The main changes are near —
and they don't reverse.

The changes chain together. The main part lands in the near future. And because the premise has inverted, it does not move back.

3-08 showed that the AI revolution starts from below: there are countless places for it to start, and no hand that can stop it. This chapter gives that unstoppability as a structure. It lays out the order in which the changes cascade, marks the boundary inside which complete replacement happens, and sets down the reason nothing moves back.

But that complete replacement is bounded to software development. The second half of the chapter makes the boundary explicit.

The chain of change

Put the claims made across the preceding chapters back in the order in which they cascade.

• AI reaches human-top execution capability (1-01) — the Codeforces 2700 tier, $20 a month
• The main battleground of maintenance moves to design (1-02)
• The role called "coder" goes away (1-03)
• A new role, the builder, emerges (1-04)
• Customers themselves become builders (1-05) — nine-tenths in-house, only one-tenth outsourced
• The SIer commission model becomes structurally uneconomic (3-04) — the same effort builds it in-house
• Lock-in dissolves (3-05) — AI-native standard code, the opposite end from Palantir's FDE
• Companies hire builders (3-06) — the senior builder moves into management, the CIO's seat, while the professional work is done by AI. The supply is not only former coders; AI plus Python plus Flet bring in the VB/VBA generation, makers, shop-floor engineers, and students as new entrants
• Multi-tier subcontracting absorbs the transition (3-07) — it can shrink without internal lay-offs

These are not independent observations. They cascade in order from one fact: AI took over execution.

Within the chain, what moves fastest is new projects and extension work. What moves slowest is the full replacement of core business systems. But both point in the same direction, and neither stops.

Why only coding moves to complete replacement

This is where the scope of this series has to be made explicit.

Software development is a broad field: requirements gathering, design, coding, testing, deployment, operations, incident response, stakeholder coordination. What AI fully replaces is only the coding inside that list. The reason is that two conditions hold at once.

1. The rules are explicit — language specs, standard-library APIs, type systems, syntax. All of it is defined formally. There is little interpretive room about what counts as the correct way to write it
2. Correctness is verifiable — whether the code compiles, whether the tests pass, whether a competitive-programming problem is solved. All of it is checkable mechanically

Where both conditions hold, AI receives, during training, an enormous volume of feedback on both "did it follow the rules" and "is it correct." That is why AI reaches superhuman levels in the coding domain.

The decisive factor is a verifier that runs free and forever

Let me restate the second condition, that correctness is verifiable, one level deeper. Whether complete replacement happens turns on whether the verifier runs free, instantly, and an unlimited number of times.

Coding has that verifier built in. Compiler, type checker, tests, runtime. Each of them runs for free, in an instant, any number of times. So AI can try a million times toward the correct answer. It can call its own grader without limit.

The desk work, self-driving, and robotics discussed below have no such verifier. The only way to check right from wrong is reality itself. And reality is expensive, slow, and irreversible. A surgery or an accident response cannot be re-run a million times for free. So AI can grind down the knowledge and pattern layers, but it cannot hold a grader that closes the last 1%.

The other parts of software development — requirements, design, operations, incident response, stakeholder coordination — carry the same structural 1% problem as the self-driving and Shinkansen cases below. This is what 1-03's "coders go away, builders remain" means. Coding gets complete replacement; builder work gets a productivity gain. Both happen inside the same field at the same time.

One thing to note: now that coding is fully replaced, the importance of the requirements side rises rather than falls. Skimp on requirements, and AI only mass-produces commonplace code, which is not the same as solving the specific business problem. AI is excellent at probabilistically reproducing what it has seen in prior samples from the same domain, but pinning down the non-negotiable conditions of this organization's particular business can only be done by the human who wrote the requirements.

The faster AI gets, the faster and the larger the cost of sloppy requirements piles up. A vast amount of code that runs but is ordinary gets produced, and maintenance collapses. It is the same failure shape as the vibe coding of 1-02, now faster and at greater volume. In a world where coding is cheap, requirements determine a software system's differentiation and its lifespan.

Other AI applications stall at the last 1%

In the reverse case — domains where the rules are not explicit, or where correctness is hard to verify — AI does not advance at the same speed. If either condition is missing, the last 1% remains. Three representative domains.

• Desk work — 99% of the work (routine documents, replying to mail, summarizing minutes, research drafts, tidying data, draft translations) can be handed to AI. But the last 1% — the unwritten rules inside the company, decisions that carry responsibility, subtle adjustments with stakeholders, the final judgment of whether to submit — decides the quality of the work and the trust in it
• Self-driving — in 99% of situations the car drives without trouble. But the last 1% — an unexpected pedestrian movement, a judgment call in bad weather, a child's ball — decides whether someone lives. The difficulty of moving from 99% to 100% is the substance of the problem
• Robotics — 99% of motion (routine assembly, picking, serving food, cleaning, repeated actions) can be mechanized. But the last 1% — unexpected object placement, handling soft items, judging how to share space with humans safely, adapting to unknown environments — decides usability on site

Look at rail, and especially at a closed system like Japan's Shinkansen, where route and obstacles are tightly controlled. Almost all normal operation can be automated. The rules are explicit, and correctness in normal operation is easy to verify. But the judgment required for accidents and equipment failures — derailment, defective equipment, natural disasters — is the kind of problem that cannot be enumerated at design time, and it stays with humans. The last 1% sits not in the openness of the system but in the unpredictability of abnormal events. No matter how closed the system is, that part does not disappear.

This judgment in abnormal events is structurally hard for two reasons.

• The expansion of what must be anticipated — enumerate one accident or failure, and its variants, its combinations, and new patterns keep appearing. The list of cases one believes has been anticipated is always incomplete, and what actually occurs in the field sits outside the design-time list. List more and the list grows; stop listing and the gaps remain
• The absence of a body — humans detect anomalies by taking in vision, touch, sound, smell, and vibration through the body all at once. AI has no body, so cameras and sensors have to be installed in its place. Each physical quantity requires its own equipment, and placement, power, networking, and maintenance costs pile up. And what to sense in the first place is itself another problem of predicting abnormal events. The anomaly nobody anticipated has no sensor on it

Because those two compound, complete replacement in the physical world stays structurally hard even when the system is closed.

In these domains AI delivers large value as a productivity tool: document drafts, driver assistance, routine work by collaborative robots. But complete replacement does not happen. There is a deep valley between being able to do 99% and being able to do 100%.

The IT industry's AI narrative often overlooks that 99/100 valley, or pretends not to see it. Every time AI comes up, lines like "it solves the labor shortage across every industry" and "all white-collar work gets automated" appear. That is overestimation.

This series stands apart from that overestimation. It argues complete replacement for one specific area, coding inside software development, where the rules are explicit and correctness is mechanically verifiable. It does not claim the same complete replacement at the same speed for the rest of software development or for other domains.

The writing of this series is itself the example

Living evidence for this claim sits in the writing process of this series itself.

As 1-04 noted, this series was written by one person plus AI in about a week. But that week contained a long list of corrections made by a human.

• The price anchor was moved back from Claude Max ($200 a month) to Claude Pro ($20 a month) — the base plan is enough to stand the toolset up and run it
• The code base written as 30,000 lines was corrected to the measured 6,000 lines
• The abacus (soroban) was added as the primary example alongside the human computer
• The "roughly a decade" written for the calculator transition could not be sourced and was replaced with how fast it spread: one million Casio Minis in ten months
• The framing of "the IT revolution completing" was added to 1-01
• The origin of multi-tier subcontracting was restated correctly as large coder head-count demand
• The section "physical goods become scarcer than software" was added to 3-07
• This chapter's scoping itself — complete replacement only in coding — was added

Every one of these is *a correction produced by a human reading the AI's draft and judging it*. Left to AI alone, factual errors, slanted arguments, and tonal slips — exactly the kind of problem that costs a reader's trust — pass through untouched. The level of this series required corrections made by human judgment.

In other words, the writing of this series carried the same structure as desk work, self-driving, and robotics. AI writes most of the draft; the human holds judgment and correction. Productivity multiplies several times over, but complete replacement does not happen.

It happens in the near future — why "near"

Here is the ground for this series' sense of timing when it says "the near future." The main changes happen within a few years. That is the outlook of this series.

Why near? Several independent time scales all converge on the near side.

• The AI capability curve — the threshold was crossed in 2024-2025 (1-01). In terms of capability, the transition is already possible
• The customer learning curve — it takes a few years for customers to learn to build alongside AI (1-05). That is moving now
• The contract-renewal cycle — SIer long-term maintenance contracts typically run three to five years. The next renewal is when replacement gets evaluated (3-05)
• The pace of multi-tier shrinkage — shrinkage without internal employment adjustment can be achieved in a few years (3-07)
• The history of the calculator and the abacus — completed in about ten years from the 1972 Casio Mini (1-03). The AI shift is faster than that

Those time scales all point to the near side. Not slow enough to deserve "ten years," not fast enough to deserve "one or two." As a band, within a few years. But pinning down the exact number is not the point. The main changes happen in the near future, and what is essential is the direction and the irreversibility rather than the precise count. This series makes no definite prophecy that the transition completes in five years.

What happens in the near future is the main changes, though, not everything. Core-system replacement in regulated industries takes longer. Some areas will still hold old models after ten years. Even so, the mainstream of the industry moves to AI-native in the near future.

The change proceeds as an irreversible one

Finally, confirm the irreversibility of the change. Its ground is not that it completes in a few years. Its ground is that the premise has inverted.

What the transition part of this series has argued throughout is the inversion of two premises.

• The economic premise has inverted — buying or outsourcing used to be the cheaper default. Now building it yourself is cheaper (3-01 on the two parallel worlds, 3-04 on the uneconomics). Because AI made execution orders of magnitude cheaper, the direction of economic rationality itself flipped
• The security premise has inverted too — vendor concentration like Microsoft 365 used to be the cheap and safe default. Now holding OSS and a sovereign AI on one's own machine is cheaper and safer (3-03 on sovereignty)

An inverted premise does not move back. Once the direction of economic rationality and of security has flipped, no force exists that rewinds the structure. That is why the change moves only one way. That one-directionality shows in the following concrete cases.

• Once a customer has experienced AI-native in-house development, they do not go back to SIer commissioning (1-05). The learning cost has already been paid
• Once an SIer has shrunk its multi-tier subcontracting, it does not hire subcontractors back at scale (3-07). Contract relationships that were closed do not re-form
• Once the builder is recognized as management, as the CIO, that role definition persists (3-06). What moved into a seat in management does not move back
• The fact that AI generates standard code cheaply does not change. The structure of reaching the top tier for $20 a month stays in place (1-01)

Each piece moves in one direction only. So the chain as a whole moves in one direction. On top of an inverted premise, once the chain starts, no structural force exists that stops it. "No hand that can stop it," from 3-08, has its ground here.

One historical comparison is worth keeping in view. The invention of the printing press in the 1450s reshaped the structures of the church, the university, and the state over two hundred years, preparing the ground for the Reformation, the scientific revolution, and the premises of the modern nation-state. The LLM holds an intensity that does not compare with that. What the printing press democratized was reading, that is, access to existing knowledge. What the LLM democratizes is making, that is, knowledge generation, judgment, and implementation. There is no wall of literacy to clear first; natural language works for anyone. The speed of diffusion is on a different order: what took the printing press decades takes years in the AI era. Read against that difference in intensity, the near-future transition this series describes is, if anything, a conservative estimate.

The free person of the Middle Ages, and the free person of the AI era

When the medieval European free person emerged out from under the feudal lords, four conditions came together at once. Economic autonomy (the free farmer who tilled his own land; the urban citizen, merchant, and craftsman who traded independently). Political self-governance (the free cities that won their charters from lords). The means of touching reality (the right to bear arms, the capacity to grow one's own crops). And education, the seven liberal arts.

The conditions that converge as the AI-era free person — the builder of this series — stands up correspond one to one.

Just as the medieval liberal arts could not stand on education alone, the contemporary liberal arts cannot stand by themselves either. *A free person comes into being only when all four converge.* And just as the free citizens of medieval cities formed guilds to strengthen their economic weight and their voice, the AI-era builder moves to the side that makes business decisions, into management, the CIO's seat in each company (3-06). The professional work is done by AI, and the human holds judgment and responsibility as part of management.

Employment is the AI era's serfdom, and the rise of self-employment is structural

Placed next to the medieval free person, one more thing becomes visible. *Modern employment, the salaried worker, sits structurally in the same position as the medieval serf.*

The stability of employment and the stability of serfdom are the same trade-off structurally: handing over the right of self-determination in exchange for predictability of survival. This is not a claim of moral equivalence — modern employment has legal protections and contractual freedom. It is the analytical observation that on the three axes of ownership of the means of production, judgment, and mobility, the structure matches.

And the reasons employment stops fitting in the AI era are structurally clear.

1. The means of production can now be owned individually — a few-thousand-yen-a-month AI, local LLMs, open source, one's own server. The employer no longer needs to monopolize them
2. One person plus AI does a team's work (1-04) — the payoff of concentration disappears
3. The boundary between judgment and execution closes within one person (1-04) — the overhead of aggregation, coordination, and management becomes pure waste
4. Those who make business decisions are intrinsically inclined to independence — owner-operators and the self-employed choose to run the business on their own judgment, and that is not an accident (3-06)

The rise of self-employment is not a policy question or a lifestyle question. It is structural necessity. The same structure under which the medieval free citizen, free farmer, and craftsman were all self-employed returns in the AI era.

Employment is the contemporary form of medieval serfdom, and *self-employment is the contemporary form of being a free person*.

The structural changes this series has argued — the structural uneconomics of the SIer commission model (3-04), customers building for themselves (1-05), the judgment-centered builder (1-04 and 3-06), the error in the specialized-engineer advice (this chapter) — all converge on one point. The industry structure organized around employment is reshaped in the AI era.

The middle layer — builders who hold physical reality

Between the pure-software free person, the builder of this series, and the pure-physical free person that the natural-farming series covers, a middle layer rises up that bridges the two.

The medieval world had the same layer: stonemasons, carpenters, smiths, weavers. They held guilds, had work in both city and countryside, and kept a foot in both the city's self-governance and the soil's reality. It was precisely that craft layer that accumulated the technical capital the Renaissance was later built on.

The middle layer of the AI era sits in the same structural place. The inputs come from physical reality (sensors, observation, material) and the outputs land in physical reality (objects, harvests, repaired machines, buildings). AI acts as the mediator and takes over the design and the analysis, but the hand that touches reality stays a human hand. Who belongs here:

• Makers and digital fabrication (AI design plus 3D printing, laser, CNC)
• Embedded engineers and robotics designers (microcontrollers, PLCs, ROS2)
• Precision agriculture and agritech (sensors, drones, local LLMs running in the field)
• Manufacturing-floor technicians (having AI rewrite the factory automation)
• *Physicians, mechanics, and repair technicians who hand the imaging and diagnostic work to AI and keep the judgment and the procedure*
• Carpenters, architects, and craftspeople who use AI design tools

What 3-06 covered as makers and field technicians entering embedded work was precisely new entry into this middle layer. The labor demand created by the era in which physical goods become scarce (3-07) is absorbed here too. The coders flowing out of the SIer industry do not all find new work inside pure software; a path sideways into the middle layer opens here.

Japan, with deep foundations in manufacturing, town factories, natural farming, electronics tinkering, and repair culture, holds a structural advantage in the move to this layer. As the alternative path to the "become a specialized engineer" advice in the next section, this gives a second route alongside stepping sideways onto the liberal arts: stepping sideways into being a builder who holds physical reality. Both are roads out of the lord's manor.

"Become a specialized engineer" misreads the structure

There is a widely repeated piece of advice: in the AI era, become a specialized engineer, hold a deep specialty that AI cannot take, such as security or ML. That misreads the structure.

What AI is taking over is the whole layer of software engineering, not a particular subdomain inside it (1-01, 1-03). Going deeper into a specialty only shifts the date at which that specialty is overtaken; the underlying structure is the same. In the medieval analogy, it amounts to telling a serf that becoming a more specialized serf will make him free. Freedom does not come from going deeper into the specialty. It comes only from stepping out of the lord's structure of control.

The path to becoming a free person of the AI era is the same. The direction is not deeper specialization within engineering. It is stepping sideways onto the liberal-arts axis of judgment, verbalization, ethics, and systems thinking. That is the structurally correct direction of movement.

This is the beginning of the Second Renaissance

The structural change argued so far — from coder to builder, from software engineering to the liberal arts, from employment to self-employment, from the lord's manor to the free city, from pure software to a middle layer that holds physical reality — lines up item for item with the structural change of the First Renaissance (14th to 17th centuries).

Nine items, all corresponding. This is not a metaphor. It is *a structural similarity*.

And the First Renaissance did not begin one morning. The self-governance of 12th and 13th century cities, the formation of the guilds, scholastic philosophy, and the rediscovery of classical texts through the Crusades accumulated as the underlying ground, and the printing press of the 1450s accelerated it. The Second Renaissance follows the same pattern. The personal computer, the Web, open source, maker culture, the revival of natural and organic farming, the data-sovereignty movement, and the AI ethics conversation have accumulated as the ground, and the LLM (from the 2022 period onward) is accelerating it. That is the phase we are in.

The near-future structural transition this series describes is *one cross-section of the Second Renaissance*. The series covered the software domain, but the same structural change proceeds in other domains of life at the same time.

The AI revolution is the completion of the IT revolution

Treating the AI revolution as a separate, new revolution is another misreading. The AI revolution is the completed form of the IT revolution. Seventy years of the IT revolution are finally fulfilling the original promise.

The IT industry until now had humans hand-writing the code that automates work. Against the original promise of the IT revolution — computers do the work, humans are freed — the side that implements the automation has been doing it by hand for seventy years. That is a strange structure. The consequences: programmer became one of the highest-paid professions; the cost of automation often exceeded the cost of doing the work by hand; and a huge industry of manual labor for automation appeared, made of SIers, consultancies, and SaaS.

Logically, it is an odd arrangement. If automation is the goal, making the automation should also be automated. The LLM dissolves that twist. Because AI writes the code, the original promise of the IT revolution is fulfilled literally. The AI revolution is not the beginning of a new revolution; it is the completion of the IT revolution.

That completion functions as the strongest accelerator of the Second Renaissance. The SIer industry shrinking, and the software engineer's role being replaced by the builder, are inevitable consequences rather than sudden shocks.

The LLM is a powerful statistical-processing tool, not a superintelligence

Viewed coolly, the LLM (Claude, GPT, Gemini, and the rest) is large-scale statistical processing of data. From an enormous body of text taken from textbooks, papers, and the Web, it predicts the token most likely to come next in context. It is an overwhelmingly powerful tool, but it is not, in itself, a superintelligence.

The pitch that AGI is coming and that white-collar work will all be automated in 12 to 18 months (Suleyman, the CEO of Microsoft AI, among others) deliberately stages that reality as a superintelligence parable, in order to lead to "so hand judgment over to AI" and "so buy Copilot."

*Handing judgment and responsibility to a statistical-processing tool is structurally wrong.* The LLM makes writing, looking things up, and organizing orders of magnitude faster, but deciding what to build, evaluating whether it is right, and taking responsibility stay on the human side. That is the logical basis of the role this series has been calling the builder.

Read the essence of the AI revolution not as the arrival of superintelligence but as a powerful statistical-processing tool finally making the IT revolution's automation promise implementable, and the structure becomes clear. The SIer industry shrinking, the builder rising, the foundational shift from software engineering to the liberal arts — all of it is explained by the structure that, as the tool got strong, the human role shifts to the judgment side. Anthropomorphizing AGI only makes the substance harder to see.

Apps do not disappear; the way of making them changes, and it comes to resemble film-making

Stated precisely, the structural change is this. *Software development as an engineering craft disappears, but apps do not disappear.*

The most precise parallel is film-making. A film is made by independent specialist roles coming together: cinematography, editing, sound, lighting, costume, set design, visual effects, scoring, acting. The audience is aware of none of it. Only one artifact, the film, appears. At the center are not the people handling each technical task but the director and the screenwriter, the people who carry the creative judgment.

App-making in the AI era takes the same structure.

A director does not learn to operate the camera. A screenwriter does not learn lighting. The audience does not know how the film was made. The film exists all the same, and carries value. Apps take the same shape. The user does not learn engineering, AI takes on the engineering work, end users do not know how it was made, and the app exists all the same.

Just as the printing press removed the scribe but not the book, the LLM shrinks the software engineer but does not remove the app. Only the way of making changes, and the way it changes is closer to film-making than to book-printing.

Film-making, though, has an enormous range. A Hollywood blockbuster still requires a large crew, an enormous budget, and years of work, while a YouTube video can be made by anyone with one smartphone. AI-era apps have the same range.

Monolithic large-scale is structurally a poor fit for the AI era. No single senior builder can hold the whole of it, it creates lock-in (3-05), maintenance is hard, and the chain of judgment is dispersed. These systems get decomposed into combinations of mid-scale focused systems.

Mid-scale is the senior builder's home territory. It is the scale at which the chain of judgment closes within one person (1-04), and it is the position of management, the CIO (3-06). Mid-scale apps themselves do not shrink in number. They grow: business apps that previously could not be cost-justified now get built in large numbers.

At the personal scale, the user is also the director.

What declines, then, is not the number of apps but the total number of workers who build them, and in particular the labor model of monolithic large-scale SIer projects. Apps themselves continue to exist across all three scales, and they grow at mid-scale and at personal scale.

This is the most precise statement of the structural change this series has argued. The SIer labor model shrinks substantially, senior builders work at mid-scale as the directors of the AI era, and at the personal scale the work is absorbed into the user.

Not only the AI revolution

This chapter has been writing "the AI era," but trying to capture the current structural change through AI alone misses more than half of it. Lay out the transitions running in parallel.

• The end of fossil resources — the collapse of the premises of an oil-dependent society (Structural analysis, chapters 02 and 14)
• Geopolitical multipolarity — the end of US unipolarity, Trump, Ukraine, Iran, China
• A generational shift in the defense industry — from large weapons platforms to drones plus AI (Structural analysis, chapter 11)
• The reconstruction of agriculture — the limits of chemical-fertilizer dependence and large-scale industrial farming, and the rise of regenerative agriculture (Structural analysis, chapter 03)
• The collapse of finance and trade premises — the dollar standard, the fragmentation of global supply chains
• The simultaneous breakdown of demographics, cities, healthcare, and pensions — the end of institutions built on the assumption of desk work (Structural analysis, chapter 11)
• And the AI revolution — the accelerator for all of the above

The First Renaissance had to be understood as a composite of the printing press, the age of discovery, the Reformation, the scientific revolution, the nation state, the rise of the commercial bourgeoisie, and the labor shifts after the Black Death. In the same way, the Second Renaissance cannot be captured by the AI revolution alone. Several independent transitions run in parallel, and their convergence point is what makes an era that is no longer the same as the one before. AI is the strongest accelerator among them, but it is not the whole cause.

An age of creation, and an age of upheaval

The Renaissance sits in the textbooks as a luminous age of creation: Leonardo, Michelangelo, Galileo, Gutenberg. But that same age was also an age of violent upheaval. The Reformation and the wars of religion (the Thirty Years' War cut Central Europe's population severely), the corruption and schism of the papacy, recurring plague, a populist demagogue like Savonarola staging the bonfire of the vanities in Florence, and a strongman politician like Cesare Borgia becoming Machiavelli's model in The Prince. While the old order is collapsing and the new order has not yet stood up, people seek refuge in strong men and extreme words.

The upheaval of the Second Renaissance is already unfolding in front of us. President Trump is the canonical case. Direct attacks on the expert class, on the judiciary, and on scientific consensus. Ad-hoc swings on tariffs, immigration, and science budgets. Streams of executive orders that override congressional checks. A governing style in which one person decides everything.

Placed next to Nadella's Copilot strategy, the structure becomes visible. Nadella concentrates a company's judgment into a single AI; Trump concentrates a nation's judgment into a single president. The means differ, but both push the old era's logic of concentrated judgment to its limit, and that is the same structure (link:/en/blog/nadella-hegel-cunning-of-reason/[related: Microsoft's Nadella and Hegel's Philosophy]).

Just as the populist demagogues of the Renaissance eventually disappeared, those who push the concentration of judgment to the extreme stop fitting the structure of the new era — distribution, the free person, judgment held close — and exit. But until then the upheaval continues. This too is the same pattern as the First Renaissance.

Summary

• The changes chain together — from the single fact that AI took over execution, the coder disappears, builder demand appears, customers build for themselves, SIers shrink, and the industry turns over
• Complete replacement happens only in coding — because the rules are explicit and the verifier runs free and forever
• Other domains stall at the last 1% — desk work, self-driving, robotics. There, AI stops at a productivity gain
• The main changes are near — the capability curve, the learning curve, contract renewals, subcontracting shrinkage, and the calculator's history all point to the near side
• The reason it does not reverse is that the premise inverted — both economics and security now point the other way
• Employment is the AI era's serfdom, and self-employment is the contemporary free person — the means of production can now be owned individually
• This is one cross-section of a Second Renaissance — the AI revolution is its strongest accelerator, but not the whole cause

Here is the conclusion of this series, compressed.

AI reached human-top execution capability. That happened because the domain has explicit rules and verifiable correctness. As a consequence the coder — the role that puts coding at the center of the work — goes away, and the builder, the judgment-side role, stands in its place. The SIer commission model cannot hold structurally, and in the near future the mainstream of the industry moves to AI-native in-house development. And because the premise has inverted, it does not move back.

But this is the story of one specific area, coding inside software development. The other parts of software development — requirements, design, operations, incident response, stakeholder coordination — carry the same structural 1% problem as self-driving and the Shinkansen, and that is where the builder works. Nor is complete replacement at the same speed claimed for other domains, whether desk work, self-driving, or robotics. In those, AI works as a productivity tool and does not reach complete replacement.

And during the same few years in which AI advances, society as a whole moves toward physical goods being scarce (3-07). AI data-center construction, manufacturing coming back, the shift to natural farming — each of them generates demand for physical labor. The coders flowing out of the SIer industry are absorbed both inside and outside the industry.

What aiseed.dev has argued across this series is the following. A structural transition centered on coding inside software development happens in the near future. That transition is irreversible, because the premise has inverted. And the argument about this specific area must not be extended automatically to the rest of software development or to other domains.

There is one more current, the one named in 1-04. The foundational discipline of the technical professions moves from software engineering to the liberal arts. Because what AI has taken over is the core of software engineering — algorithms, languages, frameworks, design patterns — what remains on the human side is judgment: the craft of logic, verbalization, ethics, systems thinking, and history that the liberal arts have always been. Just as the medieval liberal arts were defined as the arts of the free person, the person who is not enslaved, the builder is the person who does not hand judgment over to AI, the contemporary form of the same thing.

Hold on to that, and the IT industry's AI narrative no longer sweeps you along. You can read calmly what is actually happening as a structure. And from wherever you stand — as a customer commissioning software, as a coder, as a builder candidate, as an SIer executive — you can decide what to do over the next few years.

Thank you for reading to the end. aiseed.dev will continue to publish articles that read the structure.

Related articles

• 3-08: The AI Revolution Starts From Below
• 1-01: AI Solves the World's Hardest Coding Problems
• 1-03: AI Now Does the Software Engineer's Work
• 1-04: The Builder Role
• 3-07: Japan's SIer Industry Transition and Labor Mobility
• Phosphorus Depletion and Natural Farming
• Structural analysis 08: Subtracting the Enterprise-IT Tax
• Structural analysis 12: AI and the Individual Business