Preserving Judgment and Self-Worth in the Age of AI

Explores career changes, verification standards, and capability building as AI improves productivity, examining judgment, taste, and reasoning while considering how to preserve independent thought, room for growth, and self-worth.

The faster AI develops, the more we need to distinguish two things: which tasks we can hand over to tools, and which judgments and experiences we must still take responsibility for ourselves. Starting with letters from three readers, this article explores our place at work, the development of our capabilities, personal agency, and our sense of worth.

Over the past month, I came across three thought-provoking discussions on Twitter. They came from three software professionals I had never met: people working in frontend infrastructure or operations on the data side. Their concerns can be grouped into three questions:

1. Once AI takes over execution, what is left for people to do? The first discussion asked: “Do you think we will eventually hand everything over to AI? If so, will things like code review still need people?”
2. Are we creating tools, or replacing ourselves? The second asked: “Will our future work simply be building AI and then letting it replace us?”
3. What happens to our own growth once we become more efficient? The third listed reasons to embrace AI coding—less rework, less communication, higher output—but admitted at the end of the letter that the question of personal growth remained unresolved.

My answer to the first question was: soon. By the end of this year or next year; the year after that at the latest.

These questions seem to concern engineering practice and career planning, but underneath they point to the same thing: once machines take over execution, where do people stand?

When AI takes everything it can take, what do we have left?

I offered a few scattered replies in the comments, but this question deserves a fuller answer. It is also something I have been thinking about often lately. Recording these thoughts may help others wrestling with the same uncertainty.

1. From Tool Advantage to End-State Thinking

Green mountains cannot hold it back; the river flows east all the same. — Xin Qiji, “Bodhisattva Barbarian: Written on the Wall at Zaokou in Jiangxi”

More than two hundred years ago in Britain, a spinning machine called the spinning jenny appeared. One person could do what had previously required several. Yarn became cheaper, and the livelihoods of spinners who depended on their hands and an old spinning wheel collapsed in a very short time.

Later, some people went out at night to smash the machines. History calls this the Luddite movement. They are often portrayed as fools who hated technology, but that is a misreading. What they wanted to destroy was never the machine itself, but their own position, which had suddenly become worthless in its presence. They did not fail to understand the technology. On the contrary, they understood what was happening earlier than almost anyone else.

On the day Meituan announced layoffs of 30–50%, a friend asked me on WeChat what to do. I replied with one sentence: we should not become the women spinners replaced by the spinning jenny; we should become the people who use it efficiently.

He may have remembered that sentence for a long time. Today, though, I want to revise it: it was only half right.

Because the next page of history reads like this: skilled spinning-jenny operators did not remain secure for long either. Faster machines soon appeared, and the spinning jenny went into the museum along with its skilled operators. Every machine waits for the next machine. “Learn to use the new machine” has always been a reprieve, never a pardon.

So the real question is not whether you know how to use AI. Today, people who do have an advantage over those who do not. But that advantage may have a half-life of only a year or two—and among those at the forefront, perhaps no more than one or two months. As models become increasingly dominant, the window for exploration and adaptation grows shorter. Every new generation of models brings another shift in the paradigm. The workflows you painstakingly built, the prompting techniques you distilled, and the engineering solutions you developed may all become spinning jennies overnight.

I have only one approach to this. I call it end-state thinking: do not exhaust yourself on problems that belong to an intermediate stage; think and act with the eventual end state as your starting assumption.

For example, a while ago I had a serious debate with several peers: as AI writes more and more code, do engineers still need to understand that code line by line?

There are many possible answers: better documentation, better visualization, mandatory code walkthroughs. But from an end-state perspective, this is not the central question. Once intent can be translated directly and reliably into a system, the requirement that people understand every line of an intermediate artifact loses its meaning—just as nobody today requires engineers to understand every line of assembly generated by a compiler. Being trapped by a problem destined to disappear is one of the most hidden forms of waste in this era.

What does the end state look like? My judgment is simple enough to fit into one sentence: PRD is code—or, more generally, intent is implementation. A sufficiently clear expression of intent becomes a running implementation directly. Every redundant step in between—scheduling, integration coordination, cross-review, repeated translation—disappears. As I see it, this is not ten years away. It could be fully achievable within this year. Take Fable 5 from a few days ago: anyone who has tried it will understand the gap between it and Opus 4.8. It is confident and powerful, to the point that I sometimes wondered whether the questions I was asking it were too foolish.

This yields the first implication of end-state thinking, a cold but honest one: assume that every step between intent and implementation will disappear.

The remaining question is: where should people stand?

2. Preserve Verification Standards as Productivity Rises

Aim for the highest standard, and you may reach the middle; aim for the middle, and you will fall below it. — Li Shimin, The Emperor's Model

Let me begin with the change in my own work.

Over the past year, I have handed over every part of my work that I could to AI, one piece at a time. It writes the proposals. It writes the code. It also writes the first drafts of documentation and review comments. The work I can advance in parallel in one evening would have filled a quarter two years ago.

In theory, I should have more free time. The opposite is true: I am busier than ever, but the substance of that busyness has changed. I hardly produce anything directly anymore. I spend my days checking and accepting what it produces.

This is the early outline of the end-state workflow. In it, people have only two tasks left: review proposals and verify results.

Notice what these two tasks have in common: neither is production. Both are quality control.

Ninety years ago, in Modern Times, Charlie Chaplin tightened bolts beside a conveyor belt. Unable to keep up, he was pulled into the gears. That image became a classic because it captured the human condition in the industrial age so precisely: machines set the pace, and people chased after them.

Ninety years later, the conveyor belt has changed direction: now machines produce, and people inspect.

AI's production speed is almost unlimited, and the cost of starting a new task approaches zero. But the cost of verifying its output has not fallen at all, because verification requires human involvement. For the first time in human history, the production bottleneck has shifted entirely from the machine side to the human side.

This shift brings a quiet but devastating implication. You can run six or seven sessions concurrently, dance between different worktrees, switch contexts frantically, and juggle multiple tasks with elegance. But on this new assembly line, if you want to increase your throughput further, you will discover that there is only one way:

Lower your own standards.

Read one less line of code. Follow the logic one step less far. Ask one fewer “why.” When the voice in your head says, “Never mind, it is probably fine,” nod and let it through. Following this reasoning, we can practically predict that most declines in engineering quality over the coming years will be structural, rather than the result of someone being lazy. When production is infinitely fast, verification standards become the only compressible variable in the system. The incentives to compress them are everywhere: your boss's expectations, your peers' speed, the whip of performance reviews—all urging you to loosen your grip.

But think carefully about what letting go means. The very reason you still occupy the verification step, rather than having been optimized away, is that your standards are higher than the machine's. A gatekeeper who continually lowers standards is personally proving their own replaceability. My position on this has never changed: a little less speed is fine; the standards must not give an inch.

3. Find New Career Defenses

Do not rely on the enemy not coming; rely on being prepared to meet them. — The Art of War, “The Nine Variations”

But there is a more fundamental problem: not everyone can safely retreat into the role of gatekeeper. Some jobs will be submerged sooner than others, without even a chance to retreat.

Within our industry, I think frontend engineers will be among the first to feel the water temperature change.

This is not because frontend work is unimportant, but because its position is unusual. Frontend sits at the hub of the development chain: product and design upstream, backend services downstream, testing alongside. It connects more roles than any other role, like a router.

At the same time, its moat is the shallowest. The technical complexity of a single page is limited. If you build it badly, the impact stays on that page and does not spread to other systems. Even if it becomes impossible to maintain, the cost of tearing it down and rewriting it is low.

A relatively low barrier to entry—here I mean ordinary business applications, not infrastructure—a small scope of impact, and the ability to rewrite: together, these three characteristics mean there is little protection. The entire industry is watching everyone flood into the room with the lowest threshold. Designers no longer hand over mockups; they deliver HTML directly through D2C, sometimes better than engineers do. Test engineers are moving into frontend work in groups. Backend engineers build the pages while they are at it. Even people who cannot code at all are using AI to take on their first freelance projects.

But do not read this section merely as a story about frontend development. It is a story about defenses, applicable to any industry and any role. You only need to ask yourself two questions:

1. How much of your work is essentially moving information and converting formats?

   Translating requirements into code, data into reports, policies into slides, one party's words for another party—every role that makes its living by passing messages between two domains stands on the tidal flat where the water is rising fastest. Large language models are precisely the machines built for format conversion.

2. How much does it cost when you get it wrong?

   The lower the cost of failure, the easier it is to let machines learn through trial and error. Once machines are allowed to do that, their iteration speed overwhelms every human competitor.

   Conversely, in areas where failure carries enormous costs—system architecture, high performance, foundational quality affecting tens of millions of users—machines cannot fully enter yet. The reason is not that they lack intelligence, but that nobody dares to let them experiment. Payments, trading, finance, large client applications, infrastructure containers, and similar systems are examples.

Once you understand these two questions, the way forward becomes clearer. There are two directions.

Go Deeper: Take Responsibility for the Whole System

Move from being the author of a page to being the guardian of a system. Go where failure has serious consequences: performance, architecture, quality, and stability. My advice to those frontend friends in my replies was to take on some backend requirements. For small requests, handle the entire stack yourself. The purpose is not to learn one more technology, but to give your understanding depth beyond a single point. Job boundaries are breaking apart. The future needs people who can take responsibility for a whole system, rather than engineers confined to one end of it.

Go Somewhere New: Explore New Ways of Working

Build things nobody has built before, things that other people's workflows come to depend on. Here is an important rule: do not build something that can be obtained merely by burning tokens. Any engineering product someone else has already made, which you can ask AI to copy—another review bot, another process tool—is not worth your effort, because its acquisition cost is already approaching zero. What deserves effort is exploration at the level of the paradigm: things that, once they exist, change how others work.

Finally, consider the window of opportunity. A year ago, “treat AI as an intern” still seemed like a safe strategy: you set the direction, it does the grunt work, and you check the result. But the intervals between model upgrades keep getting shorter. In many respects, today's AI can already be a mentor rather than an intern. Those who treat it as an intern may remain comfortable for another year or two. After that, the real question becomes: when it is stronger than you, what can you still contribute to the collaboration?

The next section addresses that question.

4. Preserve Judgment, Taste, and Reasoning

The exemplary person is not a mere implement. — The Analects, “On Government”

If execution is destined to be handed over, what do we have left?

First, I want to clarify the nature of this question. It sounds like a lament, but it is actually a process of refinement. Michelangelo said the sculpture was already inside the marble; he simply removed the excess. The AI era is doing the same thing to every profession. What it takes away—repetitive execution, format conversion, practiced information handling—was never really “us.” It was merely the time we exchanged for a salary.

Once these layers fall away, the part that cannot be taken is the true outline of a person. In other words, what remains is what is precious. And the longer something remains, the more precious it is.

In the past, these things were obscured by enormous amounts of execution work, undervalued to the point that hardly anyone noticed them. Now the covering has been lifted away, and for the first time they stand fully in the light.

In those three replies, I kept writing about the same three things. They are not skills. Skills are implements, and implements become obsolete: however skilled a spinning-jenny operator was, their technique went into the museum with the machine. These three things sit above implements. I am fairly confident that their irreplaceability arises from the technical structure of this generation of AI, rather than simply from machines not being powerful enough yet.

Judgment: Choose Problems Worth Solving

AI is the most powerful answer machine in history, but it does not set the questions. You tell it what to do, and it does it. Deciding what to do remains a human task.

Many people think AI levels the differences between people. I believe the opposite: it amplifies them. Consider a thought experiment. Suppose everyone's prompting ability has reached perfection, and anyone can get AI to accomplish something with a single sentence. Their output will still differ enormously, because the things they ask AI to do are different. Those who choose the right problem get ten times the output. Those who choose the wrong problem merely do the wrong thing ten times faster.

Previously, execution ability could conceal mediocre judgment, because execution consumed most of the time. Now that execution approaches free, judgment is exposed in the results for the first time, with nowhere to hide.

There is no shortcut to judgment, but there is a simple, laborious method. Lay out everything you are working on, prioritize it yourself, and then interrogate that order. Why does this task come before that one? What value does it have for the business? What is the cost of not doing it? The answers come from your feel for the business and your understanding of technology. Both must be acquired firsthand.

Taste: Set Your Own Standards for the Result

AI's output is a product of pretraining. Without guidance, everyone receives the same thing: an 80-point average—functionally correct, indistinct, uniformly proper.

Moving the result from that 80-point average to a specific, distinctive 95-point outcome requires taste supplied by a person.

Here is a real story. A friend who could not code at all used AI to take on her first freelance project. The pages were built, the application ran, and the client accepted it. But one glance at her technical approach showed a mess: runnable but unmaintainable, without structure, assembled one step at a time with no plan.

I did not write a single line of code for her. I only gave AI two more sentences: what approach the backend should use, and where it should be deployed. Those two sentences immediately corrected the project's direction, and AI completed the refactoring on its own.

The functionality was the same. Where was the difference? In those two sentences. Those two sentences were taste.

The philosopher Michael Polanyi famously argued that we know more than we can tell. Taste is a classic form of tacit knowledge. It cannot be captured in a prompt template or transferred to a machine all at once, because it is not a set of rules. It grows out of thousands of rounds of feedback from doing the work yourself, from countless encounters with “this is wrong; that is right.”

Cook Ding in the Zhuangzi says: “What I care about is the Way, which goes beyond skill.” Skill can be handed to machines. The Way, for now, cannot. And the more skill we hand over, the scarcer the Way becomes.

Reasoning: Work Through the Logic Yourself

AI is a creature of probability, so the material it gives you is always abundant, relevant, and seemingly coherent. But between “seems logical” and “has sound logic” lies a road that a person must walk themselves.

AI adds; people must subtract. In a second, it can give you ten reasons, twenty references, and five approaches. Your work is to find the firmest point within this abundance of associations, then reconstruct the logical chain yourself, link by link. If we take this route, what happens first? Where will the bottleneck appear? Is that bottleneck fundamental, or is it an engineering constraint? Build the logical structure once in your own mind. A proposal you cannot build there is bound to collapse when implemented.

One of the friends who wrote to me said she was accustomed to using pen and paper. Before asking AI for a proposal, she would reason it through on paper. If the upstream team gives us code directly, how do we run it? What problems will arise? Where do those problems come from? I especially value this habit, because the process of reasoning is the source of ownership of knowledge.

AI can hand you the blueprints for an entire building in an instant. But if you have not constructed the logical chain yourself, that knowledge is not yours. It resides in your head, but you cannot transfer it to the next problem. Next time, in a different situation, you can only ask AI again.

Wang Yangming said: “There has never been knowing without acting. Knowing without acting is simply not yet knowing.” Five hundred years later, these words have acquired an unexpectedly precise footnote in the AI era: understanding that does not come through your own reasoning and action, however complete it appears, is only someone else's words lodged in your mind.

Think First, Then Let AI Expand the Idea

After discussing these three things, I must honestly face a contradiction, because one of the readers stated it plainly: “If we write less code in the future and only review it, these capabilities will be hard to develop. What do we do?”

She is right. Judgment, taste, and reasoning all come from doing things yourself, yet doing things ourselves is precisely what we are handing over to AI. Use it or lose it: this is a real, structural contradiction. I will not cover it with pretty words. I do not have a complete answer to it.

But she wrote that she had already found an approach. I think it is the best partial answer we currently have: do not immediately throw the problem at AI.

First work through the logical chain yourself, however roughly, and form a high-level idea: what do I want, how many parts does it have, and where are the key dependencies? Then give that idea to AI to expand, refine, and execute. This preserves the thinking process while borrowing the machine's speed. And because you have the blueprint in your mind, checking the result becomes faster too.

She said she had learned this approach from my article “A Claude-Based Reading Workflow.” It was slow, but it kept her from losing herself.

A little slower is fine. When the waves of the times are pushing you along, slowness can sometimes be a discipline.

5. Use AI While Remaining Present Yourself

Man is a thinking reed. — Pascal, Pensées

Judgment, taste, and reasoning are ultimately three capabilities: techniques that help you do better work. Beneath them, though, lies a more fundamental and more easily overlooked question. It concerns not whether you can do the work well, but whether you, as a person, are still present in any of it.

Let us lift our gaze a little beyond engineering.

In the Phaedrus, Plato recounts an Egyptian legend. The god who invented writing, Theuth, presented it to King Thamus and claimed that it would make the Egyptians wiser and improve their memory.

The king rejected it. He said: “On the contrary, this invention will plant forgetfulness in the souls of those who learn it. They will rely on external marks rather than remember for themselves. They will hear of many things without truly learning them. They will appear to know everything while knowing nothing.”

For twenty-four hundred years, every revolution in cognitive technology has summoned the ghost of Thamus again: writing will kill memory, printing will kill authority, search engines will kill learning. Each time, the worry has looked excessive in hindsight.

We outsourced memory, freeing mental capacity for more complex thought. We outsourced retrieval, building systems of knowledge at a higher level. Technological pessimists seem never to have won.

So what is different this time?

The difference is the level of what we outsource. Writing outsourced memory, printing outsourced reproduction, and search outsourced retrieval. What they handed over were the raw materials of thought. This time, for the first time in human history, we can outsource thinking itself: reasoning, weighing alternatives, judgment—the last part of the soul that had not yet been surrendered, the part Thamus feared for.

The mental capacity freed by previous revolutions flowed into higher-level cognitive activities. This time, nobody knows what lies above that higher level.

I have seen what the loss of agency looks like. It is neither dramatic nor even noisy. A question arrives; instead of thinking first, someone forwards it to AI. AI supplies an answer; they send a screenshot back to the person who asked. In this exchange, the person's function in the chain is that of a network cable. A production bug appears, and their manager asks what happened. They say: “I don't know. AI wrote it.”

At that moment, what has been replaced is their presence, not simply their role—although their job may follow. They are still sitting at their desk, but they are already absent.

In his analysis of alienated labor, Marx wrote that the worker does not affirm themselves in their labor, but denies themselves. The frightening thing about alienation has never been merely that the worker eventually loses their job. It is that, long before unemployment, the work has already become something unrelated to them.

The women spinners of the spinning-jenny era first lost their relationship with their work, then lost the work itself.

Today's version is cognitive alienation: before AI replaces us, we first turn ourselves into people for whom agency is unnecessary.

This is the real danger. Foucault warned long ago that discipline needs no guard: we voluntarily trim ourselves into the shape a process requires. Heidegger put it more coldly: “The real danger of technology is never that machines harm people, but that people begin to understand themselves as resources awaiting optimization.”

But pessimism is not what I want to emphasize. Quite the opposite.

Most things do not have a definite answer. Understanding does not fall from the sky. People develop it gradually through reasoning, argument, and conversation.

This process cannot be skipped, because this process is the person themselves.

If understanding is supplied directly by AI, then even setting aside whether it is comprehensive or critical, the absence of the thinking in between is enough to make agency retreat inch by inch. Eventually, a person becomes the figure standing before the spinning jenny, waiting to be made obsolete.

AI cannot replace people in sensing and experiencing the world. This is not a motivational slogan; it has a structural basis. In an earlier article about AI-assisted psychological healing, I wrote that human cognition is more than information processing. It is an experience woven from sensation and emotion. We perceive the physical world through our bodies and give things meaning through our emotions. AI is unmatched at processing information and generating text, but what it lacks is precisely this experiential dimension. It can process data without experiencing the real world that the data represents.

In one sentence: AI's understanding is instrumental, rather than existential.

It can help us understand certain aspects of the human condition, but it cannot replace the insights we gain through lived experience and inner reflection. A painting can imitate nature, but it can never become nature itself.

Logical reasoning, free association, a feel for the business, the desire to explore the unknown—these are not skills to list on a résumé. They are organs of agency. My view is that human thinking should grow in the age of AI, rather than wear away. AI takes over what is determinate. People should always have lived in uncertainty: that is where judgment, taste, and reasoning belong, and where the answer to “what makes us human?” resides.

Pascal said that man is a thinking reed: the universe need not arm itself to crush him. Yet man remains nobler than what crushes him, because he knows he will die, while the universe knows nothing of it. In the AI era, this acquires a new reading: models may surpass us in almost every task, but the one who knows what they are doing, why they are doing it, and how the world will be different afterward is still this reed.

All the reed's dignity lies in thought. Please do not surrender it yourself.

6. Let Your Sense of Worth Depend on More Than Output

What defines us is not “the voyage itself,” but “the meaning we give the voyage.” — “On the Value of Existence and the Experience of Life”

The reasoning so far has taken us steadily inward: from roles to capabilities to agency, each layer deeper than the last. Now we have reached the innermost room, where the true source of all this anxiety sits. I think I already addressed this question years ago in “On the Value of Existence and the Experience of Life.”

The danger to our jobs, the cultivation of capabilities, the survival of agency—all these anxieties converge on the same thing: we fear losing our sense of worth. We fear waking up one day to discover that we are no longer useful to the world. Layoffs are only the outer shell of this fear. At its core lies something older: a person's deepest fear has never been simply being without work, but doubting whether they are still worthy.

Yet your sense of worth is never something the world issues to you. It is something you anchor for yourself.

Many years ago, when AI had nothing to do with my life, I wrote “On the Value of Existence and the Experience of Life.” In it, I wrote that life has no inherent purpose. We arrive by chance and leave by chance; no law requires us to do any particular thing. Even so, we must bravely sail with this purposeless life aboard: “This voyage has no destination, but on the vast ocean, we continually set navigational markers for it.”

What defines us has never been the voyage itself, but the meaning we give it. Existence precedes essence. First we exist; then we search for meaning and give meaning to that existence.

I did not expect that, years later, these words about life would strike a technical problem so precisely.

Because meaning is assigned afterward, because it is a navigational marker you set for yourself, anchoring it contains an ancient trap. If you tie all your meaning to a single external source of feedback, then the moment that feedback stops, the whole structure of meaning can collapse overnight into disillusionment. The example I used then was a workaholic who stakes all meaning on performance and a boss's approval. Once that feedback stops, enthusiasm fades and self-doubt marches in.

The AI era is simply making this ancient vulnerability visible to everyone at an unprecedented scale and speed.

If your sense of worth happens to be anchored in “I can write code,” “I can execute quickly,” or “I deliver more than others,” recognize a harsh reality: you have dropped your anchor into a sea whose tide is receding. There was nothing inherently wrong with that anchoring point. It simply no longer belongs to you. It belongs to machines, and they are visibly reclaiming it.

I wrote another sentence in that old article that now reads almost like a prophecy: “The more you try to be the best at everything, the further you get from wholeness, until life becomes empty fragments.” In sprints where everything is labeled P0 and you are exhausted from rushing, in the habit of filling every saved hour with more requirements, your sense of worth is being diluted little by little rather than accumulated.

So where can your sense of worth find a truly stable anchor? In the place AI cannot take from you.

AI can process the entire world, but cannot experience it. It can simulate every meaning, but cannot truly give meaning, because giving meaning requires the presence of a subject who really lives, really feels, and really cares.

This is the one dependable source of self-worth: it comes from your ability to perceive, to care, and to give what you do a meaning that belongs uniquely to you, rather than from being useful. Usefulness is the price the market quotes you, rising and falling with technology. The other is your own worth, which nobody can take away—unless you first confuse the two yourself.

Many years ago, during a low point in my life, I wrote myself an almost stubborn sentence: whether or not I have those external things, I must believe that I am one hundred percent precious. I always have my own worth, without conditions. Today, I want to offer that sentence again to everyone who feels they are losing value in the face of AI. AI can lower your market price; it cannot touch your worth.

So when AI has taken everything it can, what remains is far more than a second-best refuge. It is where your sense of worth should have lived all along. Like an ebbing tide, AI washes away the external anchors we casually dropped over the years—rank, output, the feeling of being needed—which were never as secure as we thought. It forces us to swim back to the one center that cannot be washed away.

In that old article, I defined this center: the value of our existence lies in establishing this center, and the meaning of living lies in expanding its boundary. The center is what makes you you. The boundary is the world you spend your life experiencing, perceiving, and giving meaning to. AI can approach that boundary indefinitely, but can never reach the center, because sitting there is a person who will die, who can love, and who asks about meaning late at night.

Reclaim your sense of worth from the need to be needed by AI. This is one of the most hidden and most essential lessons this era asks of everyone.

7. Use the Time You Save to Grow

The wind rises from the earth, beginning at the tips of the green duckweed. — Song Yu, “Rhapsody on the Wind”

Let us return to those three letters and the question at the beginning.

This year, friends who know me have called me radical: proposals go to AI, coding goes to AI, the first draft of reviews goes to AI, and next I am preparing to hand over testing too. But few understand that productivity has never been the goal. If AI's entire purpose is to make one person do the work of four, endlessly piling on workload, then what? That merely turns you into a cheaper spinning jenny, saving the balance sheet another two years of costs before you are replaced.

People can never beat machines at efficiency. The end of that road is clearly marked: inevitable replacement.

The real purpose of being radical is only this: keep finding new footing before the larger trend fully arrives.

All the time saved with AI must flow into growth and exploration, rather than more requirements. This is a discipline I have set for myself, and a message I repeat in every reply: if increased workload consumes the entire productivity dividend, this transformation means nothing for the individual.

In the end-state workflow I imagine, engineers have two indispensable positions.

Build the World in Which AI Works

One side is infrastructure: build the world in which AI works. How knowledge accumulates, how context flows, how agents collaborate, how paradigms evolve—this world will not grow by itself. Engineers need to construct it. This transfers engineers' gift for understanding and building systems from business code to a higher level.

Safeguard Quality and the Bigger Picture

The other side is the gatekeeper: serve as the final checkpoint for quality and the bigger picture in the torrent carrying intent toward implementation. Use judgment to choose problems worth solving, taste to set standards for results, and reasoning to protect the logic of a proposal. These are qualities particular to engineers. Our task is to develop them to the fullest and fit them into future workflows.

As I finish, I want to return to the Luddite workers who smashed machines at night.

Their real tragedy was not choosing the wrong target. Machines did crush their livelihoods, and their anger deserves our respect two centuries later. The real tragedy was that, standing at that moment, the only option they could see was destruction. They could see what they were losing, but could not see what they might become. That was not their fault. In that era, nobody could see it.

But we are different. This is the almost harsh good fortune peculiar to our generation: change happens in years, fast enough for a person to witness the end state within their own career, and fast enough to leave nobody the luxury of waiting to see. We are the first ordinary people in history who can participate in a transformation with a sense of its end state.

So where do we go from here?

Return to the question in the first letter: “Will we eventually hand everything over to AI?”

Yes. And sooner than most people expect.

But the answer to that question ultimately depends less on AI than on where you are standing when that day arrives.

Will you stand at the end of the assembly line, stamping approval on the machine's output with ever-lower standards, waiting for even the stamping to be optimized away? Or will you move further upstream—to where questions are set, standards established, logic safeguarded, and worlds built?

Wherever you choose to stand, remember the other side of this transformation. It is a tremendous act of subtraction, but what it subtracts has never been the most precious part of us. It removes the execution and repetition that have long concealed those precious things.

When the tide recedes, what appears on the beach has always been there, merely submerged beneath endless busyness: your judgment, your taste, the logic you have worked through yourself, your perception of and curiosity about the world.

So we no longer need to ask “what do we have left?” in a voice of anxiety. What remains is what is precious. What remains is us.

The wind begins at the tips of the green duckweed. The larger trend is never a vast force descending from the sky. It forms from countless individuals' choices here and now. Today, you decide to work through the logical chain on paper before opening the chat box, to preserve your standards instead of sacrificing them for speed, to invest the time you save in an exploration nobody has attempted before. These small decisions are themselves part of the larger trend.

We are not simply individuals swept along by it.

The wind begins at the tips of the green duckweed. We should be the larger trend itself.

Run forward. Do not look back.