, but the direction is clear: companies are turning the know-how that used to live scattered across employees, processes, documents, and systems into AI-callable organisational context.

The current goal-setting is still narrow. In the dominant corporate narrative, this work is positioned for two goals. First, efficiency lift: let employees use AI to finish current work faster, treat the distilled knowledge base as assistant, search entry, agent context, or automation material. Second, AI-use management: measure whether employees are using AI, how deeply, whether adoption rate and ROI are improving.

Both goals stand, but they are still optimisations from the current view. The deeper layer is that companies have not fully treated those distilled workflows and know-how as the foundation of a future training system. What really matters is not “let today’s employees finish work faster,” it is: when roles, processes, and decision chains are all rewritten by AI, how does the company turn the organisation’s most scarce judgments, paths, experiences, and tacit rules into training materials that new hires can learn, old hands can transfer, and teams can reuse.

The real meaning of this is not in the knowledge base, it is in the training source.

Recall the earlier point: AI extends the high-cognition teacher’s reach. Inside a company, “high-cognition teacher” isn’t necessarily a teacher — it is the company’s best middle managers, most senior experts, the old hands with the most accumulated know-how. Their judgment methods, experience, and decision criteria could only be transmitted through master-apprentice and long co-presence one-on-one or in small groups before. That is why Japanese corporate training works: tacit knowledge deposits over years of co-presence. It is also why the Japanese model can’t scale: one-on-one transmission has too low coverage.

Distilled into AI, this changes. What a new employee touches via AI is no longer just AI’s default output, but output shaped by the company’s best people, most effective processes, and longest-accumulated cases. What a middle manager learns via AI is no longer just generic management knowledge, but the specific judgment frameworks, decision criteria, and process logic of this firm.

That mechanism pushes AI from efficiency tool to training tool: AI’s default is generic median output; corporate know-how distillation calibrates it to high-quality experience inside the organisation. Originally the argument was AI helping employees work; now we can also say AI helping employees be trained.

The real corporate-training transformation. Three moves.

First, treat know-how distillation as the core of the training system, not just a current efficiency tool. The distilled material has to be designed against the standard of “use AI to train new hires”: structured enough that AI can understand it, can use it to author problems, can use it to assess learners.

Second, shift training from classroom lecture to AI-assisted on-the-job practice. New hires, on entering the role, go straight into “AI-assisted real-task” mode and skip the week of classroom training. AI tutors them, corrects them, sets them problems, and audits their collaboration process based on the company’s distilled know-how. This mechanism lets new hires reach in three months what the old system needed two years to produce.

Third, shift assessment from “scoring outcomes” to “scoring AI-collaboration processes.” Same as the assessment reform above. Audit the new hire’s conversation with AI: question quality, follow-up direction, judgment depth, ability to identify AI errors, formation of the final conclusion. These dimensions reflect real judgment.

Where most companies still are. Most are still at “teach employees to use AI tools.” A small set of leading firms are doing know-how distillation, but the goal is still current efficiency, not future training. Companies that treat distillation as the training base — at this moment in 2026 — are essentially none.

This will happen, because the logic is unavoidable. Three factors gate the speed.

First, the quality of distilled know-how depends on the company’s pre-existing documentation level. A company with a messy internal knowledge base produces messy AI training material.

Second, whether senior management recognises this as the training base. Most companies still treat it as an efficiency tool.

Third, whether employees cooperate. A senior employee who fears AI learning his know-how then making him replaceable will not genuinely cooperate with distillation.

These three gate the speed of corporate-training transformation. It will happen, but slower than the technology can move. Until companies complete this upgrade, the sandwich generation has to rebuild on its own.

The Non-Standardisable Layer: Domain-Deep Schemas and Domain Metacognition

Class C training will never scale. As argued: domain-deep schemas need long real tasks plus real failures in repeated feedback; in-domain metacognition cannot be assessed, it happens inside each person’s own judgment process. These two layers are forever individual responsibility.

The shape of AI-assisted self-training. Picking up the moves ISSUE 01 and ISSUE 02 already discussed: have a refutable initial judgment first; use AI to find counter-examples, do reviews, decompose variables; revise the judgment model under refutation; abstract experience into structurable rules; reconnect learning to real tasks.

Self-training for each mechanism.

Gf and its activation frequency: set yourself daily small tasks that “must be solved by reasoning.” Don’t let yourself ask AI by reflex; think for 5 minutes first. AI’s role here is counter-example provider. This isn’t just mechanically lifting Gf’s activation frequency; pushing Gf invocation repeatedly also pushes you closer to your Gf ceiling.

The depth of cognitive decoupling: when you have a judgment about a specific problem, ask AI to “abstract the core structure of this judgment” and check whether that abstract structure applies in other domains. This is the “cross-domain isomorphism” training from ISSUE 01.

Domain metacognition: after every important judgment, have AI run a post-mortem with you. What are the key assumptions of this judgment, what evidence supports it, what evidence opposes it, how would the judgment shift if the evidence changed.

Domain-deep schemas: depend on real tasks and real failures. AI cannot replace the failure itself, but it can speed up reflection. After each completed real task, have AI walk you through “if I redid it, which step would change.”

AI is a calibration tool, not a substitute. The lived experience: you have a judgment on a problem; you ask AI to refute it. Once AI gives the refutation, you decide whether the refutation holds. If it holds, revise the judgment; if not, argue why not. Each pass through that loop trains in-domain metacognition. Cost: AI’s token cost is very low, but your time cost is not. Each post-mortem costs 30 minutes to a few hours. This cannot be scaled. It also doesn’t need to be scaled. It is a series of specific conversations between an individual and AI — the individual’s own cognitive rebuild.

Who Can Do This

One concrete baseline picture: the person who scores 70-80 without drilling. Everyone has met this kind of person. In secondary school and college, doesn’t drill problems, doesn’t pay attention in class, remembers a limited number of formulas and methods, and yet on test day pulls 70-80.

Three things in their cognitive setup.

Higher Gf ceiling. Faced with the same structurally unfamiliar problem, he doesn’t need to have seen the type before to reason it out on the spot.

A small but high-quality strategy repertoire. The “few formulas and methods” he remembers are deeply internalised core structures, not shallow tricks. He uses domain-deep schemas: knows what to use when and what to avoid.

Stronger in-domain metacognition. He knows when he’s deceiving himself, knows which step of reasoning needs another check, knows how to allocate exam time.

These three are precisely the abilities the age of AI prices on their own. This kind of person will be in a far better position in the age of AI than someone who reaches the same score by drilling.

But there are two boundaries.

First boundary: the transfer from exam scenes to real work scenes is not automatic. Exam scenes are structured with model answers; real work is unstructured and you have to define your own metrics. Some people who scored high on Gf in exams stay in “wait for the task to be assigned” mode at work.

Second boundary: this kind of person is still rare at the top of the score distribution. Most gaokao top scorers and recommended-admission students are heavy-drilling-plus-high-Gf combinations. “Low-drill, high-score” is statistically a minority.

A more precise judgment: in the age of AI, the market value of the “low-drill plus mid-to-high score” cognitive mode will be sharply higher than the “heavy-drill plus high-score” combination. The former is close to the cognitive shape the age of AI prices on its own; the latter is precisely what AI takes over. But to genuinely stand out in a career, the former still has to actively transfer the cognitive mode to real tasks after leaving the exam system. That doesn’t happen automatically.

The threshold isn’t age, credential, or industry. Inside the sandwich generation, the difference between those who can complete a self-rebuild and those who can’t is not age, not credential, not industry. A 50-year-old engineer may complete it; a 30-year-old PM may fail.

The threshold is five things.

A. Can produce a refutable initial judgment. B. Can revise the model from being refuted. C. Can abstract experience into rules. D. Uses AI to find counter-examples, do reviews, run post-mortems. E. Can reconnect learning to real tasks.

None is hard alone. Together they take long persistence and real pain.

The “low-drill, score-70-80” person naturally has some of A, B, C. D and E need active cultivation. The AI tooling is new; reconnecting learning to real tasks is everyone’s individual choice.

A relatively positive judgment. The number of sandwich-generation people who can complete the non-standardisable (Class C) self-rebuild is small. That is unavoidable, because that layer will never scale.

But Class A and B training can, in the age of AI, hit high coverage at stable quality. Which means: as long as the Class A and B AI turn happens, large numbers of “basic-usable + with basic metacognition” people will be produced. They may not have completed deep self-rebuild, but they can complete basic AI use, hit above-median work output with AI assistance, and keep basic judgment. Those people still have a place in the labour market.

The premise is that the Class A and B turn actually happens. AI-use paradigm as new foundational automation, AI knowledge retrieval, general-metacognition training, problem-bank-plus-AI-auto-authoring-plus-AI-collaboration-process-audit, AI extending the high-cognition teacher’s reach, corporate know-how distillation. All of these turns are early in 2026, with no mature practice yet, but the path is clear.

If those turns succeed, the age of AI will not produce “most people can’t make a living.” What will appear is layering: a small group of deep-rebuilders at the top, a large group of basic-turners stable above the median.

The premise is that AGI hasn’t arrived. If AGI arrives in 5-10 years, all of these paths need rewriting. The AGI discussion isn’t unpacked here.

Education’s Lag, and the Scale of This

Education systems are, by design, lag products. The people they train enter the labour market 5-10 years later, on the assumption that the labour market is relatively stable in the meantime.

That assumption held for the past 100 years. The age of AI breaks it for the first time. AI rewrites work on a monthly clock; any 5-10-year curriculum is already out of date when its graduates reach the labour market.

Every wave of reform arrives as a solution to the previous era.

The concrete meaning for the sandwich generation: don’t wait.

Don’t wait for government reform (10-year cycle, doesn’t target working adults). Don’t wait for corporate training (stuck at the tool layer). Don’t wait for market products (anxiety, commodified).

The only thing that can start now is AI-assisted self-training. It doesn’t depend on any external system being ready.

Back to the Preface. Yang Jian wanted usable officials. Frederick wanted usable subjects. Ford wanted usable workers. Every time teaching and training were established, they were answering “can this person be used.”

April 2026. A 35-year-old sits at his desk and watches an AI agent finish what he had spent ten years getting fluent at. He knows the abilities he was trained on are no longer the abilities the market prices on their own.

This generation has not been given a new map. The map they were given describes a terrain that no longer exists.

Teaching and training have not lost meaning. They have lost the credentials they used to confer.

Some things remain necessary inside the old credentials: foundational knowledge, rules, ethics, basic logic, problem-bank activation. Some things have to be built new: the AI-use paradigm, AI knowledge retrieval, general-metacognition training, AI-collaboration process auditing, AI-assisted self-training, corporate know-how distillation.

This rebuild will happen. It will be slower than the technology, and far slower than the sandwich generation can bear.

For today, do today’s work. Tomorrow’s work, tomorrow’s people will do.

Sources and Citations

The principal factual claims in this essay rely on the sources below.

On the three historical pivots in the Preface

• Yang Jian (Emperor Wen of Sui), in the seventh year of Kaihuang (587 CE), issued a decree instructing each prefecture to “annually present three” candidates for the xiucai examination. Sui Shu, Annals of Gaozu. There is long-standing scholarly debate over the start year of the imperial examination; some scholars place its formal establishment in the Daye era of Emperor Yang of Sui (605 onward), with the jinshi category. This essay takes the 587 “annual presentation of three” as the proto-form of subject-based recruitment — the conventional narrative.
• Frederick II (“the Great”) issued the Generallandschulreglement (General School Regulations) on 12 August 1763 — a key text in the institutionalisation of Prussian state-mandated schooling: James Van Horn Melton, Absolutism and the Eighteenth-Century Origins of Compulsory Schooling in Prussia and Austria (Cambridge University Press, 1988). Note: Frederick William I had already issued an early compulsory-schooling edict in Prussia in 1717; the Massachusetts Bay Colony had a still-earlier compulsory-education act in 1642. The 1763 regulation is not “the world’s first national-level compulsory-education law” but a landmark text in the early institutionalisation of state compulsory schooling.
• Henry Ford’s 1913 Highland Park moving assembly line compressed automotive assembly time from 12.5 hours to 93 minutes; the $5/day wage policy was announced in January 1914; the Sociological Department (50 home-visit investigators) was formed in 1914; the Ford English School taught English and American civics. Sources: The Henry Ford digital archives (Sociological Department & English School collections); Stephen Meyer III, The Five Dollar Day: Labor Management and Social Control in the Ford Motor Company 1908-1921 (SUNY Press, 1981). The founding year of the Ford English School appears as both 1913 and 1914 in different archival entries; this essay uses the integrated narrative of “1913 line + 1914 Sociological Department expansion.”

On the seven cognitive mechanisms (Chapter 1)

• The fluid (Gf) / crystallised (Gc) intelligence split: Raymond B. Cattell, “Theory of Fluid and Crystallized Intelligence: A Critical Experiment,” Journal of Educational Psychology 54 (1963): 1-22, and the Cattell-Horn-Carroll lineage that followed.
• Tacit knowledge (anmokuchi): Michael Polanyi, The Tacit Dimension (Routledge, 1966) is the source; Ikujiro Nonaka’s application in corporate knowledge management appears in Nonaka & Takeuchi, The Knowledge-Creating Company (Oxford University Press, 1995).
• The “cognitive decoupling” framework comes from Offbook Press ISSUE 01, On Cognitive Decoupling. “Activation frequency,” “strategy repertoire,” “foundational automation,” and “metacognition” are organising labels this essay uses to integrate related concepts in educational psychology; they do not map exactly to a single academic theory.

On the international assessments (Chapter 1)

• PISA 2022: OECD (2023), PISA 2022 Results (Volume I): The State of Learning and Equity in Education, OECD Publishing, Paris, https://doi.org/10.1787/53f23881-en. Mainland China did not participate in 2022. The mainland-China B-S-J-Z (Beijing, Shanghai, Jiangsu, Zhejiang) figures of 591 / 555 / 590 used here are from PISA 2018: OECD (2019), PISA 2018 Results.
• PISA 2022 creative-thinking assessment: OECD (2024), PISA 2022 Results (Volume III): Creative Minds, Creative Schools.
• PIAAC 2023 (the OECD’s second-round Survey of Adult Skills): OECD (2024), Do Adults Have the Skills They Need to Thrive in a Changing World? Survey of Adult Skills 2023. Korean literacy is down 23 points relative to 2012; numeracy is also down. https://www.oecd.org/en/publications/survey-of-adults-skills-2023-country-notes_ab4f6b8c-en/korea-republic-of_5f95963c-en.html
• TIMSS Video Study 1995 / 1999: James W. Stigler & James Hiebert, The Teaching Gap: Best Ideas from the World’s Teachers for Improving Education in the Classroom (Free Press, 1999); James Hiebert et al., Teaching Mathematics in Seven Countries: Results from the TIMSS 1999 Video Study (NCES, 2003). The 1995 study covered Germany, Japan, and the U.S.; the 1999 study expanded to seven systems: Australia, Czech Republic, Hong Kong, Japan, the Netherlands, Switzerland, and the U.S. (without Germany). “Shallow teaching syndrome” comes from the 1999 Australian national report: Hilary Hollingsworth, Jan Lokan & Barry McCrae, Teaching Mathematics in Australia: Results from the TIMSS 1999 Video Study (ACER, 2003).
• NAEP 2024: U.S. Department of Education, Institute of Education Sciences, National Center for Education Statistics, 2024 NAEP Reading & Mathematics Assessment Results, https://www.nationsreportcard.gov/. The 4th-grade reading 10/90 percentile gap is 107 points.

On the four-system education and training reforms (Chapters 1 & 4)

• China’s Double Reduction policy: July 2021, Opinions on Further Reducing the Workload of Compulsory Education Students and the Burden of Off-Campus Tutoring (CCP Central Committee General Office and State Council General Office).
• China’s Education Strong Nation Plan (2024-2035): issued by the CCP Central Committee and the State Council in January 2025.
• China’s AI + Education Action Plan: jointly issued in April 2026 by the Ministry of Education, NDRC, MIIT, MoST, and the National Data Administration; reference number Jiao Ke Xin [2026] No. 1.
• Shanghai’s grades 4-7 mandatory AI curriculum and Huangpu District’s “Climbing the Peak” program: published documents of the Shanghai Municipal Education Commission and the Huangpu District Education Bureau.
• U.S. federal AI-education executive order: “Advancing Artificial Intelligence Education for American Youth” (The White House, April 2025).
• Japan’s GIGA School plan (2020-2024 nationwide K-12 device deployment): MEXT.
• Japan’s MEXT December 2024 Guidelines on the Use of Generative AI at the Primary and Lower-Secondary Stages: MEXT.
• Japan’s AI Education Accelerator Program: expanded teacher AI training through industry partnerships (including SoftBank Robotics among others). Specific scale figures (e.g. “about 50,000 teachers trained”) currently appear only in secondary sources (industry blogs and media reports) and have not been directly verified against MEXT official documents; the essay therefore uses qualitative phrasing.
• Korea’s AI Digital Textbook (AIDT): rolled out in stages in March 2025 across selected subjects; downgraded by the National Assembly within 2025 to “supplementary materials” (no longer statutory textbooks) in light of effectiveness concerns and teacher feedback, with budget and coverage substantially reduced. Sources: Korean Ministry of Education announcements and related National Assembly legislative records.

On the cognitive-mechanism analysis of gaokao math (Chapter 2)

• The author’s team’s analysis of cognitive mechanisms in 2021-2025 gaokao math. Sample: 567 problems across 7 distinct paper variants — Shanghai, Beijing, Zhejiang, Quan-Guo-Jia (Sciences), Quan-Guo-Yi (Sciences), New Gaokao Paper I, New Gaokao Paper II. Each problem, under a v2 calibrated classifier, outputs a verdict (GOOD / MIXED / BAD), a primary_mechanism (across the seven mechanisms), and a crammer_gap (drilling premium), among other dimensions. Data assets: scored.jsonl (567 lines) plus findings.md (project conclusions document).

On AI and the labour market (Chapters 3 & 5)

• The judgment that “AI lifts efficiency dramatically while keeping quality at or above the median” is drawn from the analysis of AI’s impact on workplace output in Offbook Press ISSUE 02 (Rebuilding Learning) and ISSUE 03 (Breakdown of Firms).

Continuities with earlier issues of Offbook Press

• “Cognitive decoupling,” “rumination consumption,” and “cross-domain isomorphism” come from ISSUE 01, On Cognitive Decoupling.
• “AI review loop” and “AI lifts efficiency dramatically while keeping quality at or above the median” come from ISSUE 02, Rebuilding Learning.
• “Organisational dissolution,” “the resistance to AI landing inside large organisations,” and the early discussion of “know-how distillation” come from ISSUE 03, Breakdown of Firms.
• “Form being able to be produced independently of substance” and the training of “authorship” come from ISSUE 04, Mirage of Form.

This essay paraphrases the sources above conceptually rather than quoting them directly. For verification of original wording, consult the original links and publications.

ISSUE 05 ends · Offbook Press · 2026