Editor's note:
Welcome to Tabbit's Tinsight column, where we share in-depth pieces worth spending time with. Previously we shared the first part of The Real Divide in the AI Era, published by offbook.press in April 2026, Are You Consuming Information or Starting Cognition?, and the second part, The Illusion of Learning: When Knowledge Management Becomes an Anxiety-Relief Ritual. This issue brings you the third part. By unpacking concepts such as generalists, taste, and cognitive folding, the author explains which abilities become the real points of differentiation once AI starts flattening knowledge, memory, and fluent expression.
Author: Dawei Geng. Source: https://offbook.press/essays/on-cognitive-decoupling/
Most people mistake knowledge, experience, and fluency for cognitive ability itself. They have never identified what real cognitive activity is.
Chapter 5 · What They Got Right: Generalists, Taste, and the Refolding of Intelligence
By this point, we have taken apart the two forms of ruminative consumption: passively scrolling information and actively organizing knowledge. They look different, but at bottom both are anxiety management rather than learning. What emerges is the fundamental question that has been avoided all along: what abilities actually matter?
This question had been touched many times even before AI. In different circles, several popular explanations describe the same thing from different angles, but each captures only part of it and none realizes it is describing the same underlying object. Put them together, and a more complete picture appears.
▌ Generalism
In recent years, the idea of the generalist has returned to the foreground among creators and indie developers. The most frequently cited version comes from Dan Koe. His core claim is that specialization is depreciating, while multidisciplinary generalists are becoming rarer and more valuable.
What he gets right: breadth is not dabbling. A real generalist is not someone who knows a little about everything, but someone who has accumulated enough depth across multiple fields to see isomorphic structures between them. Someone who has deeply studied economic incentives, biological evolution, and organizational behavior will find that the underlying mechanisms overlap heavily, because all three ask what emerges at the system level when multiple actors pursue their own goals under constraints. Recognizing this cross-domain isomorphism is the true value of the generalist, and the only path by which breadth becomes judgment.
What he does not make clear: what kind of depth counts. This is the blurriest part of generalist theory, and the place followers most easily fall into a trap. Some people follow the advice by expanding their reading, finishing one introductory book in each of ten fields, and ending up with ten surface narratives in their heads. That kind of depth is useless for cross-domain mapping, because surface narratives are different by nature and have no isomorphism to map. What can truly be mapped is deep structure: the causal mechanisms, incentive constraints, and feedback loops that actually determine behavior in a domain. One workable test for depth is this: can you name one thing practitioners in a field collectively believe that is actually wrong? If you can, you have depth. If you cannot, you are still at the level of the standard narrative.
The gap he leaves: generalist theory does not isolate the underlying mechanism that performs cross-domain mapping. It reads like a methodology: read across several fields and build connections. But it hides a premise: the reader must already be able to make those mappings. That ability itself is what is scarce. It explains why the same methodology works for some people and not for others. The difference is not effort; it is mapping ability itself.
▌ Taste
Another line comes from creators and design circles: the taste repeatedly discussed by Paul Graham, Steve Jobs, and Rick Rubin. In Chinese, taste is easily understood as a vague aesthetic feeling, but what they are really describing is a more specific mechanism.
What they get right: before you can clearly explain why, you can already tell whether something is good. This pre-verbal judgment is not mysticism. It comes from the long-term internalization of many high-quality samples. A person who grew up in museums can look at a painting and judge its quality within seconds, not because they know art history, but because their visual system has been calibrated by a large number of excellent works. The product of that calibration is taste.
But what they describe is low-order taste: pure pattern recognition. A person can judge that a typeface looks bad, but cannot explain why. This taste is real and useful. It lets someone quickly filter many options in a familiar domain. But it has a serious limitation: it cannot be transferred, taught, or verified. You cannot pass your taste to someone else, and you cannot build taste from scratch in a new domain without enough samples.
What they do not make clear: taste has two layers. Low-order taste is pre-verbal pattern recognition. High-order taste is pattern recognition plus the ability to unpack structured reasons when needed. A truly strong designer can not only say that a typeface looks bad, but also explain that its x-height and weight ratio break the visual rhythm. High-order taste matters because it can be taught, transferred, and used to accelerate the formation of new taste in new domains. It makes the structure behind intuition explicit, and therefore learnable.
At bottom, high-order taste and generalism are the same ability: abstracting a concrete judgment into an operable structure, then using that structure somewhere new. The difference is only the angle. Generalism discusses it through cross-domain mapping; taste discusses it through aesthetic intuition. They are two sides of the same mechanism.
▌ Cognitive Folding
A third line has become increasingly common in tech and indie creator circles. Its core claim is blunt: AI will not narrow gaps; it will widen them at unprecedented speed. The usual analogy is the Industrial Revolution. Steam engines and later assembly lines sharply reduced the relative value of manual labor, and the gap between people who could design machines, organize production, and coordinate systems and people who could only sell physical labor moved from linear expansion to exponential divergence. Now it is cognition's turn. AI is flattening a group of cognitive abilities: knowledge, memory, retrieval, and fluent expression. The remaining parts that are not flattened will be explosively amplified. The result is not simply a wider wealth gap, but a folding of different groups' output, within the same unit of time, into entirely different orders of magnitude.
What they get right: the gap will widen, and the speed will be unprecedented. This judgment is correct, and the Industrial Revolution analogy is apt. After every general-purpose technology leap in history, the gap between those who mastered the new leverage and those who did not has not expanded linearly; it has diverged exponentially. This was true of steam, electricity, and the internet. AI will be more extreme, because for the first time the technology acts directly on cognition itself, and cognition is the source of judgment, creation, and decision-making. The amplified part will be more visible than in any previous transition.
What they do not make clear: exactly which cognitive ability is being amplified.
This is the blurriest part of the argument, and the place most followers fall into a trap. Their default assumption is usually high cognition equals high IQ, so the conclusion becomes that high-IQ people will become richer while low-IQ people will be left further behind. The first half of that conclusion is roughly right, but the reason is completely wrong.
What gets amplified is not IQ. IQ measures raw processing power: working memory, processing speed, information extraction, rule-based inference. AI is already stronger than most people at these. Someone with high IQ who never engages in real cognitive activity may be eliminated fastest in the AI era, because the things they used IQ for in the past, such as fast learning, memory retrieval, and fluent reasoning, are now done faster, more accurately, and more cheaply by AI. Their advantage is flattened directly.
What is truly amplified is another layer: identifying deep structure in an unfamiliar problem, judging what question is worth asking, and carving an operable object out of ambiguous reality. This is related to IQ, but it is not the same. A person with average IQ who identifies what real cognitive activity is and keeps investing in it will be amplified by AI more than a high-IQ person who spends their life inside the comfort zone of pattern recognition.
The gap they leave: high cognition is treated as an innate, fixed attribute. That makes the whole folding narrative sound like an irresistible force: high-cognition people win, low-cognition people fall behind, with no middle ground and no room to act.
But the real divide is not an innate divide between high cognition and low cognition. It is an awareness divide between those who have identified it and those who have not. The former sounds like an unchangeable fate; the latter is a threshold that can be crossed. Crossing it is not easy, but it is open in principle. Much of the despair in cognitive folding theory comes from mistaking a crossable threshold for an unchangeable innate trait.
▌ Three Lines, One Thing
Put the three lines together.
The cross-domain mapping ability described by generalist theory, the underlying mechanism that turns depth across fields into judgment, is the ability to strip away the appearances of multiple domains and grasp their shared deep structures. Without this ability, ten fields produce only ten piles of fragments. With it, three fields can generate new judgment.
The high-order taste described by taste theory, the ability to make the structure behind intuition explicit, is the same ability operating on top of many internalized samples. Low-order taste needs only samples; high-order taste needs samples plus the ability to abstract patterns into structures.
The layer described by cognitive folding theory as not replaced by AI but amplified by it is also this ability. After AI flattens knowledge, memory, fluent expression, and instruction-following, the only thing not flattened is the act of stripping reality out of its appearances and operating on it as structure. That is what gets amplified.
The three lines touch the same thing from completely different angles, and simply give it different names: generalism, taste, folding. The core is the same: the ability to abstract concrete situations into operable structures.
Cognitive science has a name for it: cognitive decoupling, the ability to detach a representation from the reality it refers to and operate on it as an independent object. All abstraction, hypothesis, counterfactual reasoning, and self-examination are built on it. The three popular explanations describe it from different angles, but each only reaches part of it.
This ability has always existed, but it was never priced separately. In the pre-AI era, it was mixed together with many other abilities: knowledge reserves, memory, fluent expression, and proficiency. A smart person usually had many of these at once, but no one knew which one was truly doing the work. So every popular explanation was like feeling an elephant in the dark: whichever part it touched, it described in that part's language.
The AI era makes this ability visible for the first time. AI replaces knowledge reserves, because it knows more than anyone. It replaces memory, because it can retrieve on demand. It replaces fluent expression, because it writes better than most people. It replaces proficiency, because it does not tire, err, or need practice. Once these abilities are peeled away one by one, what remains is the underlying mechanism that had been hidden all along: the ability to cut reality into structures, judge which structure is right, and build new structures in unfamiliar domains.
This is what generalism, taste, and cognitive folding all point toward. It is not a new ability, but an existing ability that has been independently named for the first time.
▌ What AI Can Do, What Remains for People, and What to Learn
Push this judgment into practical terms.
The range of things AI can do is expanding quickly: pattern synthesis, information retrieval, fluent expression, instruction-following, and complex tasks in verifiable-reward domains such as code, math, and structured reasoning. This range expands every few months, with no obvious boundary.
The range of things AI cannot do is shrinking, but there are several things it cannot do now and will not do in the near term: carve ambiguous reality into operable problems, meaning problem construction; identify structural errors in its own output, meaning meta-judgment; judge what is good without an external reward signal, meaning aesthetics and value trade-offs; and perform real abstraction on completely unfamiliar structures, meaning decoupling itself.
The first three depend on the last. Without real decoupling ability, the first three are only interpolation simulations in a high-dimensional representation space. They look similar, but are not the same in essence.
So what truly remains for people in the AI era? The layer that directs the model: asking the right question, judging whether the output actually solves that question, pulling the model back when it drifts, and making the value decision on the final result. All of that work rests on cognitive decoupling as the underlying mechanism.
Then comes the key question: what exactly should we learn?
Not more knowledge, because models know more than anyone. Not more fluent expression, because models write better than most people. Not more thinking frameworks, because most frameworks on the market are rhetoric dressed as tools.
What we really need to learn is a small number of concrete things: a few core reasoning tools that let judgment work without relying purely on intuition; several regularities from fields outside our own so cross-domain mapping has something to map; environments where reality can calibrate us so all of this does not slowly decay. Add to that the partly untrainable underlying mechanism: cognitive decoupling.
The four elements are independent and interactive. Together they form a complete system.
To be continued...
Chapter 6 · The Four-Part Formula, Explained Item by Item
Chapter 7 · The End
