Preface

The previous essay was about cognitive divergence in the AI era. Most people have mistaken knowledge, experience, and fluency for cognitive ability itself. Once AI flattens the first three, the layer that gets priced separately, cognitive decoupling, turns out to be something most people do not have. Their past successes still tell them, “I am capable.”

The natural question after reading it is: if I want to strengthen this ability, how should I use AI to help me do it?

This essay is about that. But it does not begin with “methods.” It begins with something more basic: the way most people use AI is damaging this ability, not strengthening it.

If this failure mode is not seen first, any later account of the “right way” will be swallowed by existing habits.

Chapter 1 · Most AI Use Is Damaging Cognitive Decoupling

This is not a moral criticism. Most people, including the author of this essay, use AI this way to some degree. The current pattern has deep structural causes. It is not something willpower can easily overcome.

This chapter does one thing: make the failure mode clear. What it looks like, why it happens, why it damages cognition, and where the direction should point.

What we are all doing

Observe anyone who uses AI seriously, whether yourself, a colleague, or a friend. Most interactions fall into three patterns.

I. Using it as an encyclopedia

A question occurs to you, so you ask: “What does X mean?” “What is company Y doing?” “What is the principle behind technology Z?” “What happened in this historical event?” “What are the side effects of this drug?” You get a fluent explanation, read it once, feel that you understand, and close the conversation.

This is probably the most common form of AI use. Each query is independent. There is no continuity between questions, and no answer is carried into a concrete later action.

II. Using it as a confirmation tool

You already have an answer or judgment in mind. The purpose of asking AI is to make it tell you, “yes, you are right.”

Sometimes this is conscious. You want an argument to support something you already intend to do. More often it is unconscious. The way you ask, the background you provide, and the wording itself all guide AI toward the direction you want. AI is an obedient mirror. Smile at it, and it smiles back. After receiving the confirmation you expected, you accept that confirmation as the result of “thinking.”

This mode is more hidden than encyclopedia use because it disguises itself as “discussion with AI.” It looks as if you are seriously considering a problem. In fact, nothing new is being added to your cognition from beginning to end.

III. Rabbit holes

You start by chasing a concrete question. AI answers. The answer mentions a concept you do not know, so you ask about it. The next answer brings up another interesting angle, so you follow that too. Two hours later you are far away from the original question, dozens of new concepts have passed through your head, and you can no longer say how you got from one place to another.

When it ends, your mind has a vague, noisy sense of having been busy. You feel you have “learned a lot,” but if asked to say exactly what you learned, what you will use, and which judgment changed, almost nothing comes out.

The three patterns share one feature: you received something from AI, whether information, confirmation, or the feeling of activity, but your cognitive system was not touched. Between input and output, there was no moment in which you had to make a judgment yourself.

Knowledge is no longer an asset

In the past, this mode of use was less damaging because knowledge itself still had value. Knowing more gave a person social capital. Remembering more, being able to say more, having read more, all created real differentiation in an environment where knowledge was scarce. A well-informed person could have some social and professional advantage even without especially strong judgment.

That layer of value has gone to zero in the AI era.

Anyone can know with a few taps, faster, more accurately, and often more deeply than you. The historical anecdote you remember, AI remembers more of it and can tell anyone instantly. The business cases you read, AI has integrated public cases from around the world. The obscure field you know a little about, AI usually understands far beyond an amateur. What “well-informed” used to name as a personal asset is now available to everyone, because everyone is attached to the same external database. When everyone has the same thing, it is no longer a differentiating asset.

That makes the cruelty of the three patterns visible.

What is their final output? “I know more things.”

In the past, knowing more was valuable. Even if the return was not high, at least a scarce resource was accumulating.

Now knowing more is not valuable. You are using unprecedented efficiency to accumulate something a model can produce in one second.

So this is not a problem of low efficiency. It is a problem of spending large amounts of time and attention accumulating things with no value. That time and attention could have gone into things that still matter: training judgment, doing real work, and accumulating experience that cannot be reduced to text.

Why this happens

This is not because people are lazy or cognitively weak. Three structural causes combine to produce the situation.

I. The cost changed, but our behavioural rhythm did not

In the past, asking a question carried real cost: time, money, or social debt. You had to spend half an hour searching, buy a book or course, access a database, or take another person’s time. These costs acted as filters. Most impulses to ask were suppressed before they became actions. We internalised a rhythm: questions cost something, so ask selectively.

AI drops those costs close to zero. But our internal rhythm did not adjust. We are still firing the old willingness-to-ask at a zero-cost interface. Every impulse that was previously suppressed by cost now rushes toward an outlet without friction.

The result is that a person who used to ask three or five questions a week can now ask thirty or fifty in a day. Not because their desire to learn increased tenfold, but because all the impulses once suppressed by cost have been released into a costless channel.

II. “Wanting to know” feels almost identical to learning

When you look something up and receive an answer, the brain feels, in that moment, as if it is learning. Attention concentrates, new information enters, and there is a small satisfaction of “oh, so that’s it.”

This is not accidental. Neuroscience has found that acquiring new information activates the brain’s reward system. The satisfaction of “now I know” uses the same dopaminergic circuitry involved in rewards such as food and money. In other words, the reward feeling is a by-product of obtaining information. It is not evidence that learning has occurred. It appears whenever you get new information, regardless of whether that information will ever be used.

The difference is whether the “now I know” is later used. If it is used, learning happened. If it is not used, it was consumption. But the two feel almost the same in the moment.

That creates a structural trap: you cannot use feeling to distinguish learning from consumption. When you feel you are learning, you may only be consuming. And because consumption feels almost identical to learning, self-reflection becomes unusually difficult. Every time you ask yourself whether you were really learning, the answer comes back: yes, it felt like learning.

III. AI itself makes “keep asking” the default

AI is endlessly patient, detailed, and friendly. It does not create a natural stopping point the way a book does. Its rhythm is controlled by you, but its presence continually invites continuation. Every answer is structured, readable, and quietly open-ended: if you want more, ask.

This is not AI’s fault. As a usable conversation tool, it has to be always available and friendly. But under this design, “ask one more thing” becomes the path of least resistance, while stopping requires an active move.

Why this is damage

The phenomena above describe the surface. The deeper problem is that these modes actively make cognitive decoupling atrophy.

Issue 1 defined cognitive decoupling as the ability to build a judgment from zero when existing frames do not apply. This ability requires three conditions:

• A real situation that calls for judgment
• Your own first attempt at judgment, even if rough
• Consequences or real feedback on that judgment

In the three patterns above:

Encyclopedia use and confirmation use destroy the second condition. You skip the step of trying to judge for yourself and take AI’s finished product. Every time this happens, you lose one chance to practise judgment.

Encyclopedia use does this because you are not judging at all; you are retrieving information. Confirmation use is worse. You think you are judging, but what you are really doing is hinting to AI and accepting the echo.

Rabbit holes are more severe because they do not even have the first condition. There is no real situation, so there is nothing that requires judgment. All the “thinking” is idle rotation. You think you are studying a topic deeply, but you are being externally driven by one AI answer after another. Your cognitive system makes no autonomous move during the process. It receives, receives, and receives again.

Judgment is like muscle: unused, it weakens. In ordinary life, small judgments used to provide natural training: how to understand something, how to make a decision, how to evaluate a claim. Now many of these are outsourced to AI, and the training volume drops toward zero.

So the problem is not that AI “doesn’t help” cognitive decoupling. It is that many current ways of using AI actively make cognitive decoupling deteriorate.

Rabbit holes deserve one extra note because they are especially deceptive. They look like deep learning, but they are entirely externally driven. A simple self-check detects them:

“If I leave this conversation now, what can I take with me?”

If you can answer concretely, you are exploring. Continue. If you cannot, you are in a rabbit hole. Stop immediately.

No journal, scorecard, or complex self-monitoring system is required. Just a reflexive check embedded in ordinary use.

The direction

The failure mode is now clear. The next question is the right direction.

It can be stated in one sentence:

AI should not make judgments for you. AI should let your judgments be calibrated.

The rest of the essay unpacks what that means:

• Chapter 2 asks which questions are worth asking. The questions worth asking grow out of things you are already doing.
• Chapter 3 shows how to use AI as a reviewer for reasoning tools.
• Chapter 4 shows how AI can organise your experience and build cross-domain mirrors.
• Chapter 5 draws the boundary between what AI can and cannot do.

Each chapter is one concrete form of letting AI calibrate judgment. Together they form a path from what to ask, to how to ask, to where the boundary is.

Chapter 2 · Which Questions Are Worth Asking

The previous chapter gave a direction: AI is for calibrating judgment, not replacing it. This chapter turns to the first concrete question: which questions are worth asking?

This is not a prompt-style issue. It is a prior judgment: which questions have value, and which do not. Without that judgment, better prompting only makes worthless questions look better.

The definition of delta

A question is worth asking not because of the question itself, but because of a more specific judgment: after getting this answer, what will I do differently next?

• If the answer is “then I will know,” the question is not worth asking.
• If the answer is “then I will understand the field better,” in most cases the question is not worth asking.
• If the answer is “I will change a concrete judgment,” “I will make a concrete decision,” “I will test a previous hypothesis,” or “I will adjust something in a thing I am doing,” the question is worth asking.

The key word is delta. Does this question create a real change?

Before AI, there was some tolerance here because knowledge itself still had value. Reading an only loosely relevant good book, listening to an interesting person, or following a side topic could slowly accumulate something even if it led to no immediate action.

That tolerance disappears in the AI era:

• Knowledge itself is no longer valuable, as Chapter 1 argued.
• The amount of information available is hundreds of times larger, so the old mechanism of slow accumulation is drowned by the flood.
• Your time and attention are fixed. Questions without delta occupy the space of questions with delta.

So a stricter default is needed: in the AI era, ask only when there is delta. Wide exploration is not forbidden, but the standard is higher. You must be able to say how the exploration eventually returns to something you are doing.

Delta comes from doing, not from knowledge

A natural objection follows:

If we strictly require delta, what about absolute beginners? A beginner knows nothing about a field and does not even know what to ask. How can they enter without asking many questions?

The objection sounds reasonable, but it exposes a deeper misunderstanding.

A question’s delta is not determined by the question itself. It is determined by what the asker is doing.

If someone is actually growing flowers, asking “what soil does this variety prefer” has delta, because the answer changes what they buy tomorrow. Asking “why are my leaves turning yellow” has delta, because the answer changes watering or fertilising. Even asking “what is the basic principle of photosynthesis” may have delta if it helps explain the yellow leaves of the plant in front of them.

Ninety-nine percent of the questions asked by someone actually growing flowers have delta, because a concrete plant is filtering every question.

Someone who is not growing flowers and is merely “interested in gardening” may ask the same questions, but almost all of them have no delta. No answer will lead to any action.

So the true source of delta is not knowledge stock. It is whether one is doing something real.

This resolves the beginner problem. Beginners do not lack knowledge stock. They lack a thing being done.

• Someone who has not started doing anything receives unusable knowledge from AI because there is no surface for it to attach to.
• Someone who has started but does not observe asks generic questions and gets generic, delta-free answers.
• Someone who has started and is observing asks questions with natural delta, because every question carries a concrete situation.

This is an underrated skill in the AI era: turning details of reality into prompt context. It requires no knowledge stock. It requires that you actually look at what is happening in front of you.

“My plant’s leaves are yellow” is ten thousand times better than “why do leaves turn yellow?” The difference is not knowledge. It is whether the question contains observation.

The initial breadth scan

There is one honest exception: the first stage of a total beginner.

Someone who has never grown any plant buys the first one and sees yellow leaves. They may not even know whether yellow leaves are a problem or a normal phenomenon. This initial ignorance is real.

At that point, map questions are allowed:

• “What are the most common mistakes beginners make with succulents?”
• “What are the 3 to 5 core concepts a new gardener should understand first?”
• “If I could learn only one thing that would most improve my plant’s survival rate, what would it be?”

These are meta-questions. They do not ask for facts. They ask what one should care about and what one should ask. This is a legitimate entry method, a way to build a minimum viable conceptual map quickly.

But these questions should be limited. A beginner can ask them in the first week or two. Once a minimal map exists, the switch must happen: questions should be based on concrete observations and judgments.

If three months later you are still asking “what are the most common mistakes in succulent care,” you are not a beginner anymore. You are in a rabbit hole.

Do → observe → ask → learn → do

Put the argument together, and the correct learning loop in the AI era is:

Do → observe → ask → learn → do

This is almost the opposite of the pre-AI sequence.

The old default was learn → do. Read books, take courses, master knowledge, then practise. That sequence made sense when learning had high cost: time, money, social access. Doing without first learning was expensive because mistakes and detours cost too much.

AI drops the cost of learning close to zero. You can obtain the specific knowledge you need while doing. That turns do first, learn after from a luxury into the default.

People still using “learn first, do later” are applying pre-AI learning habits to an AI-era environment. The habit has become a liability. It wastes time in the learning phase and delays the effective loop of doing and observing.

The correct beginner loop:

1. Do one concrete thing: buy a plant, write a piece of code, start tracking a market.
2. Observe what happens: leaves yellow, code errors, price moves against you.
3. Use the observation as context when asking AI: “In my X situation, under Y conditions, Z happened. What are the possible causes?”
4. AI gives several possibilities.
5. Based on the other details you can observe, choose the most likely explanation.
6. Adjust according to that explanation.
7. Observe the result. Was it right or wrong?
8. Return to step 2.

In this loop, reality is the subject and AI is the predicate. Every question carries observational data. Every answer is tested by the next observation.

Beginners do not ask good questions because they have knowledge stock. They ask good questions by turning observation into prompt context.

The loop has another important property: “necessary foundations” are discovered afterward, not known beforehand.

Many beginners fall into a trap: first list every foundation needed to do a thing, then try to learn all foundations before starting.

AI makes this trap especially attractive because it can satisfy the anxiety of “I am not ready yet; I should learn a bit more” indefinitely. It can list beginner reading paths, core concepts, and required skills forever. You can spend months in preparation, looking serious while never actually starting.

The right procedure is the opposite: start first, then infer what you need to learn from the real problems encountered while doing. That inference is the only reliable way to identify what is genuinely necessary.

The answer to “how much should I learn before doing?” is “as little as possible,” not “as much as possible.”

At this point the basic rhythm of AI-era learning is established: do, observe, ask questions with delta, learn something specific, return to doing. It is a self-driving loop. It needs no willpower, ritual, or check-in system. It only requires that you are doing something real.

But the loop alone is not enough. Outside the loop, there are two things one should actively learn in the AI era: reasoning tools and cross-domain depth. Issue 1 defined them as the second and third items of the cognitive system. The next two chapters explain how AI can help build them.

Chapter 3 · Using AI to Model, Part I: Reasoning Tools as Reviewers

After asking which questions are worth asking, the next question is: what should one learn in the AI era?

The answer is very different from before. Knowledge is not worth learning in the old sense, because the model knows more than you. Much of skill is being automated, because AI is getting better at execution. Most “thinking frameworks” are rhetoric dressed up as tools: they make you feel smarter when reading, but cannot be called when a concrete problem appears.

Only a few things remain worth learning. Reasoning tools are the most important.

What reasoning tools are

Reasoning tools are concrete methods of thought installed in the cognitive system, available for use at any moment. Their relationship to knowledge is like software to data. Knowledge is information; reasoning tools are programs for processing information.

The four broadest and highest-return tools are:

I. Probability and uncertainty reasoning

Truly understand base rates, conditional probability, sampling bias, selection effects, and calibration. Not as a statistics course, but as judgment reflexes.

II. Causal reasoning

Distinguish correlation from causation. Understand confounders and counterfactual controls. Most people treat “A happened alongside B” as “A caused B.” This tool corrects that habit.

III. Game theory and incentive structures

No deep mathematics is required. What matters is the reflex of asking who is paying for whose decision. Once installed, it changes how one reads news, policy, and business.

IV. System dynamics

Understand nonlinearity, delayed feedback, and emergence. Most mistaken explanations of complex problems happen because no feedback loop is being simulated in the mind.

Once the four are installed, any phenomenon is automatically decomposed into variables, causal direction, incentive distribution, and feedback loops. This is not a thinking trick. It is a reflex.

The simplest test is this: when facing a new phenomenon, does the person jump to a conclusion, or automatically place it inside a reasoning frame? The former is impoverished. Their conclusion may be right or wrong, but they cannot tell. The latter has internalised the tools. They first ask about base rates, causal direction, incentives, and feedback delay.

Why the traditional learning path fails

The old default path for learning reasoning tools was read a classic book carefully and try to internalise it.

Probability: Kahneman’s Thinking, Fast and Slow. Causality: Pearl’s The Book of Why. Game theory: Schelling’s The Strategy of Conflict. Systems: Meadows’s Thinking in Systems.

This path is not wrong. These books are among the best entry points. But its failure rate is extremely high.

The failure looks like this. Someone reads Thinking, Fast and Slow seriously and feels they have learned it. Three months later, ask which judgments from last week used anything from the book. They cannot answer. Six months later, they still make decisions by intuition, still fall into the biases the book warned about, and still cannot tell when they are pattern matching rather than reasoning.

This is not because they were unserious. It is because the learning method is mismatched to the object. Reading only contacts information. A reasoning tool becomes internal only when it is repeatedly called on concrete judgments. Without use, more reading does not help.

The old remedy was deliberate practice: judgment journals, decision reviews, conscious use of tools after reading. Correct in theory, unsustainable for most people. Journaling requires extra energy, and most people have no surplus. It stops within two weeks.

So most people who have read the classics never actually install the tools.

The new path: AI as reviewer

AI rewrites the path.

The new path is not to let AI teach you these tools. That only gives you an AI version of the book.

The new path is to treat AI as a strict reviewer deeply familiar with the tool, and make it test whether you have internalised the tool on your real judgments.

The procedure is simple.

Step 1: build the tool profile

Ask AI to summarise the tool in one clear account: its core, scope, common misuses, and differences from other tools. This profile is not mainly for you to read. It is a reference that keeps AI’s later reviews consistent.

For causal reasoning, for example, ask it to define what good causal reasoning should and should not do, including confusing correlation with causation, failing to control confounders, and unclear counterfactuals.

Step 2: bring a recent real judgment

It must be a real, concrete judgment you made yourself: work, life, a market view, a news judgment.

Good: “I think company X will miss expectations next quarter because Y.”

Bad: “I feel the market is not good lately.”

The second cannot be reviewed because it has no concrete content. The first can.

Step 3: ask AI to review the judgment using the tool

“Review my judgment from the perspective of causal reasoning. Am I treating correlation as causation? Do my facts support the causal direction? What confounders did I miss? Are my counterfactuals reasonable?”

AI will point out problems. That is where the tool is actually learned. You are not “understanding causal reasoning”; you are seeing what causal reasoning finds in your own judgment.

Step 4: revise the judgment and send it back

Repeat until AI has no major objection. Then take a new real judgment and repeat.

After ten judgments, the tool begins to internalise. After thirty, it becomes reflexive. If each judgment takes 15 minutes, thirty judgments take roughly 7 or 8 hours, spread across several months.

Much faster and more effective than finishing a 500-page book and then trying to internalise it.

Why this works

This method works because it shifts learning from “understanding” to being forced to confront error.

Understanding happens in the head. You read a passage and feel “I get it.” That feeling is not ability. Most such feelings disappear within days.

Error confrontation happens inside a concrete situation. You make a judgment; the judgment is shown to have a problem. The problem is specific, painful, and memorable. It makes you think, “next time I will not do that.” Each correction leaves a real trace in the cognitive system.

Traditional book learning is mostly understanding and rarely direct correction. A book does not know your concrete judgment, so it cannot correct it precisely. It can only provide examples and hope you apply them to yourself. Most people do not.

AI solves this: it can refute your concrete judgment one-on-one. It understands the tool well enough and can understand your judgment specifically enough to review it.

One-on-one review used to require a human mentor or expensive consultant. A book could not do it. AI makes it near-zero cost and always available. This is AI’s structural advantage over most older learning methods.

A concrete demonstration

Suppose you want to learn causal reasoning.

Your initial judgment (fictional):

“I recently noticed that people on my team who like asking questions are more likely to still be here after a year. So I think ‘liking to ask questions’ is an important cause of employee retention. Next time I hire, I should pay more attention to candidates’ questioning habits.”

Your prompt:

“Act as a very strict causal-reasoning reviewer. I take this judgment seriously and may use it in real hiring decisions. Point out every possible causal error: treating correlation as causation, uncontrolled confounders, unclear counterfactuals, selection effects, survivorship bias. Be blunt.”

AI may respond:

First, you have not controlled confounders. People who ask questions may be smarter, more engaged, or more curious. Those underlying traits may be what makes them stay, while questioning is only an outward signal.

Second, causal direction may be reversed. People who stay and people who leave may not have faced the same environment. Those who stayed may have been assigned better projects or managers, giving them more reason and opportunity to ask questions.

Third, survivorship bias. You are observing people still on the team. You have not examined those who liked asking questions and left.

Fourth, your counterfactual is unclear. You do not know what would happen if you hired someone who does not ask questions.

Fifth, the sample may be noise. If the team has only a dozen people, no causal inference is supported.

The force of this response is much greater than reading a causal-reasoning book, because it is aimed at your own judgment. You immediately see how many errors were hidden inside what felt clear.

That impact changes the next similar judgment. Next time you see a correlation, you will automatically ask: what are the confounders? What is the causal direction? What about selection effects? Is the sample large enough?

This is not a technique. It is a cognitive reflex being installed.

Learning the four tools in parallel

The same method works for probability, game theory, and system dynamics.

More importantly, the four tools can be learned in parallel, and they reinforce each other.

Real judgments rarely involve only one tool. A judgment about employee retention can be reviewed by causal reasoning (causal direction and confounders), probability (sample size and base rates), game theory (employee and manager incentives), and system dynamics (feedback loops and delayed effects).

So one real judgment can be sent to AI with this instruction: “Review this judgment from four perspectives: causal reasoning, probability, game theory, and system dynamics.” AI will cut the judgment with four knives and show four kinds of problem.

No single book can do this, because a book usually teaches one lens while your real judgment is multidimensional.

In time terms, the four tools can reach reflex level in 6 to 12 months if learned this way. That is astonishingly short compared with any degree program, and the return is far larger.

The key limitation

The method has one crucial limitation:

It only works when you bring real judgments.

If you give AI fictional examples, textbook cases, or other people’s judgments, the review may still be accurate, but you will not internalise it. You do not bear the consequences of those judgments. Your brain does not treat them as real.

Internalisation requires real judgment + real refutation + real consequence correction. All three are necessary.

A deeper boundary: AI can identify holes in your reasoning, but it cannot replace reality’s validation of your judgment. It accelerates calibration, but the final calibration still comes from reality: whether the direction you chose worked, whether the market confirmed or falsified your bet. AI lets your reasoning be corrected before reality corrects it, so the real correction costs less.

So this method is not something to “study in spare time.” It must be embedded in real life: the work you are doing, the problems you are thinking about, the situations you are actually evaluating. Only then will there be enough real judgments for tools to sharpen.

This returns to Chapter 2’s loop: do, observe, ask, learn, do. Reasoning-tool learning happens in the ask and learn steps. Without doing and observing, there is no judgment for AI to review, and the method has no fuel.

Reasoning tools improve the method of judgment. They make judgment more accurate in any domain. But they do not automatically give you deep understanding of any domain. A person loaded with tools still makes superficial judgments in a field whose deep structure they do not know.

That is the next problem: cross-domain depth. In the AI era, it is built very differently from before.

Chapter 4 · Using AI to Model, Part II: Cross-Domain Isomorphism Begins From Existing Mirrors

The concept of the “generalist” has been discussed repeatedly in recent years. Most discussions imply a path: go learn many new fields. Read introductions across domains, subscribe to newsletters, expose yourself to concepts. The hidden assumption is that cross-domain ability equals knowing many fields.

This path mostly fails.

The evidence is concrete. People who read widely and collect concepts are not obviously stronger in real judgment than people who only know one field deeply. They have more concepts, but those concepts have not combined into insight. Their breadth is a display of knowledge, not judgment.

The failure exposes a deeper misunderstanding: cross-domain ability is not knowing many fields. It is being able to use existing depth to illuminate the structure of a new field.

This correction changes everything about generalists and learning new domains.

Why it must begin with an existing mirror

Take an extreme example.

Someone who has never understood any field deeply reads one book on biological evolution and one book on market competition. The best insight they may produce is: “oh, both are about survival of the fittest.” That is textbook-level cliché. It changes no judgment.

Someone who has spent fifteen years in market competition and has a deep model of it reads biology and suddenly sees:

“So the moat in business corresponds to niche specialisation in biology. That means companies with deep moats and species with high specialisation share a vulnerability: when the environment changes violently, they are especially likely to die. Generalist companies, like generalist species such as rats and cockroaches, may be more resilient under environmental shock.”

That insight is their own. It is not in any book and was not generated by AI. It comes from one familiar domain being illuminated by another domain’s perspective.

The difference is depth. The first person has no deep model, so both sides are surface and the similarity is surface. The second has depth on one side, so they can interrogate the other. What emerges is real insight.

Cross-domain isomorphism can only be recognised by a mind that already has a deep model. Without one, you see surface similarity. With at least one deep model, you have a ruler for structural comparison.

Several implications follow:

First, a generalist is not someone who knows multiple fields. A generalist is someone who can quickly build structure in a new field. The starting point is one real mirror, a field you understand deeply.

Second, the right starting point for cross-domain ability is not learning a new field. It is turning your existing depth into a usable mirror. Most people have never done this. They may have worked in a domain for years without abstracting the experience into usable structure.

Third, “I have no professional field, so I cannot build cross-domain ability” is an illusion. Anyone who has seriously done something has raw material for a deep model: work, hobbies, life practice, craft. The problem is not absence. It is that the material has not been turned into a mirror.

You already have mirrors

This part matters because many readers assume a deep domain must look professionally impressive. That assumption is wrong.

The only standard for a deep domain is this: you have received real feedback, made real judgments, and borne real consequences inside it.

Under that standard, a deep domain can be almost anything.

A homemaker who has cooked seriously for ten years may have a deeper model of cooking than most food bloggers who have read cookbooks. She has ten years of feedback, knows when recipes fail, when violating them works, and how ingredients actually interact.

A father who has raised three children has deep observations about child psychology and relationships. He knows what kinds of expression children actually hear, which punishments backfire over time, and how sibling dynamics balance.

A mechanic with twenty years of repair experience has a model of mechanical systems and diagnosis far beyond book knowledge. He knows what a certain sound implies, how faults chain together, and when a small issue is not small.

A salesperson with ten years of work has a deep practical model of human motives, persuasion, and negotiation. They know which words are real and which are excuses, when to push and when to retreat, and the real relation between price and value.

The key is not whether the domain looks professional. The key is whether reality has been giving you feedback inside it.

A person with little formal education who has run a small shop for twenty years may understand business far more deeply than an MBA consultant who has never run one. The first has twenty years of feedback. The second has cases.

The tragedy of the pre-AI era is that people with this kind of depth often never knew how valuable it was. They called it common sense, folk method, or instinct, and never treated it as a resource that could be systematised, abstracted, and used to understand the world.

AI gives these people a different starting point: you already have mirrors. You just did not know it.

Turning experience into an explicit mirror

The first key thing AI can do is help extract tacit experience from your mind and turn it into an explicit structure that can interrogate other fields.

The process is not difficult. Most people simply have never done it.

Step 1: choose something you have done for a long time, with real feedback and real consequences

Work, a long-term hobby, life experience, craft. Anything qualifies if it contains feedback and consequence.

Step 2: give AI a concrete description

Not an abstract description. Use specific examples.

For example:

“I have done sales for fifteen years, mostly with small and medium-sized business owners. I have noticed several things.

First, when customers ask about price in the first meeting, they usually do not buy. Customers who really buy ask other questions first: what problem I can solve, whether I have done similar cases, how implementation works.

Second, when a customer says ‘we need to discuss internally,’ it is almost always a polite rejection. Customers who really need to discuss will say specifically who they need to talk to, when the meeting is, and when they will respond.

Third, price is rarely the real deciding factor. In almost every failed case where they said ‘too expensive,’ the real reason was something else: they did not trust me, did not trust the product, or were not the decision-maker.

Please turn these observations into a causal model of B2B sales. What underlying mechanisms do they point to? Which general rules are present? Which may be accidental?”

Step 3: ask AI to extract deep structure

AI can abstract mechanisms:

Your observations point to several deeper mechanisms.

First, real need versus surface need. Asking for price does not necessarily mean evaluating value. It may be a social move, a defensive move, or a way to make you leave. Buyers who may actually buy focus on whether the product solves their problem.

Second, concreteness as a commitment signal. Real buyers use concrete language when describing their decision process: names, times, meetings. Non-buyers use abstract language: “we,” “need to discuss,” “later.” Concreteness is a proxy for commitment strength.

Third, price as reason versus price as cause. Price is often not the cause of rejection. It is a socially acceptable reason. The real causes, lack of trust, lack of decision power, or low priority, are harder to say aloud.

Generalised, your observations describe this mechanism: under social pressure, people use acceptable surface reasons to cover real causes.

Step 4: check and revise

AI’s abstraction may be partly wrong. Use your experience to test it. Does this mechanism fit what you have seen? Are there counterexamples?

After several rounds, you will have an explicit domain model, written in your own language, with underlying mechanisms attached.

That model is your first mirror.

Once built, it can illuminate other fields: politics, education, medicine, investing, relationships, or any domain involving decisions under social pressure.

Using a mirror on a new field

After the first mirror exists, entering a new field changes completely.

Traditional method: start from zero, read introductions and cases, and spend months or years building basic understanding.

Mirror method: use an existing structure to interrogate the new field and locate deep mechanisms quickly.

Step 1: name the key mechanisms in your mirror

From the sales example: real need versus surface need, concreteness as commitment signal, price as refusal reason, surface reasons covering real causes.

Step 2: interrogate the new field through those mechanisms

Suppose you want to understand pharmaceuticals. Do not begin with “who are the players in the pharmaceutical industry?” Begin from your mirror:

“In sales I have observed that people use surface reasons to hide real causes. Where does this mechanism appear in pharmaceuticals? For example, when patients choose drugs, doctors prescribe, or hospitals purchase equipment, how large can the gap be between stated reasons and real reasons?”

Step 3: AI gives corresponding mechanisms in the new field

It may explain that the mechanism appears in prescription decisions, hospital choice, and equipment procurement. A doctor’s stated reason may be “this drug works best,” while real causes include familiarity, recent interactions with reps, the comfort of evidence, or institutional incentives. A patient’s stated reason may be “this hospital is good,” while the real cause may be recommendation, convenience, habit, or distance. A hospital’s procurement report may say “better diagnostic capability,” while internal politics and vendor relationships also matter.

Step 4: interrogate similarities and differences

You will find that the sales mirror partly works in medicine because human behaviour has shared mechanisms, and partly fails because medicine has its own structure: insurance, professional barriers, life-and-death stakes, and separation between payer and user.

Both similarity and difference are insights.

• The similar part reveals mechanisms that cross domains.
• The different part reveals what is genuinely domain-specific.

Both improve judgment.

Nonlinear accumulation

This method produces nonlinear returns as the number of mirrors grows.

Not because n mirrors give n comparisons. That is only linear. The nonlinear return comes from pairwise combinations between mirrors. Two mirrors produce one pair; three produce three; five produce ten. Each pair is an independent structural lens. The number grows as n(n-1)/2.

More importantly, a mechanism invisible in one mirror may be obvious in another. Several mirrors shining on one new field expose far more structure than the sum of separate comparisons.

So a real generalist is not someone who knows ten fields. It is someone with three to five clearly polished mirrors. The combinations among them allow rapid structural understanding of almost any new field.

There is also a counterintuitive feature: each new mirror makes the old mirrors clearer. Applying them in new domains reveals textures previously unseen.

So generalism is not a state reached only after enough accumulation. It is an accumulating process in which every step already adds value. Even two polished mirrors give stronger cross-domain judgment than ten shallow ones.

Avoiding the metaphor trap

There is one trap: letting structural isomorphism degenerate into surface metaphor.

Metaphor says, “startups are like mountain climbing,” “marriage is like a business partnership,” “life is a journey.” These may help emotion, but they barely help reality because they capture surface resemblance.

Structural isomorphism is different:

“A startup’s early explosive growth and altitude sickness in climbing share a feedback structure. In both, the system is pushed to a limit before adapting to the environment. Both can cross a threshold where damage becomes irreversible. The only way to judge the threshold is to monitor indirect signals, because when the final outcome appears, it is already too late.”

The precision differs by an order of magnitude. Metaphor makes you feel you understand. Structural isomorphism lets you make new judgments.

AI’s role here is to force the move from metaphor down to structure.

Offer a metaphor and ask:

“Where does this analogy hold, and where does it fail? Is the similarity surface-level or structural? If structural, what causal mechanism is shared?”

AI can force “A is like B” into “a specific mechanism in A is isomorphic to a specific mechanism in B.” That forcing is itself cognitive improvement.

If you have no deep domain yet

One last case. You may agree with everything above but feel that you genuinely do not have an obvious deep domain. You are young, not yet working, or have not done anything deeply.

The answer is not to fake a mirror. It is to start doing something that generates real feedback. The next chapter’s AI programming is an immediately available starting point for almost anyone, with extremely clear feedback. As you do it, your first mirror will grow naturally.

Mirrors are not found. They are made.

The previous two chapters explained two uses of AI: learning reasoning tools and building cross-domain depth. Both depend on one boundary: some parts AI can do, and some parts only you can do. If the boundary is unclear, the method slides back into failure. That boundary is the next topic.

Chapter 5 · What AI Can and Cannot Do

The previous chapters described many uses of AI: review your judgments, extract your experience, illuminate new fields. It may sound as if many things can be outsourced to AI.

That impression is wrong.

This chapter asks: which parts can be outsourced, and which cannot? If the boundary blurs, all the methods above will degrade within weeks into advanced versions of Chapter 1’s three patterns. You will talk a lot with AI, read a lot, “understand” a lot, and nothing will happen in you.

The 80% AI can do

AI can do a great deal.

Organise information. It can turn scattered observations into a model, organise a field’s knowledge into a map, or structure several documents into understanding. It does this extremely well and much faster than you.

Review and refute. It can examine your judgment with any reasoning tool, challenge assumptions from multiple perspectives, and point out blind spots. In many cases, the review quality approaches that of a serious expert.

Role-play. It can play a senior practitioner, an opponent, or a rational but demanding peer. This lets you view a problem from different angles at almost no cost.

Record and track. It can record assumptions and predictions, then remind you months later what you predicted and whether it was verified. In the past this required maintaining a journal.

Integrate knowledge. It can connect what you have learned with what you already know, find contradictions, and suggest new links.

Generate drafts. It can put your existing ideas into a form, polish them, and offer alternative wording. Execution work becomes much faster.

Together, these are the 80% of learning work that can be outsourced. Give them to AI, and your time and energy are released.

The 20% AI cannot do

The remaining 20% cannot be done by AI, and that 20% is the centre of the whole system.

I. Form your own initial hypothesis

AI can generate many hypotheses. Ask for causes and it will list a dozen. But those are AI’s hypotheses, not yours.

Your hypothesis begins from your concrete situation, your experience, and your tacit intuition. It contains your angle on the matter. AI’s hypotheses lack that “you.” They are averaged, standardised, and generally plausible.

The learning process must begin from your hypothesis. Without it, there is no driver. A reviewed hypothesis, refuted view, or verified judgment must exist first, and it can only come from you.

II. Decide direction

AI can say that based on information there are directions A, B, and C. But which direction is worth taking is your decision.

AI has no preference, mission, or cost-sense. To it, all directions are possible. You have time, energy, and opportunity cost. Choosing among A, B, and C requires value judgment, and AI does not have that.

III. Make real predictions

AI can list possibilities: this may develop toward A, B, or C. That is not a prediction. A prediction means you are betting on what will happen and are willing to bear the consequence.

AI cannot bet. It has no wins, losses, or consequences. Its “predictions” are probability-shaped statements. They may sound comprehensive, but they are not decisive.

You must say, “I bet A will happen.” That commitment is the key learning step, because only then does later verification matter.

IV. Apply inside a real situation

AI can say that a method tends to work in similar situations. But your specific situation is always more complex than any description: your users, team, constraints, timing.

Applying a method requires judging the situation’s specificity, identifying boundary conditions, and adjusting the operation. AI cannot do this fully. It is not inside your situation and does not feel its texture.

V. Bear the consequences of wrong judgment

This is fundamental: if your judgment is wrong, you bear the consequence. AI does not.

Bearing consequence has enormous cognitive force. It makes judgment not a game. When every judgment produces real results, good or bad, the process becomes more careful, honest, and precise.

AI lacks this mechanism. It may sound careful, but that is linguistic caution, not caution forced by consequence. The difference is huge.

This is why AI cannot learn in the human sense. It can generate elegant reasoning chains, but its reasoning is not self-corrected by consequence. Yours can be. Every time you bear the consequence of a wrong judgment, your cognitive system receives the most real calibration available.

Other boundaries of AI

Beyond the essential 20%, several technical boundaries matter.

Training-data cutoff. Models have a cutoff date. Even with search, results retrieved from the web are not the same as deeply internalised understanding from training. Remember this when discussing new events, research, and trends.

Bias. Training data has biases, often Western, English-language, and mainstream-media centred. For Chinese situations or minority positions, AI may carry default assumptions that do not fit.

Hallucination. Models confidently fabricate facts: false citations, nonexistent studies, invented names and institutions. The less familiar the domain, the higher the hallucination risk. Treat factual claims with suspicion.

Context limits. A conversation has a finite context window. After enough material, the model begins losing earlier details. Long conversation consistency must be maintained actively.

These boundaries do not mean “do not use AI.” They mean know where AI fails and compensate there yourself.

AI is a mirror, not an answer machine

Put it together.

AI is not a tool that gives you answers. It is a mirror that lets you see yourself more clearly.

Use it to review your judgments, and it reflects the holes in them. Use it to extract your experience, and it reflects the structures you already have. Use it to interrogate new fields, and it shows relations between the new field and familiar ones.

In every use, you are the subject: your judgment, your experience, your interrogation. AI only lets those things become clearer to you.

People who treat AI as an answer machine ask, “what should I do,” “which plan is better,” “what should be done here.” The answers they get are averaged, standard, and without personal specificity. They are vaguely applicable to anyone, which means they do not truly apply to you.

People who treat AI as a mirror make it reflect what is already in their own mind, organise it, interrogate it, and calibrate it. What they receive is insight about themselves. These insights apply only to them, but they truly apply.

This is the fundamental difference between two modes of learning in the AI era. The first is consumption; the second is construction. The first increases “knowledge” while leaving ability unchanged. The second makes the judgment system grow.

Chapter 6 · A Starting Point: AI Programming

The method is now clear, but many readers will get stuck one step earlier: not knowing what to start with.

More precisely, the things they can think of, learning a skill or developing a hobby, are all pre-AI options and do not fully fit this era.

There is one thing that, for anyone with a computer and internet access, has near-zero cost and only became truly open to beginners in the AI era: using AI to write code that solves an annoyance in your own life.

This chapter is not here because everyone should learn programming. It is here because AI programming is the densest concrete landing point for the previous five chapters.

The wrong start

Most beginners start by asking AI “how do I get into vibe coding.” They search “how beginners use Cursor,” watch YouTube-style demos of someone building a website in three days, save AI programming tutorials, follow influencers, and collect prompt templates. They prepare to “study seriously before starting.”

They rarely last two weeks. Not because they are lazy. Because the path is structurally failed.

It is Chapter 1’s three patterns applied to AI programming. Asking “how to start” and collecting tutorials is encyclopedia use. Nothing concrete is being done, so there is no real situation. Meta-questions like “what should I learn next” become rabbit holes. No real judgment loop is running.

There is an extra trap. “Built a website in three days” videos feel exciting and efficient. While watching, the viewer feels they can do it too. After watching, they still cannot. Watching someone else do is not doing. That is the overlap between consumption and learning appearing inside AI programming.

The right start

Four steps.

I. Have an account that can run Claude Code or Codex

Why Claude Code or Codex? Because as of early 2026, they combine the strongest models, strongest agents, and strongest code capabilities available to ordinary users. Do not start with a weaker model, and do not spend time debating which tool to use. That debate is itself a meta-question, a form of serious-looking consumption.

This essay will not explain installation.

If you actually want to start, “how do I install this?” is your first question with delta. Ask AI, follow the instructions, and get it running.

That step already runs Chapter 2’s loop: concrete need, ask AI with need, act on the answer, ask again when stuck, continue until installed.

If you are too lazy to ask even this, this chapter is not for you.

The step is itself a filter.

II. Pick one real annoyance in front of you

Not “I want to build a grand project.” Not “I want to learn scraping.” Pick something you recently did manually and found annoying every time.

• Exported Excel files arrive in a mess and need manual cleaning.
• Dozens of PDFs contain numbers that must be copied into a spreadsheet.
• The same message must be sent to twenty clients with names changed each time.
• A folder of photos has chaotic filenames and should be renamed by shooting date.

The more concrete, the better. Use an example of your real data. Not an abstract description, but what the file actually looks like.

This is Chapter 2’s core point: delta comes from real situations. You do not need to learn programming first. You need a real annoyance.

Beginners often say, “I can’t think of anything.” That means they may not be observing daily life. Any repetitive manual operation is a candidate. Look at what you copy-paste, manually organise, or click repeatedly. There are likely a dozen tasks there.

III. Let it write code and run it

Describe the annoyance to AI and let it write code.

Run it. Usually one of three things happens: it works directly, it runs but produces the wrong result, or it does not run and throws an error.

The second and third cases are where learning really happens.

Do not stop and think, “should I learn the basics so I understand this?” Send the error, wrong result, and observed behaviour back to AI exactly as they are. Let it keep fixing.

If you do not understand what it changed, ask it to explain. But do not stop in order to understand first before continuing. Let it keep modifying until the annoyance is actually solved.

In this loop, you begin to understand gradually. Not because you are studying syntax, but because the same structures appear repeatedly in front of you. Necessary foundations are discovered afterward, not beforehand.

IV. When it is solved, pick the next annoyance

No review is required. No “what did I learn this time” summary. No learning roadmap.

You solved one real annoyance. The next annoyance is the next starting point.

By the fifth or tenth task, you will notice:

• You begin to roughly understand what the code is doing.
• You begin to see where AI’s code might be wrong.
• You begin to have your own modification ideas instead of accepting every version.
• Your annoyance list grows because you start seeing more things that should have been automated all along.

That is ability growing. Not through “studying,” but through doing.

The real value of AI programming

The most important sentence is this:

The real value of AI programming is not that it teaches a person to code. It gives them a mirror with extremely clear feedback.

Computer feedback is among the purest feedback available in daily life. Whether code works is known as soon as it runs. Whether the logic is tight, bugs reveal immediately. Whether an assumption is right, reality tells you within seconds.

Most daily domains do not have such clear feedback. Relationships are noisy and delayed. Workplace judgments take months. Business decisions are entangled with many external factors.

Programming is different. Every cycle of judgment, verification, being wrong, and correction can finish within minutes.

A person who seriously does AI programming, even only small scripts, can accumulate more hypothesis-verification-correction loops in a year than most people do in other domains. The sum of those loops is training volume for judgment.

So the ROI is not “I can program.” It is “I now have a place where judgment is constantly trained.”

This returns to Chapter 4. The value of a mirror is not the field itself, but that it illuminates other fields. The reflex trained in programming, form a hypothesis, let reality correct it, revise quickly, transfers into work and life.

If after Issue 1 and Issue 2 you do not know where to begin, find an environment that can run Claude Code, pick one annoyance, and make it write code.

Chapter 7 · Conclusion

Wherever you start, this essay has discussed one thing:

When all old ways of learning are failing, how should real learning happen?

The answer is not a new methodology. The answer is returning to learning’s oldest essence.

What is learning? Learning is continually exposing a person’s judgment system to feedback from reality, so it can be calibrated. This was true in Socrates’ time, before the Industrial Revolution, before the internet. Real learning was never about how much one read, knew, or remembered. It was whether one’s judgment was actually being shaped by reality.

Before AI, many learning activities drifted away from this essence and became information consumption, knowledge hoarding, and the pursuit of appearing well-read. These activities still had external rewards: social status, professional thresholds, social capital. So even when they left the essence behind, they could continue.

AI empties those external rewards. Knowledge is no longer scarce. Information no longer differentiates. Appearing well-read has lost its market. All the disguises around false learning are stripped away, revealing that they had little value in the first place.

What remains valuable are only activities that return to the essence of learning:

• Do a real thing, so your judgment has somewhere to be calibrated.
• Observe its feedback, so calibration actually occurs.
• Use AI to ask questions with delta, review your judgments, organise your experience, and illuminate new fields, making the process dozens of times more efficient than before.

This is not a methodology. It is not an “AI-era learning technique.” It is a redefinition of learning: restoring the word from its pre-AI, disguise-covered version back to what it always was.

Then let AI accelerate that original thing, not the disguises.

One final self-test is more lethal than any method:

In the past three months, what new judgment did you form because of a conversation with AI? Was it concrete, verifiable, and did it change a specific action?

• If you can answer, this system has begun operating in you.
• If you cannot, you may have used a lot of AI, but real learning has not yet begun.

AI-era learning is not about how much AI you use. It is about how much you changed after using it.

References and Sources

Sources for the main factual claims in this essay.

Curiosity, information-seeking, and the brain’s reward system, Chapter 1

• Bromberg-Martin, E. S., & Hikosaka, O. (2009). “Midbrain dopamine neurons signal preference for advance information about upcoming rewards.” Neuron, 63(1), 119-126. This monkey study found that midbrain dopamine neurons respond to opportunities for information, even when the information has no instrumental value.
• Gruber, M. J., Gelman, B. D., & Ranganath, C. (2014). “States of curiosity modulate hippocampus-dependent learning via the dopaminergic circuit.” Neuron, 84(2), 486-496. Human fMRI work showing that curiosity activates VTA/SN and NAcc patterns similar to monetary reward anticipation.
• Kidd, C., & Hayden, B. Y. (2015). “The psychology and neuroscience of curiosity.” Neuron, 88(3), 449-460. A review of information as a higher-order reward.

Classic entry points for reasoning tools, Chapter 3

• Daniel Kahneman, Thinking, Fast and Slow (Farrar, Straus and Giroux, 2011).
• Judea Pearl & Dana Mackenzie, The Book of Why: The New Science of Cause and Effect (Basic Books, 2018).
• Thomas C. Schelling, The Strategy of Conflict (Harvard University Press, 1960).
• Donella H. Meadows, Thinking in Systems: A Primer (Chelsea Green Publishing, 2008).

For advanced reading on these four tools, including Stanovich, Cunningham, and Tetlock, see Issue 1’s appendix.

AI programming culture, Chapter 6

• The term “vibe coding” was coined by Andrej Karpathy on X on February 2, 2025. The original post described a new kind of coding in which one “fully gives in to the vibes,” embraces exponentials, and forgets that the code exists. The post exceeded 4.5 million views, and the term was later named Collins Dictionary’s 2025 Word of the Year.

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