Complete English translation.

Chapter 1 · Four Phenomena

If one audits a year of one’s own information consumption, several patterns appear at once. Any one of them on its own is unremarkable; read together, they point at a structural fact that has not been discussed enough.

Four phenomena, then.

I. The rumination chain

Most of the AI-related content you see on Instagram, Substack, or X each day is produced like this. A piece of primary material, whether a paper, a snippet of code, an experiment result, or some practitioner’s firsthand observation, is summarised once by an account using an LLM. Then a marketing account picks it up and gives it a hook-shaped headline. Then the next account’s LLM rewrites it in its own voice. Then it is pushed to the end reader.

You’ve seen the headlines: “You’re using ChatGPT wrong,” “Gemini just cooked GPT-5,” “The white-collar bloodbath has begun,” “Sam Altman just confirmed it,” “How this 19-year-old is making $100K a month with AI,” “Silicon Valley is quietly panicking,” “This time is actually different.” Each one manufactures urgency; click through, and you find a few hundred words of nothing.

This is the rumination chain. At every layer, three things happen at once. Information density drops, as details, caveats, and uncertainty are eaten by summarisation. Error accumulates, because every pass introduces small distortions that reinforce each other over multiple layers. Emotional intensity rises, because every handler has to crank up the drama to survive algorithmic distribution. What arrives at the end reader is something processed, nutritionally depleted, but easier to swallow: a prefabricated junk food of the content world.

The chain used to be entirely human, and its speed and volume were nothing like today’s. Now we’re watching a vast quantity of ruminated garbage being reinvented and manufactured at scale.

When people think they’re “taking in information” every day, what they’re taking in is the tail end of this chain, probably three or four layers removed from the primary source. This fact is not itself news; it’s been said many times. It’s the premise for the three layers that follow.

II. The tooling-up of ruminated consumption

The more arresting thing to watch is not the rumination chain itself. It is the consumer’s response.

Faced with an infinite supply of ruminated content, a subset of people responds not with “consume less, find a primary source” but with “I need better tools to consume more.”

Take OpenClaw. Its underlying capability goes well beyond this, but its most virally shared use case is not “help me write code” or “help me analyse data.” It is “help me summarise what happened in the AI world today.” Users enthusiastically show off their workflows on social media: AI reads dozens of sources for me each day, auto-generates summaries, pushes key updates to my phone.

On the surface this looks like an efficiency gain. It is in fact a strange self-reinforcing loop: an AI tool helping me consume, more efficiently, the rumination another AI has spat out. The throughput of actively eating ruminated matter has itself become a publicised achievement.

The reader’s position has shifted subtly, from passively receiving rumination to actively optimising the ruminated-consumption pipeline. What looks like a step up in agency is, at bottom, an amplification of consumption volume.

III. The producers split

Meanwhile, in a different and barely overlapping circle, a different group of people is doing something entirely different.

Andrej Karpathy has described this split precisely on X: judgments of AI capability are extremely bimodal across populations. On one side are people who mainly use free or older ChatGPT. Their impression sits somewhere around “this thing hallucinates, it produces AI slop.” The viral social-media examples of AI not being sure whether to walk or drive to the car wash are the AI these people see every day. On the other side are people paying a few hundred dollars a month and using Claude Code or Codex for real technical work daily. They see “refactor this entire codebase in an hour,” “find a vulnerability in this system autonomously.” The ceiling of this capability has been growing explosively over the past year.

The two groups inhabit radically different realities. The first group’s information source is summaries from the tail of the rumination chain; the second group’s information source is their own real-time collaboration with the model. The first judges AI by “what did I read about it”; the second judges AI by “what have I gotten it to do.”

Two groups use the same word, “AI,” and are not referring to the same thing. One says “AI is whatever”; the other says “this is exploding.” Both sentences are true. The difference isn’t in the model; it’s in what the user is doing.

IV. Three layers at once

Any single layer, read in isolation, is just a local observation. Read the three together, the rumination chain, the tooling-up of ruminated consumption, and the producer split, and a question surfaces that hasn’t been singled out for discussion.

People consuming rumination are not short of information. They are exposed to more of it than anyone in any previous era. Nor are they “failing to learn,” because the product was never designed to teach them anything. What they are actually doing is using the act of “consuming information” to manage the fear of having fallen behind the times, a kind of self-soothing against anxiety.

People doing real cognitive work (Karpathy’s second group) aren’t doing so because they consume more information. It’s that they are doing something categorically different: their cognitive activity is running. Faced with a problem they push on it, test it, verify it, and form their own judgment. Rumination consumers don’t do any of this. They read the finished product, register a keyword or two, close it, next one.

So what the three layers together point to is not “information overload” (a cliché), nor “AI is making people lazy” (an evasion). They point at a sharper fact: most people have been mistaking knowledge, experience, and fluency for cognitive ability itself. They have never identified what a real cognitive act is.

What counts as real cognitive activity? Start by ruling out several familiar mistakes. It’s not a brain turning over (brains are always turning over; you’re alive). It’s not having read or memorised a lot (that’s knowledge storage). It’s not being skilled in a field after years of practice (that’s experience). It’s not fluent expression (that’s linguistic capacity). These were once confused for cognitive ability because, in the old environment, they tended to show up together with it. But they aren’t it.

Real cognitive activity means producing structural judgments that reality can calibrate. Forming concrete judgments that future facts can verify or refute. Recognising deep structure in an unfamiliar problem. Treating the distance between one’s current judgment and reality as a manipulable object. Cognitive activity in that sense is the thing the AI era has separated out and priced on its own. Most people have never registered that this distinction exists. Without registering it, there is no opportunity to do it.

The old evaluation system did not require it to run. Experience, patterns, fluency, broad reading, expressive ease, social sensitivity: none of the old expressions of “smart” or “capable” required real cognitive activity. A shrewd businessman, a well-read editor, a competent middle manager. Their cognitive systems may never have performed a single genuine reflection. Yet in the old environment they were genuinely high-ability people, because the old environment rewarded those traits, and rewarded them well.

The AI-era shift isn’t “AI makes people lazy.” It’s that for the first time, cognitive activity itself has been peeled off from every other ability and priced separately. Once knowledge storage has been flattened by the model, repetitive tasks swallowed, and the production cost of finished judgments driven to zero, what remains is cutting reality into structure, catching errors in the output, and doing real abstraction on unfamiliar problems. Most people simply don’t have that. And their own past successes tell them “I am capable.”

This isn’t a new problem. It’s a fact that has always existed, masked by the old evaluation system, now exposed for the first time under new technical conditions. Not because people suddenly got dumber, but because the old sources of “feeling capable” are disappearing.

---

Chapter 2 · AI Isn’t Uniformly Strong

Before the main argument, a common misconception needs to be cleared out.

The misconception is this: AI has surpassed humans across the board, so most people feeling small in front of it is natural, and passively consuming rumination is a reasonable response.

This is half right and half completely wrong.

Where AI is strong. On structured, measurable tasks, AI has stably outperformed the median human and, in many cases, the domain expert. Math olympiad (AIME), PhD-level science questions (GPQA), real software-engineering tasks (SWE-Bench), various language-understanding tests: these benchmarks have been broken one after another by frontier models over the past year. Add knowledge retrieval, fluent expression, and instruction-following to the list, and AI far outstrips most people. For work in these directions, the best use of AI is as a super-lever. It can amplify a capable person’s output by more than ten times.

Where AI is weak. An equal amount of evidence shows that AI’s strength is sharp and uneven. Change a problem’s surface conditions slightly. Substitute a scenario that looks unrelated but is structurally identical. The model’s performance can suddenly collapse. This brittleness shows up repeatedly in benchmarks like SimpleBench and BrainBench, where models underperform the median adult on ordinary common-sense questions. And there is a whole class of tasks AI is bad at: judgments with no clear right or wrong (is this design good; is this decision correct), processing murky reality (what is this tangle of user feedback actually saying), abstracting on a completely unfamiliar structure (how do you decompose a problem type you’ve never seen before).

Why it’s uneven. There is a technical reason. The frontier models’ progress is driven mostly by reinforcement learning, which gives the model “right / wrong” signals so it can calibrate. That kind of training is extraordinarily effective in domains with unambiguous correctness. Does the code run? Is the math answer correct? Does the test pass? These are verifiable rewards: clean signals, high training efficiency. Conversely, in domains without clean correctness, such as good judgment, appropriate restraint, or taste, there is simply no way to produce a reliable reward signal. How do you teach a model that “this paragraph feels off”? You can’t. So progress in these domains lags far behind the verifiable ones. Add commercial incentives: code and structured tasks generate direct B2B revenue, so the vast majority of compute and research effort flows there. The result is that AI capability grows exponentially in “verifiable + commercially valuable” domains, and slowly everywhere else. This is the technical story Karpathy has been repeating on X. Chapter 1’s bimodal capability split between the two groups is, at bottom, the two groups using different axes of the same model.

The real corollary: AI is an amplifier, not an equaliser. What it amplifies is not “human ability overall” but “the structure the user can feed into it.” Give it to a person who can already break a problem down and judge output quality, and AI amplifies their work more than tenfold. Give it to a person who can only ask AI to “write something,” and AI returns rumination. What gets amplified is the frequency of their anxiety management.

The crux: AI has not replaced the act of constructing the problem.

A reasonable objection needs to be addressed here. “AI has already demonstrated structural abstraction under reinforcement learning: the o-series reasoning models, working on hard math problems, overturn their own wrong assumptions and rebuild their approach mid-solve; AlphaGo was inventing novel joseki a decade ago. Isn’t saying ‘AI can’t do structural abstraction’ already out of date?”

The observation is correct. But there is a critical distinction that has to be drawn explicitly.

AI’s structural abstraction always happens under a given reward function. Go’s objective is to win; a math problem’s is to be correct; code’s is to run. Inside those well-bounded spaces, where the victory condition has already been defined by humans, AI can indeed search for structure better than humans can. This is instrumental decoupling: finding the optimum inside a defined game. AI is already past humans at that layer, and will keep pulling further ahead.

But there is another layer of decoupling: defining what is worth pursuing, what question is worth asking. What hidden human need does this product actually solve? What kind of human reflex will this policy provoke? What is this tangle of user feedback really trying to say? Where is the real root of an organisation’s long-running problem? This layer has no preset reward function, because “what counts as right” here has to be adjudicated by a human in the first place. AI can generate ten thousand logically airtight structures, but the step of judging which one actually corresponds to the real world’s pain requires an adjudicator who is an embodied agent, living inside reality, bearing consequences. Without that agent, where does the reward function come from?

So what AI has taken over is “finding the optimal structure inside a defined game” (something humans were not particularly good at anyway; handing it to AI is reasonable). What it has not taken, and cannot generate on its own, is “deciding which game to play.” The structure of that second question is, by nature, not in a layer AI can reach.

Back to the division of labour: which questions are worth asking, which direction is right, whether the output actually solves the problem. These judgments still rest entirely with humans. And this class of judgment is exactly what the last section of Chapter 1 was pointing at: the kind of thing only real cognitive activity can do.

So “AI is too strong, I can’t keep up” is a lazy story. AI has indeed surpassed you at the things you should have let go of (memorisation, fluent expression, instruction-following); it has not surpassed you at the things you should actually be participating in (judgment, construction, taste, structural abstraction). The people who have opted out where they could participate are the ones AI is actually leaving behind. And they often don’t notice, because ruminated consumption gives them the illusion of “still keeping up.”

---

Chapter 3 · Where the Reader Stands on the Chain: Learning, or Anxiety Management?

If AI is an amplifier rather than an equaliser, then “the output gap between users is widening” is a necessary corollary. An amplifier blows up whatever gap was already there on the user’s side. But that alone doesn’t explain why the gap between the two groups is so large, and pointed in the opposite direction: on one side, people building complex systems with AI; on the other, not “people who use AI less” but people who consume more and more through AI and produce less and less.

This reverse dynamic needs an additional explanation. It isn’t about ability or skill with the tool. The real reason is that at the tail of the rumination chain, most people aren’t learning at all. They’re doing something that looks like learning but whose mechanism is the exact inverse.

Learning and anxiety management

Put the two definitions cleanly side by side.

Learning is this loop: encounter something you don’t understand → form a preliminary judgment on your own → go find information with specific questions in mind → use the information to calibrate your judgment → form a new understanding that is now yours. All four moves are required. Of them, the prior judgment and the calibration are the core of the process. Without those two steps, incoming information won’t be absorbed into any existing structure. It’s like pouring water where there is no container: a little while later it’s dry, nothing left.

Anxiety management is a different loop: feel behind → consume information → feel “I’m keeping up” → anxiety briefly eases → because nothing was internalised, the feeling of being behind returns a little later → consume again. This loop has no prior judgment, no calibration, no internalisation. Its operating logic is the opposite of learning’s. Learning produces cognitive structure; anxiety management produces emotional relief.

On the surface the two look alike, because both involve the act of “reading something.” Mechanically they are inverses. Worse: the brain cannot distinguish between “reading a conclusion” and “arriving at a conclusion oneself.” Read a well-constructed article and the brain produces the feeling of “I now understand this,” nearly identical to the satisfaction of having reasoned to a conclusion yourself. But the first is a moment’s illusion. Close the tab, and a few days later nothing is left. The second is genuine cognitive internalisation, and changes how the person judges the matter for a long time afterward.

This intracranial mechanism is the foundation that lets the rumination chain keep running. It makes anxiety management feel identical to learning, so the consumer gets a daily “today wasn’t wasted” confirmation while in fact nothing has been taken in.

What the rumination chain is actually for

Chapter 1 already laid out the mechanics of the chain: every layer drops information density, accumulates error, and raises emotional intensity. The more important fact is this: the chain’s design objective is not to convey information; it is to hold attention.

That distinction matters. A thing produced to convey information (an academic paper, a rigorous empirical report, technical documentation) keeps, by design, the content that lets the reader truly understand and use it: definitions, conditions, counterexamples, uncertainties. In the attention economy those things are liabilities. They reduce reading fluency and raise cognitive load. So a thing produced to hold attention will, by design, systematically delete exactly that kind of content.

Which means: no matter how hard you read this kind of material, you will not learn from it. Not because the reader isn’t trying hard enough. Because the product was never designed for learning to occur. The more carefully you read, the more you are just walking circles inside a room with no outlet.

The reader’s position inside the loop

Now stack the anxiety-management loop on top of the rumination chain, and you can describe precisely what most people do each day:

Open an app → scroll into a hook-shaped finished judgment → read quickly → retain one or two keywords or an emotional impression → close → land on the feeling “I kept up today” → anxiety briefly eases.

Not one step in that entire sequence is learning. No prior judgment (before reading, they didn’t know what they were trying to verify), no calibration (they didn’t check the information against any existing judgment), no internalisation (close the tab and nothing remains; ask them later, they can’t recap it). Every action in the sequence is emotional regulation, using the act of “contacting information” to complete a ritual of “I didn’t fall behind.”

Here’s a cruel verification: find someone who scrolls AI news every day, and ask them to recap the core argument, and their own judgment of it, for any single piece they read three days ago.

Most of them can’t.

This isn’t a memory problem. It’s that the content was never processed, so there is nothing there to remember. The brain only holds onto what it has processed itself: what it has judged, argued with, tested against its existing understanding. Information that has flowed past passively, no matter how “information-dense” it felt at the time, is gone a few days later.

The contrast with people doing actual AI-adjacent work is complete. They don’t consume ruminated content. The day’s experiments, model feedback, and debug output already provide information at a density far higher than any secondhand summary. Their judgment of what a model can and cannot do comes from the specific thing they tried and failed to get it to do yesterday, not somebody else’s review.

And here is the counterintuitive part: the more rumination consumers consume, the further behind they feel; the more doers do, the clearer they feel. Not because the doers know more. Because the two activities move information in opposite directions. Anxiety management is input-only: information becomes emotion, burned up. Real work processes input and produces output; every pass reinforces cognitive structure. Consumer impressions of AI: “there’s too much, I can’t keep up, something new dropped again.” Doer impressions: “I noticed yesterday it’s especially strong at X, still weak at Y.” Specific, clean, bounded.

The two groups use the same word while referring to entirely different mental objects.

When consuming rumination becomes an achievement

Push the diagnosis all the way, and a phenomenon worth staring at appears.

OpenClaw’s most virally shared use is “help me summarise what happened in the AI world today.” It looks like an efficiency gain; it’s something that wasn’t possible before. But think about it. Is this actually an achievement worth broadcasting?

Under the framework built up so far, the meaning of that use case becomes: a tool that lets me more efficiently, more continuously consume low-density ruminated content; and I treat it as a use case worth sharing. The throughput of actively eating ruminated matter is itself presented as a productivity expression.

This is not a problem with a particular tool. It is the externalisation of a deeper collective belief: keeping up with information equals being capable. In the pre-AI era that belief had some basis. Information was scarce; sustained access to primary sources was itself a rare skill. After AI collapses the production cost of finished judgments to zero, that belief becomes a purely manufactured need. The content you can consume is infinite, and not one piece of it was produced so that you could learn something.

So we get a structurally ironic picture. AI takes the supply of information from scarce to infinite, and a portion of the population responds not with “I can finally shift from consuming information to producing judgment,” but with “I need stronger tools to consume more.” The stronger the tool, the faster the consumption; the faster the consumption, the more information dimensions brushed against; the more dimensions brushed against, the deeper the anxiety. Every piece consumed implies ten more unconsumed.

Once the production cost of finished judgments is zero, “efficient consumption of finished judgments” becomes the last citadel of a manufactured need. The citadel stands not because it provides any real value, but because it provides a kind of busyness that can be displayed, a performance of “keeping up” that the AI era still accepts.

People doing real cognitive work are not in that performance. They aren’t even watching it.

---

Chapter 4 · A Subtler Trap: Knowledge Management

Ruminated consumption is the obvious form: scroll, read, next. But it has a subtler variant: not passive absorption but active construction, not scrolling information but organising it. Because it camouflages itself better as “serious work,” it does more damage to serious people.

The variant is called knowledge management.

Why it has become a false premise in the AI era

The phrase “knowledge management” hides an older era’s assumption: that knowledge is a static object that can be managed, like a book in a library, able to be categorised, indexed, retrieved. The entire tool-chain of the past few decades (Evernote, Notion, Roam, Obsidian, Logseq) is built on that assumption.

Real cognition doesn’t work that way. What’s in your head as “knowledge” is not statically stored information; it’s a relational network constantly being re-composed. “Understanding” is not “remembering.” “Being able to use it” is not “being able to find it.” That misalignment meant knowledge management was doing a misaligned thing from day one.

Several mechanisms came together in the past few years and pushed the misaligned thing into full false-premise territory:

Retrieval has been solved by AI. One of the core reasons to organise notes used to be “so I can find it later.” Today’s AI knows almost every piece of text in the world, and retrieval cost has gone to nearly zero. About 80% of the traditional “organise so you can retrieve” purpose no longer holds.

The memory-outsourcing trap. There’s a psychological phenomenon called the Google effect: knowing that “the information is saved” makes the brain remember it less. The effect is extreme in heavy note-takers: a note containing X ≠ your head containing X. Knowledge that can actually be called on lives as an active model in your head, not as a tag in a notebook. Many heavy note-takers’ active internal knowledge is in fact atrophying, because they treat the brain as an index rather than a workbench.

Organising notes is the most sophisticated form of anxiety management. It scores exceptionally high on the anxiety-relief instrument: it looks like work, produces a sense of accomplishment (“added twelve backlinks today”), carries no judgment risk, and supplies the feeling of “growing” without requiring any output. It is more dangerous than scrolling ruminated content, because its “I’m doing real work” disguise is stronger.

Organising notes and forming judgments are different activities. Luhmann’s Zettelkasten generated the core material for seventy-plus books not because of whichever system he used, but because every card held his own thinking. Each card was a small judgment, a calibration against existing knowledge. The essence of that system is not “managing knowledge” but “forcing thought.” Modern Obsidian users are mostly managing other people’s thinking: highlighting, quoting, citing. That has nothing in common with what Luhmann did. It only borrows the same visual form.

Karpathy’s wiki: what it actually solves

In that context, in early April 2026 Andrej Karpathy proposed an LLM Wiki scheme: drop raw material into a raw/ folder, let the LLM compile it into a structured markdown wiki, and let the human curate only the input. His own wiki for a single research topic has grown to roughly 100 articles and 400,000 characters; he barely edits it by hand.

The scheme solves three things: loss of context between sessions, the unsustainability of note maintenance (LLMs do the bookkeeping better than humans do), and the fact that knowledge doesn’t compound (every new piece of material automatically updates multiple existing pages and establishes cross-references).

This sounds like a refutation of “knowledge management is a false premise.” After all, he’s doing a form of knowledge management, and in his case it clearly works.

But a bottom-level fact determines why it works for him in particular: Karpathy is a scientist.

His raw sources are arXiv papers, experiment results, code, and his own unpublished work. All of these are either absent from, or already stale in, the training data of the large models. Model training data has a cutoff; the frontier research the model doesn’t know about is exactly the information space a scientist lives in. He isn’t “managing knowledge”; he’s maintaining a delta against the model’s knowledge base, layering the frontier on top of the existing foundation.

That is categorically different from the knowledge-management community’s version of “I have to manage everything I’ve ever read.”

Why ordinary people fail when they copy it

Once Karpathy’s precondition is clear, you can see why ordinary followers replicating his scheme end up producing new anxiety management.

The topics ordinary people care about, such as business news, AI developments, popular science, management frameworks, and industry analysis, are things the model basically knows better than they do. The “knowledge” they’re trying to “manage” is actually a ruminated version of material the model already has in its training data.

Copying Karpathy’s scheme under those conditions becomes what? Having an AI generate a structured-looking wiki out of already-ruminated content. You’re adding another layer to the rumination chain. The output looks more like “knowledge,” more like “a system,” more like “research.” But it’s further from the primary source and emptier.

And because a wiki’s visual structure is stronger than that of a notebook, the illusion of “I’m learning” it produces is also stronger. A person who watches their wiki grow by the day, cross-references multiplying, will feel more strongly that they’re “growing” than one scrolling Instagram. What’s actually happening is the same as what happens to the Instagram reader: they are consuming information they cannot actually call on.

The only difference is that one consumes on a public product and the other consumes inside a self-built system. The latter is more of a cope, not a healthier habit.

What a reasonable personal information infrastructure looks like

To avoid the trap, ask a fundamental question: relative to the model, what am I short of?

Most knowledge-management practice never asks this. It asks only “what should I record,” and ends up recording a pile of things the model already knows more and more accurately.

There are only two reasonable classes of answer:

Class I · The frontier delta. You work in a fast-moving frontier field: research, cutting-edge engineering, emerging practice not yet covered by training data. In that case maintaining a continuously updated knowledge base has real value, because what you are accumulating is what the model doesn’t know. Karpathy’s wiki fits this scenario. Preconditions: you are actually doing frontier work, your raw sources are primary, and you have the judgment to check whether the wiki has drifted.

Class II · The personal-cognition delta. You’re not working at the frontier, but you have your own judgments on specific matters, preferences, non-consensus models, personal lessons. None of these can be generated by a generic model from its training data. In this case the right move is a minimum alignment layer: record only “things the model doesn’t already know,” so that next time you collaborate with the model it can start from where you left off instead of from zero.

The two schemes look very different; the underlying logic is identical. Both are answering the same question: “relative to the model, what am I short of?” The “short of” simply falls in different information distributions for different people.

The second class has a counterintuitive but important property: it cannot be performed. Its entry requirement is “I can identify what the model doesn’t know.” People without that judgment can’t use the method (they discover they have nothing worth recording); people with it end up recording material that’s guaranteed to matter. That self-selection is worth more than the method itself. It structurally prevents the method from becoming a new rumination ritual.

For most people, Class II is the reasonable starting point. If, during use, you discover “the model doesn’t know enough about my field,” you can consider adding a Karpathy-style disciplinary base. The opposite order lands people straight in ritual knowledge management. This is the single biggest mistake the knowledge-management community has made: advising beginners to build the base first. The beginners then spend months organising other people’s judgments, and the organising substitutes for the cognitive activity itself.

A more general principle

Lift knowledge management to its most general level and you get a principle that evaluates almost every future “new scheme”:

In the AI era, any methodology that doesn’t require the user to exercise independent judgment has become, or is becoming, a variant of ruminated consumption.

The distinction is not in the content of the methodology. It can be about productivity, learning, knowledge organisation, creativity. The distinction is whether the user has outsourced judgment to it.

Which also explains a common pattern: every “popular methodology” is, at its core, a collective ritual of cognitive outsourcing. The more popular a methodology, the more it indicates that what it satisfies is not a concrete tool need but a “don’t have to judge for myself” need. Genuinely effective methodologies tend not to go viral, because they depend too heavily on the user’s specific context to be replicated in bulk.

Karpathy’s wiki works for him. But the wave of “Karpathy Wiki Tutorials,” “Second Brain 2.0,” “AI Knowledge Management Workflows” that followed is already turning into the next round of performance schemes. The test is simple: how long into using this scheme does it take before you start feeling “I don’t have to judge for myself”? That moment is the moment it starts rotting.

---

Chapter 5 · What They All Got Right: Generalists, Taste, and the Refolding of Intelligence

By this point both shapes of ruminated consumption have been pulled apart: passive scrolling of information and active organising of knowledge. They look different; underneath, both are doing anxiety management rather than learning. That surfaces the question that has been circled all along: what actually matters as a capability?

This question had already been brushed against several times before AI. A few widely circulated framings, from different corners, were describing the same underlying thing from different angles. Each one described only a fragment, and none of them realised it was describing the same thing. Laid side by side, a more complete picture emerges.

The generalist thesis

Over the past few years, in creator and indie-developer circles, the concept of the “generalist” has been pushed back to the front. The most-quoted version comes from Dan Koe. The core claim is that specialisation is losing value, and the multi-domain generalist is becoming scarce and valuable.

What he got right: breadth is not miscellany. A real generalist is not “someone who knows a little bit of everything,” but someone who has accumulated enough depth in multiple fields to see the isomorphic structures between them. A person who has studied economic incentives, biological evolution, and organisational behaviour in depth simultaneously discovers that the underlying mechanisms of the three fields overlap heavily. All three describe what happens at the system level when multiple actors pursue their own goals under constraints. Recognising that cross-domain isomorphism is the generalist’s real value, and the only route from breadth to judgment.

What he didn’t make explicit: what kind of depth counts. This is the vaguest part of the generalist thesis, and the part where followers most often slip. Some people, following the thesis, read one introductory book from each of ten fields and end up with ten fields’ worth of surface narratives in their head. That “depth” is useless for cross-domain mapping, because surface narratives are different across fields by definition; there is no isomorphism to be found. What can actually be mapped is deep structure: the causal mechanisms, incentive constraints, and feedback loops that genuinely drive behaviour in a field. A workable test for depth: can you name one thing that practitioners of a field collectively believe but that is actually wrong? If yes, you have depth. If no, you’re still stuck at the standard-narrative layer.

The gap he left behind: the generalist thesis never singles out the underlying mechanism of doing cross-domain mapping. It reads as a methodology: read books from several fields, draw connections. But it rests on an unstated precondition: that the reader is capable of making such mappings. And it’s that capacity that’s actually scarce. Which explains why, given the same methodology, some people pull it off and others don’t: the difference isn’t effort, it’s the mapping capacity itself.

Taste

A second line comes out of the creator and design circles: the taste that Paul Graham, Steve Jobs, and Rick Rubin have been circling for decades. The Chinese translation of the word (品味, pǐnwèi) makes it sound like a fuzzy aesthetic sensation, but what they are describing is a more specific mechanism.

What they got right: before you can articulate the reason, you can already tell whether a thing is good. That pre-linguistic judgment is not mysticism; it comes from the long internalisation of a large body of high-quality samples. A person who grew up in museums can judge a painting’s quality within seconds, not because they know art history, but because their visual system has been calibrated by a large body of high-quality paintings. The output of that calibration is taste.

But what they are describing is low-order taste: pure pattern recognition. Someone can tell that “this typeface is ugly” but can’t say why. That taste is real and useful. It lets a person rapidly filter options inside a domain they know well. But it has a severe limitation: it doesn’t transfer, can’t be taught, can’t be verified. You cannot pass your taste to someone else; you can’t build taste from zero in a new domain, because you don’t have the sample base yet.

What they didn’t make explicit: taste actually has two layers. Low-order taste is pre-linguistic pattern recognition. High-order taste is pattern recognition plus the ability to extract structured reasons on demand. A genuinely strong designer doesn’t only say “this typeface is ugly”; they can analyse why. It may be that the ratio between x-height and weight breaks the visual rhythm. High-order taste matters because it is teachable, transferable, and lets you build taste in a new domain faster. It has made the structure behind the intuition explicit, so the structure can be learned.

High-order taste and the generalist thesis turn out to be, at bottom, the same capability. Both are abstracting a concrete judgment into an operable structure, then applying that structure somewhere new. The generalist thesis talks about it from the cross-domain angle; the taste argument talks about it from the aesthetic-intuition angle. Two faces of one mechanism.

The cognitive-folding thesis

A third line has become common in tech circles and among indie creators. Its core claim is blunt: AI will not narrow the gap; it will widen it at an unprecedented speed. The analogy is usually the industrial revolution. Steam engines and assembly lines crashed the relative value of physical labour. The gap between the people who could design machines, organise production, and schedule systems, and the people who could only sell their physical labour, went from linear widening to exponential splitting. Now it’s cognition’s turn. AI is flattening a batch of cognitive abilities (knowledge, memory, retrieval, fluent expression), and the remaining part that isn’t flattened will be blown up explosively. The result isn’t “inequality grows”; it’s “different populations’ output in the same unit of time gets folded onto completely different orders of magnitude.”

What they got right: the gap will widen, at a speed without historical precedent. The judgment is correct, and the industrial-revolution analogy is apt. Every previous general-purpose-technology leap has produced not linear widening but exponential splitting between people who seized the new lever and people who didn’t. True of steam, true of electricity, true of the internet. AI will only be more extreme, because it is the first time the technology acts directly on cognition itself, which is the source of judgment, creation, and decision. What gets amplified will be more visible than it has ever been.

What they didn’t make explicit: which kind of cognitive capability gets amplified.

This is the murkiest part of the argument, and where most followers slip. The default assumption is usually “high cognition = high IQ,” so the conclusion slides into “high-IQ people will grow richer; low-IQ people will be left behind faster.” The first half of that is roughly right; the reason behind it is completely wrong.

What gets amplified is not IQ. IQ measures raw processing: working memory, processing speed, information extraction, rule-based inference. AI already beats almost everyone at these capabilities. A person with high IQ who never does real cognitive activity is, in the AI era, the one eliminated fastest, because the things they used to use their IQ for (rapid learning, memory recall, fluent deduction) are now done faster, more accurately, and more cheaply by AI. Their advantage is flattened.

What actually gets amplified is another layer: recognising deep structure on an unfamiliar problem, judging which question is worth asking, carving operable objects out of a fog of reality. This is related to IQ but not identical. A person with moderate IQ who has identified what real cognitive activity is and sustained effort at it will be amplified more by AI than a person with high IQ who has spent their life inside the comfort zone of pattern recognition.

The gap they left behind: “high cognition” gets treated as an innate, fixed attribute. This makes the whole folding narrative sound like an unstoppable force: high-cognition people win, low-cognition people get left behind, no middle ground, no agency.

But the real divide isn’t the innate “high cognition vs. low cognition” divide; it is the conscious one: “recognised it vs. didn’t.” The first is an immovable fate; the second is a crossable threshold. Not easy to cross, but in principle open. Most of the hopelessness that colours the cognitive-folding narrative comes from mistaking the crossable threshold for the immovable attribute.

Three lines are describing one thing

Put the three together.

The generalist thesis’s “cross-domain mapping ability,” the underlying mechanism that turns multi-domain depth into judgment, is really the ability to peel the surface off multiple domains and grasp the deep structure they share. Without it, reading across ten fields produces ten piles of fragments; with it, three fields can produce a new judgment.

The taste argument’s “high-order taste,” the ability to make the structure behind intuition explicit, is really the same ability operating on top of a deeply internalised sample base. Low-order taste needs only samples; high-order taste needs samples plus the capacity to abstract the pattern into structure.

The cognitive-folding thesis’s “the layer AI doesn’t replace but amplifies” is also this ability. Once AI has flattened knowledge, memory, fluent expression, and instruction-following, the one thing left unflattened is the act of peeling reality off its surface and operating on it as structure. That is what gets amplified.

Three lines, three different angles, touching the same thing and giving it three names: generalist, taste, folding. The core is one thing: the ability to abstract a concrete situation into an operable structure.

Cognitive science has a name for it: cognitive decoupling. It means peeling a representation off the reality it refers to and operating on it as an independent object. Every abstraction, hypothesis, counterfactual, and act of self-examination is built on it. The three popular framings are each touching a different facet.

This capability has always existed, but it was never priced on its own. In the pre-AI era it sat mixed in with a large number of other capabilities: knowledge storage, memory, fluent expression, skill through practice. A “smart” person typically had all of them, but nobody knew which of them was actually doing the work. So each popular framing was a blind person touching an elephant, grabbing a part and describing it with the vocabulary of that part.

The AI era makes this visible for the first time. AI has substituted for knowledge storage (it knows more than anyone), memory (it retrieves instantly), fluent expression (it writes better than most), skill through practice (it doesn’t tire, doesn’t slip, doesn’t need reps). With those capabilities stripped away one by one, what remains is the underlying mechanism that was always masked: the ability to cut reality into structure, judge which structure is right, and build new structure in unfamiliar territory.

That is what the generalist thesis, the taste argument, and the cognitive-folding thesis are all converging on. It isn’t a new ability; it’s an ability that has always existed and is only now being named in isolation.

What AI can do, what is left to humans, what to learn

Take the judgment to the practical level.

The range of what AI can do is expanding quickly: pattern synthesis, information retrieval, fluent expression, instruction-following, and complex tasks in verifiable-reward domains (code, math, structured reasoning). That range is expanding every few months with no visible ceiling.

The range of what AI can’t do is shrinking, but a few things it cannot do now and cannot do soon: carving murky reality into operable problems (problem construction), recognising structural errors in its own output (meta-judgment), judging what is good in the absence of an external reward signal (taste and value trade-offs), true abstraction on a completely unfamiliar structure (decoupling itself).

The first three all rest on the last one. Without real decoupling, the first three are just interpolative mimicry in a high-dimensional representation space. They look like the real thing but aren’t.

So what is actually left to humans in the AI era? The layer that directs the model: posing the right question, judging whether the output actually solves that question, pulling the model back when it drifts, making the value decision about the final result. Every part of that layer’s work rests on cognitive decoupling.

Then the crucial question: what, specifically, should one learn?

Not more knowledge. The model knows more than anyone. Not more fluent expression. The model writes better than most people. Not more “thinking frameworks.” Most of the thinking frameworks on the market are rhetoric dressed up as tools.

What one should actually learn is a small number of concrete things: several core reasoning tools (so judgment can be made without relying on intuition); regularities from a few domains outside one’s own (so cross-domain mapping has something to map); and putting oneself in environments where reality calibrates (so everything above doesn’t slowly rot). Plus the partly-untrainable underlying mechanism: cognitive decoupling.

Four things, independent of each other, acting on each other. Together, a whole system.

---

Chapter 6 · The Four-Part Formula, Item by Item

Decoupling itself is only an operator. A person’s capacity for cognitive output is jointly determined by four things. Their relationship is not additive. It is multiplicative.

Hardware × software × database × operating environment.

If any one item is zero, the whole thing is zero. This explains why most attempts to “improve cognition” fail. They usually work on only one of the four items, while the multiplicative structure lets the other three weaknesses cancel the gain completely. A naturally smart person who never learns reasoning tools and a person who has loaded themselves with reasoning tools but never exposes judgment to feedback will both produce close to zero. They just fail in different ways.

Take the four items in turn.

Item I · Cognitive decoupling as hardware

Cognitive decoupling is the ability to peel a representation away from the reality it refers to and operate on it as an independent object. Every abstraction, hypothesis, counterfactual, and act of self-examination rests on this ability. Without it, a person remains bound to the concrete stimulus in front of them. “My thought” and “the fact” blur together. “This situation” and “the structure behind it” cannot be separated.

The closest psychological concept is fluid intelligence: the ability to recognise structure when facing a completely unfamiliar problem, without relying on prior experience. Notice that it has nothing to do with knowledge storage. A well-read person may have low fluid intelligence; a person who has read very little may have high fluid intelligence. This is the underlying capacity for asking, when meeting something unseen, whether one can produce a structure for it by oneself.

There are several observable signals in daily life. Give someone two problems with different surface descriptions but the same underlying structure. A strong decoupler performs similarly on both. A weak decoupler collapses as the surface difference grows, and usually does not notice the collapse. Stability in counterfactual reasoning is another signal. If X had not happened, and everything else stayed the same, what follows? A person who can hold “X did not happen” steady and reason from it has decoupling online. A person whose mind is immediately overwritten by the actual world does not. There is also a very simple proxy: the first reaction to a completely unfamiliar problem. Strong decouplers begin abstracting the structure. “What is the general form of this problem?” Weak decouplers get stuck at “I don’t understand this,” or force-fit it to the most superficial similarity available.

This part has to be brutal.

The upper limit of fluid intelligence is largely genetic. Twin studies comparing identical twins, whose genes are almost the same, with fraternal twins, whose genetic difference is like that of ordinary siblings, show that this ability is far more heritable than people like to admit. There is also a counterintuitive pattern: the older you get, the more genetic it looks. Childhood and adolescence still leave some room for environmental effects. As people age, genetic influence becomes stronger, and by adulthood this ability is largely locked.

Over the past twenty years, many products have claimed to “raise intelligence”: brain-training games, working-memory apps, thinking courses. Later large-scale studies have mostly rejected that path. You improve at whatever you train, but the improvement does not transfer to real reasoning tasks. Practising a brain-training game makes you better at that game. It does not make you think more clearly when you face an unfamiliar problem.

That means almost every course or product promising to “improve cognitive ability,” “train thinking,” or “raise IQ” is selling an illusion. In adulthood, this item is basically locked.

So there are only two things to do.

First, identify honestly where you stand on this item. Not in order to give up, but in order to allocate effort rationally across the other three. A person with middling fluid intelligence and a person with very high fluid intelligence will have different ceilings of output complexity after installing the same reasoning tools. Pretending the ceiling does not exist only makes people burn themselves on the wrong target.

Second, do not let the limits of this item contaminate your judgment of the other three. The next three items are trainable, and they determine how far a person is from their own ceiling. For most people, that distance is far larger than the gap between ceilings.

An honest footnote has to be added here. Fluid intelligence does not only set the ceiling for decoupling. It also affects the training efficiency of the next three items. The speed at which reasoning tools are installed, the depth to which they are installed, and whether they transfer to new domains do not develop at a uniform rate. People with high fluid intelligence install them faster, understand them more deeply, and transfer them more widely. This pattern has solid empirical support in cognitive science. The more complex a learning task is, the stronger the correlation between fluid intelligence and learning rate. In the early phase of skill acquisition, fluid intelligence explains roughly 30% to 40% of the variation in complex problem-solving ability.

So the more honest version is this: this is not a story in which everyone can catch up with the smart people by trying harder. The space below the ceiling is open to everyone, but the shape of that space and the slope of the climb differ. A person with below-average or lower-middle fluid intelligence may need longer than the 6 to 12 months given later to install the basic reasoning tools. The eventual depth may also be shallower than that of a person with high fluid intelligence.

But this does not change the central fact. Most people, including many people with high fluid intelligence, have not used the reachable space they actually have. Not using it is the real waste. Whether one can reach the very top is a separate question, and for most people it is irrelevant.

The self-help literature that makes people feel “I am actually very smart, I just haven’t used it yet” mainly serves anxiety management. It is not help. Useful information is cold: the ceiling exists, and after adulthood it mostly does not move. But below that ceiling, most people have left enormous space unused.

Item II · Reasoning tools as software

Reasoning tools are concrete methods of thought installed in the cognitive system and available for use. Their relationship to decoupling is like software to hardware. Hardware determines how complex a program can run. Software determines what the hardware actually outputs on a given problem.

This is the most severely underrated item in the four-part formula. Most people who describe themselves as “people who like to think” have an empty reasoning-tool stack. What they call thinking is intuition rationalising a phenomenon, then fluent language expressing the rationalisation. No tool participates in the process, so the output is mostly a by-product of emotion and pattern matching. It is not judgment.

The four reasoning tools with the widest coverage and highest return are these.

Probability and uncertainty reasoning. Truly understand base rates, conditional probability, sampling bias, selection effects, and calibration. This does not mean taking a statistics course. It means turning these ideas into judgment reflexes.

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

Game theory and incentive structures. This does not require mathematical depth. What matters is the reflex of asking, in any phenomenon, who is paying for whose decision. Once this tool is installed, the way one reads news, policy, and business changes completely.

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

Each tool has classic entry-level books, listed in the appendix. Once the four are installed, the cognitive system automatically decomposes any phenomenon into variables, causal directions, incentive distributions, and feedback loops. This is not a thinking trick. It is a reflex.

There is a simple test for whether the tools are installed. When facing a new phenomenon, does the person jump straight to a conclusion, or do they automatically place it inside a reasoning frame? The first type is impoverished. The conclusion may be right or wrong, but the person cannot tell which. The second type has internalised the tools. They first ask about base rates, causal direction, incentive distribution, and feedback delay.

Another diagnosis is more direct. Ask the person to estimate a quantity they are completely unfamiliar with, such as how many bicycles in a city get rained on in a year. A well-equipped person will spontaneously decompose the estimate into several independent factors and multiply them. An empty-stack person will give a number directly or say, “How would I know?” The difference is not knowledge. It is tooling.

This item is fully trainable, and its return is the highest of the four because most people start near zero. Installing the minimal set takes roughly 6 to 12 months of serious reading plus deliberate use in daily judgment. That time span is surprisingly short for most people. It is far shorter than any degree program, far shorter than learning a craft, and the return is much larger.

The key is not finishing the books. The key is being able to call the tools automatically in daily judgment. The verification method is simple. Next time you form a judgment about some phenomenon, pause and ask which tool is operating here. If you can answer accurately, it is installed. If you cannot answer, it is not. If you can answer but realise your judgment bypassed the tool and went straight through intuition, you are in the most common intermediate state. You need to practise putting the tool before the intuition.

The relationship between this item and the first has to be made explicit. A repeated academic finding is that the correlation between IQ and actual judgment quality is surprisingly low. High IQ does not prevent stupid decisions, because IQ measures processing power. It does not measure whether reasoning tools are installed. A person with middling fluid intelligence and a complete tool stack can steadily outperform a person with high fluid intelligence and an empty stack.

For people who did not get a strong starting point on the first item, this is genuinely good news. For people who did get a strong starting point and are too lazy to install the second, it is bad news.

Item III · Cross-domain regularity depth as database

Decoupling needs something to operate on. That something is an understanding of regularities across multiple domains. But “depth” has to be defined precisely, because it is easily mistaken for mastery or erudition.

There is a core distinction here. A domain’s deep structure is the causal mechanism and constraint relationship that actually determines behaviour. Its surface features are terminology, process, cases, and industry jargon. What decoupling can use is deep structure, not surface features.

So cross-domain regularity depth is not mastery. Mastery means being able to execute work inside a domain: arguing a case, performing surgery, writing production-grade code. Regularity depth means being able to explain why things in that domain happen the way they do: the incentive structure, information asymmetry, feedback delays, survivorship bias, and where the standard narrative lies. The first requires more than ten years of specialised investment. A good observer can reach the second in a few months.

The harshest diagnostic is this: can you name one thing practitioners in a field collectively believe that is actually wrong?

The logic behind this test is that every field has its own standard narrative. It is constructed by stakeholders, beautifies itself, and is sometimes inverted from reality. A person who has grasped a field’s regularities can identify where that narrative lies. If they can name it, they understand. If they cannot, or if all they can repeat is generic meta-complaint everyone in the industry already knows, such as “capital chases profit” or “the system is broken,” they are still at the standard-narrative layer. They have not entered deep structure.

Another diagnostic: when hearing about a new phenomenon in that field, can the person predict where it will go without looking anything up, and know which assumption was wrong when the prediction fails? If yes, there is already a causal model in the mind. If not, the person knows facts about the field, but does not have a model of it. Having facts without a model does nothing for decoupling.

This item is fully trainable, but the threshold is higher than for reasoning tools. It requires time and sustained curiosity. Each domain takes roughly 6 months to a year to reach the level where one can predict and identify error.

Three types of domains have the highest return.

Domains with obvious incentive distortion. Medicine, education, academic publishing, government procurement, insurance, charity. These fields have the largest gap between appearance and actual dynamics. Understanding any one of them installs a set of transferable thinking tools.

Domains rich in historical data. Financial markets, military history, epidemic history, technology iteration. They have real feedback, falsifiability, and regularities pressure-tested over long periods.

Domains adjacent to your home field. Neighbouring disciplines around a field where you already have depth. The learning slope is fastest, and the value transfers back to the home field most directly.

The minimal learning path for each domain is simple. Read one insider’s critique of the field, not an introductory textbook. Read 2 or 3 empirical studies or systematic reviews. Follow one high-information-density source regularly: not news, not a KOL, but a deep newsletter or researcher blog from inside the field. Then make at least 10 verifiable predictions and track them. Without the last step, the first three are entertainment.

What should be actively avoided are domains driven mainly by story and narrative: fashion, celebrity gossip, political commentary, inspirational business books. Their “regularities” are mostly post-hoc rationalisation. Studying them fills the mind with more surface features, not deep structure.

Item IV · Feedback exposure as operating environment and exponent

The first three items determine momentary ability. The fourth determines whether the first three can be maintained and keep growing over time.

Feedback exposure means the degree to which a person’s judgments are systematically tested by reality. In a high-feedback environment, each judgment is confirmed or corrected by facts. In a low-feedback environment, judgments can float indefinitely without being calibrated.

Why is this an exponent rather than an additive item? Because without feedback, any level of the first three decays over time. Decoupling turns into self-indulgence: the person thinks they are abstracting structure, but they are producing deep-sounding nonsense. Reasoning tools turn into ritual: Bayesian language without any actual prior update. Cross-domain regularity becomes scholarly collecting: the person knows many patterns but cannot tell which one applies in the current situation. Conversely, even if the first three items are only moderate, a high-feedback environment keeps calibrating them. Over time, output quality can exceed that of someone fully loaded on the first three but living in low feedback.

The simplest diagnostic is this: among the important judgments you made in the past year, how many were clearly confirmed or refuted by reality?

If the number is close to zero, then no matter how much the person thinks, how many books they have read, or how fluent their expression is, the cognitive system has not been calibrated for a long time. They may once have reached a fairly high level, but that level is slowly distorting, and they cannot see it. A low-feedback environment is defined by the absence of anything that tells you the distortion is happening.

A famous long-term study tracked nearly 300 “experts” across fields and collected a large number of their predictions over twenty years. The core finding was not which people predicted well. It was that most experts’ predictive accuracy was close to a coin toss, while their self-assessment of accuracy was far higher than their actual performance. The gap came from one thing: their predictions had never been systematically recorded. No scoreboard, no calibration.

Strictly speaking, this item is not about training. It is about environment selection. Feedback exposure cannot be raised by effort alone. It can only be selected by entering some environments and refusing others. This makes it the most hidden of the four items. People can roughly infer the first three from conversation. Feedback exposure is structural. Outsiders cannot see it, and the person often cannot see it either. Someone who has spent ten years doing “strategy” at a large company may be strong on the first three items. But if none of those judgments has ever been tested by the market, users, or concrete outcomes, most of that decade’s “thinking” output is noise.

The environments worth entering include entrepreneurship, where sales and retention test judgment every week; trading and investing, where decisions carry profit and loss; empirical research, where hypotheses are routinely falsified; building products for real users rather than KPI products; and competitive games such as Go, poker, and esports, where outcomes are clear.

Those are feedback environments at the level of career. But feedback loops are not limited to professional life. Daily life contains many equally strict feedback structures, although they are easy to ignore. Cooking tests salt, heat, and timing at every meal. Raising children gives feedback within seconds through crying, laughter, and behaviour. Physical training lets the body report on movement, intensity, and diet adjustments over short cycles. Caring for pets resembles raising children but often gives faster feedback because reactions are more direct. Learning an instrument gives immediate feedback on pitch and rhythm. Repairing things leaves no room for self-beautification: it works or it does not. Negotiation tests whether your judgment has aligned with what the other side actually wants. Gardening is equally indifferent. The plant lives or dies.

These are high-density feedback loops. By contrast, someone who has spent ten years in a big-company meeting room making strategy slides and never executing may score far lower on feedback exposure than a parent who cooks daily for two children while caring for a dog. This is not rhetoric. It is structure. The former’s judgments never touch reality. The latter’s judgments are confirmed or corrected within hours.

Culturally, we tend to bind “cognitive ability” to certain occupations: scientists, engineers, investors. But feedback exposure does not respect occupational labels. A person genuinely invested in a life practice that requires continuous feedback, even if it is cooking, parenting, or training, may keep their cognitive system more online than many people in respectable white-collar roles whose work is never validated by outcomes. The first three items, decoupling, reasoning tools, and cross-domain depth, do require specialised training. But the fourth item, keeping one’s judgment in contact with reality, is available in any life form.

What should be avoided are pure opinion production such as KOL commentary, columns, and punditry; consulting-style “strategic thinking” where advice is given and execution is never tracked; large-company roles with low-frequency decisions and long-delayed feedback; and any domain where speaking well is enough to win. Persuasiveness and judgment accuracy are completely decoupled in those domains.

If changing the main environment is temporarily impossible, one can build feedback mechanisms. But first a common illusion has to be broken. The popular trio of prediction logs, decision journals, and adversarial peer groups is something almost nobody sustains. The problem is not that those tools are wrong. The problem is that they turn feedback into an extra ritual detached from real life. A person busy with work and life who relies on willpower to write five predictions a month will stop within two weeks. People who lack feedback do not lack a spreadsheet. They lack a life structure that forces judgment to meet reality. A spreadsheet cannot solve a structural problem.

Sustainable feedback comes from two things: closing judgment-verification loops inside domains one already cares about, and turning the learning of new domains into a process with verification built in. Neither is a journal. Both are actions.

First: close loops in existing domains. Most people are already making judgments in their own field every day. The judgments simply remain implicit, so they cannot be verified. A colleague asks whether a plan will work. A friend asks whether a company is worth joining. You decide whether to build a feature. You read a news item and feel that it will reverse in two weeks. Each of these is a judgment, and each could be tested by later facts. But most people never say or write the judgment concretely. It remains a vague “I feel,” and afterward, whatever happens, the brain can rewrite the memory into “that is what I thought all along.”

There is only one thing to do: make the judgment concrete at the moment it occurs. When telling a colleague that a plan has a problem, say specifically, “I think it will get stuck on X within two months.” When making a product decision, write directly in the document, “We expect DAU to rise by Y after this ships. If it does not, I was wrong.” When making a prediction about a news event, tell one concrete person, “I bet X will not happen.” No journal, no scoring system, no Brier score is required. Saying the judgment out loud already forces later self-calibration, because what was said will be remembered by the other person or by you a few months later.

The threshold is accepting the risk of being wrong in public. Most people do not say their judgments because if they are wrong, they lose face. Exactly that pressure makes the judgment real. It forces you to think again before issuing it. Every such concretisation is a small act of cognitive calibration.

Second: turn learning a new domain into verification. When trying to understand a new field, most people default to reading, then reading more, then reading still more until they reach a vague sense of “I think I get it.” That “understanding” is never tested. The correct path is to form a prediction that can be wrong before verifying it.

The procedure is concrete. Choose a field you want to understand. Before reading anything, write down three to five basic assumptions about how the field works, what factors drive it, and how it may change soon. After writing them down, begin reading. The reading process is not “absorbing information.” It is colliding the material with your own prior assumptions. Which assumptions are confirmed? Which are refuted? Which need revision? Afterward, what remains in the mind is not a pile of other people’s conclusions. It is your own model, calibrated by contact with material.

The point of this method is not efficiency. It is the processing path of information. Reading before thinking is passive absorption, and its product is a repetition of other people’s judgments. Thinking before reading is active calibration, and its product is your own causal model. The time cost is roughly the same, but the output quality is completely different. The first path leaves almost nothing after a few months. The second builds a reusable judgment tool in that field.

Why these two methods work while journals fail. Journal methods require extra energy for the act of recording itself. Except for a small number of unusually disciplined people, that is not sustainable. The two methods above require no extra life content. They simply change the structure of actions you are already taking. You are already discussing things with colleagues, making decisions, reading, and following news. These methods make those actions carry a verification loop inside them, instead of adding a separate ritual beside them.

Sustainability comes from not adding new life content, but changing the structure of existing content. That principle matters more than any specific tool.

---

Of the four items, only one determines the ceiling. The other three determine how far a person is from that ceiling.

Most people are very far from their own ceiling.

This is bad news for people who do not want to move. It is good news for people who do.

---

Chapter 7 · Conclusion

A person willing to move, after reading the framework above, will naturally ask: “So where should I start?”

The question itself deserves suspicion.

If the first move after reading the four-part formula is to ask for a starter checklist, that means the article is still being consumed as a methodology. However good the checklist is, the person holding it will probably stop within two weeks. The actions described by the checklist all depend on something deeper, and that deeper thing cannot be started by a checklist.

That deeper thing is only this:

Stop treating ruminated consumption as learning.

This does not mean “scroll half an hour less” or “wake up early to read every day.” Those frame the problem as discipline. What actually has to stop is a form of self-deception: mistaking the feeling of “I kept up with the AI world today” for “I learned something today.” Admit that the former is only emotional regulation, in the same category as relaxing with a TV show. That admission itself is the first step.

After that admission, something concrete happens. Cognitive bandwidth that was not occupied by rumination begins to appear. Once it appears, it finds its own use. Not because you have “resolved to learn,” but because a mind not drowned in noise naturally moves toward the problems it actually cares about.

The next questions are outside the scope of this essay: how to work with models, which questions are worth asking, how to use AI to sharpen one’s own judgment, and how to grow cross-domain depth out of existing experience. This essay is only a diagnosis. It explains what the divide in front of you is, what it is made of, and how far most people are from their own ceiling.

---

This system is not for everyone.

The hard ceiling of cognitive decoupling means some people cannot install this system. That is cruel, but it has to be said. This essay has not offered universal hope. It has offered a framework for locating oneself.

But the people who can read this far are not that group. If you can recognise “I am consuming rumination” and follow the argument all the way here without closing the page, you have already completed the basic self-test. Some of the brutal statements above did not hit you.

The real re-stratification of the AI era is already happening. The old evaluation system rewarded crystallised intelligence and pattern libraries, so people who mistook knowledge and experience for cognitive ability could still appear capable. After AI flattens those things, the remaining part, real cognitive activity, is being priced on its own for the first time. Most people have not noticed. By the time they do, the gap will no longer be catch-up sized.

One final self-test is more lethal than any framework above.

Think back to the last time you genuinely changed an important view. When was it, and why did it change?

If you can answer with something concrete, recent, and non-trivial, this system is already running in you.

If you cannot, the problem is not AI, and it is not the era. It is you. And the problem was always there. It simply had not been exposed before.

---

Appendix · Starter Reading List

The four core reasoning tools mentioned in Chapter 6 each have classic entry-level books. They are arranged from easier to harder. Start with the first one.

Probability and uncertainty reasoning

• Thinking, Fast and Slow, Daniel Kahneman
• What Intelligence Tests Miss, Keith Stanovich

Causal reasoning

• The Book of Why: The New Science of Cause and Effect, Judea Pearl
• Causal Inference: The Mixtape, Scott Cunningham

Game theory and incentive structures

• The Strategy of Conflict, Thomas Schelling
• The Elephant in the Brain, Robin Hanson and Kevin Simler

System dynamics

• Thinking in Systems, Donella Meadows

Feedback exposure and judgment quality, the theoretical basis for Chapter 6’s fourth item

• Expert Political Judgment, Philip Tetlock
• Superforecasting, Philip Tetlock and Dan Gardner

A practical note for reading: after finishing each book, do not immediately start the next one. Spend 1 to 2 weeks deliberately using that book’s way of thinking in daily judgments. Without this step, finishing the book is close to not reading it at all. The knowledge has not run through actual reasoning in the mind, so it will not become a reflex.

---

References and Sources

Sources for the main factual claims in this essay.

AI capability divergence, Chapters 1 and 2

• Andrej Karpathy’s X post about the “growing gap in understanding of AI capability” from April 2026. The original post discusses the bimodal difference between free ChatGPT users and paid Claude Code / Codex users, as well as the technical explanation involving reinforcement learning and verifiable rewards. A secondary report: The New Stack, “Karpathy says developers have ‘AI Psychosis’” (April 2026).
• The “walk or drive to the car wash” example was not proposed by Karpathy. It was a viral test that circulated independently on Threads / Twitter in February and March 2026 and was covered by several outlets, including Newsweek and Cybernews.
• SimpleBench: simple-bench.com (Philip AI Explained, 2024), 213 multiple-choice questions, with a human baseline around 84%, remaining above SOTA models.
• BrainBench: “Exposing the Commonsense Reasoning Gap in Large Language Models” (arXiv:2603.14761, 2026), a systematic account of LLM failure modes in common-sense reasoning.

Rumination chains and knowledge management, Chapters 3 and 4

• Google effect / digital amnesia: Sparrow, Liu, Wegner, “Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips,” Science 333 (2011): 776-778. Note that the effect size has been disputed in later replication work.
• Karpathy’s LLM Wiki scheme: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f (published April 3, 2026). The wiki scale mentioned is roughly 100 articles and 400,000 characters.
• Luhmann’s Zettelkasten system: roughly 90,000 cards accumulated over a lifetime, producing 70+ books and 400+ papers. See zettelkasten.de and Sönke Ahrens, How to Take Smart Notes (2017).

Generalists, taste, and cognitive folding, Chapter 5

• Dan Koe’s generalist thesis: several essays at thedankoe.com, including “The Rise of the Generalist” and “The Future of Work.”
• Paul Graham’s theory of taste: paulgraham.com/taste.html, “Taste for Makers” (2002).
• Rick Rubin, The Creative Act: A Way of Being (2023).
• Cognitive decoupling, reflective mind, and the low correlation between IQ and rational judgment: Keith E. Stanovich, Rationality and the Reflective Mind (Oxford, 2011) and What Intelligence Tests Miss (Yale, 2009). The Stanovich & West line of work shows that correlations between thinking dispositions and IQ are usually below 0.30.
• Fluid intelligence and structural abstraction: François Chollet, “On the Measure of Intelligence” (arXiv:1911.01547, 2019).

Plasticity of cognitive ability, Chapter 6 item I

• IQ heritability and the Wilson Effect: Haworth et al., “The heritability of general cognitive ability increases linearly from childhood to young adulthood,” Molecular Psychiatry 15 (2010): 1112-1120; Plomin & von Stumm, “The new genetics of intelligence,” Nature Reviews Genetics 19 (2018): 148-159. Adult IQ heritability is commonly estimated at 0.70 to 0.80.
• Brain training and the failure of transfer to fluid intelligence: multiple meta-analyses by Melby-Lervag & Hulme and others; Simons et al., “Do ‘Brain-Training’ Programs Work?” Psychological Science in the Public Interest (2016).

Expert prediction and feedback exposure, Chapter 6 item IV

• Philip E. Tetlock, Expert Political Judgment: How Good Is It? How Can We Know? (Princeton University Press, 2005/2017). From 1985 to 2003, the study tracked 284 experts across fields and collected 27,451 verifiable predictions. Core conclusion: expert predictive accuracy was close to random, while experts’ self-assessments of accuracy were far higher than actual performance.

Frontier AI benchmarks, Chapter 2

• AIME (American Invitational Mathematics Examination): an olympiad-track mathematics benchmark.
• GPQA (Graduate-Level Google-Proof Q&A): a benchmark of PhD-level science questions.
• SWE-Bench: a benchmark of real GitHub issue software-engineering tasks.
• These benchmarks were approached or reached at expert level by frontier models in 2025-2026, including OpenAI o-series models, Anthropic Claude 4/4.5/4.6-series models, and Google Gemini 2.x/3.x-series models.

---

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