Editor's note:
Welcome to Tabbit's Tinsight column, where we share in-depth pieces worth spending time with. In the previous issue, 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? This issue brings the second part, perhaps the most painful part of the whole series.
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 3 · Where People Sit in the Rumination Chain: Learning or Anxiety Management?
If AI is an amplifier rather than an equalizer, then widening output gaps between different users is an inevitable conclusion: the amplifier magnifies the differences already present on the user side. But this is still not enough to explain why the gap between two groups is so large and why the directions are completely opposite. One group uses AI to build complex systems; the other is not merely using AI less, but consuming more and more through AI while producing less and less.
This reverse dynamic needs an additional explanation. It is not a problem of weak ability or poor usage technique. The real reason is that at the end of the rumination chain, most people are not learning at all. They are doing something that looks a lot like learning, but operates by the opposite mechanism.
▌ Learning and Anxiety Management
First, put the definitions of the two activities side by side.
Learning is a closed loop: encounter something you do not understand, first form a preliminary judgment yourself, search for information with a specific question, use the information to calibrate your judgment, and form a new understanding that belongs to you.
All four actions are necessary.
The prior judgment and calibration action are the core of the entire process. Without them, incoming information is not absorbed by any existing structure. It is like pouring water where there is no container: after a while it dries up and leaves nothing behind.
Anxiety management is a different loop: feel behind, consume information, get the feeling that you are keeping up, feel temporarily relieved, then because nothing has been internalized, feel behind again after a while and consume again.
This loop has no prior judgment, no calibration, and no internalization. Its operating logic is completely different from learning. The product of learning is cognitive structure; the product of anxiety management is emotional relief.
The two look similar on the surface because both involve reading something. Mechanically, they are opposites. More troublesome still, the brain cannot distinguish between reading a conclusion and arriving at one yourself.
When you read a well-organized summary, the brain produces a feeling of now I understand this, almost identical to the satisfaction of deriving a conclusion yourself. But the former is only a momentary illusion. A few days after closing the article, nothing remains. The latter is real cognitive internalization, and it changes how a person judges the subject over the long term.
This mental mechanism is the basis that keeps the whole rumination chain running. It makes anxiety management feel like learning, so consumers receive daily confirmation that today was not wasted, while in reality nothing has truly been absorbed.
▌ The Real Purpose of the Rumination Chain
Chapter 1 already explained the mechanism of the rumination chain: every layer of processing lowers information density, accumulates error, and raises emotional intensity. But the more important fact is that the design goal of this whole production chain is not to transmit information, but to retain attention.
This distinction matters. Something produced to transmit information, such as an academic paper, a rigorous empirical report, or technical documentation, is designed to preserve the content that lets readers truly understand and use it: definitions, conditions, counterexamples, and uncertainty.
In the attention economy, these things are liabilities, because they reduce reading fluency and increase cognitive load. So something produced to retain attention systematically removes them by design.
That means that even if you spend a great deal of time reading such content carefully, you still cannot learn from it. It is not that readers are not trying hard enough. The product was not designed to help people learn. The more seriously you read it, the more you are circling inside a room with no exit.
▌ The Person's Position Inside the Loop
Now overlay the anxiety-management loop with the rumination chain, and we can describe precisely what most people do every day:
Open an app, scroll to a finished judgment with a hook title, read it quickly, remember one or two keywords or emotional impressions, close it, gain the feeling that you kept up today, and briefly reduce anxiety.
No step in this process is learning. There is no prior judgment, because before reading you do not know what you are trying to verify. There is no calibration, because the information is not compared against any existing judgment. There is no internalization, because after closing it, nothing remains and you cannot restate it when asked later.
Every action is emotional regulation: using one contact with information to complete the psychological ritual of I have not fallen behind.
Here is a cruel verification method: find someone who reads AI news every day and ask them to restate the core point of anything they read three days ago, along with their own judgment of it.
The vast majority cannot do it.
This is not a memory problem. It is because that content was never processed, so there is nothing to remember. The brain only remembers what it has processed itself: what has been judged, refuted, or compared against existing understanding. Information that passively flows by, no matter how information-rich it feels at the time, does not exist a few days later.
Compare this with people doing real AI-related work and you find the opposite phenomenon. They do not consume ruminated content, because their daily experiments, model feedback, and code debugging already provide information density far higher than any secondhand summary. Their judgment about what a model can and cannot do comes from something they personally failed to make it do yesterday, not from someone else's review.
Even more counterintuitively, the more people consume rumination, the more behind they feel; the more people do real work, the clearer they feel. This is not because the doers know more. It is because the direction of information processing in the two activities is completely opposite.
Anxiety management takes input without output, turning information into emotional consumption. Real work processes input and produces output; every act of processing reinforces cognitive structure. The former group's impression of AI is, there is too much, I cannot keep up, something new came out again. The latter group's impression is, yesterday I found it is especially good at X and still weak at Y: specific, clear, and bounded.
The two groups use the same word, but the mental objects behind it are entirely different.
▌ When Consuming Rumination Becomes an Achievement
Push this diagnosis to the end and you meet a phenomenon worth staring at.
The most widely shared use of tools like OpenClaw is helping me organize what happened in the AI world every day. It looks like an efficiency gain, something that was hard to do before. But think carefully: is this really an achievement worth proudly promoting?
Under the analytical frame above, its meaning becomes this: a tool helps me consume low-density ruminated content more efficiently and continuously, and I treat that as a use case worth sharing. The efficiency of actively swallowing ruminated material itself becomes an expression of productivity.
This is not a problem with one tool. It is the externalization of a deeper collective belief: keeping up with information equals capability. In the pre-AI era, this belief had some basis, because information was scarce and continuous access to first-hand information was itself a rare skill. But after AI drove the production cost of finished judgments to zero, the belief became a pure pseudo-need. The content you can consume is infinite, but none of it is built to help you learn.
So a structurally ironic picture appears. AI turns information supply from scarce to infinite, and some people do not respond by finally moving from consuming information to producing judgment. They respond by saying: I need stronger tools to consume more information. The stronger the tool, the faster the consumption; the faster the consumption, the more dimensions of information are touched; the wider the contact, the deeper the anxiety, because every consumed item implies ten more that were missed.
When the production cost of finished judgments falls to zero, the efficiency of consuming finished judgments becomes the last fortress of pseudo-need. The fortress still stands not because it provides any real value, but because it provides a displayable busyness, a performance that looks like keeping up in the AI era.
People doing real cognitive work have long since left this performance. They do not even watch it.
Chapter 4 · A More Hidden Trap: Knowledge Management
Ruminative consumption is the obvious form: scroll, read, next. But it has a more hidden variant: not passive absorption, but active construction; not browsing information, but organizing it. Because it looks more like proper work, it is more damaging to serious people.
This variant is called knowledge management.
▌ Why It Becomes a Pseudo-Problem in the AI Era
The phrase knowledge management hides an old-era assumption: knowledge is a static object that can be managed, like books in a library, something that can be classified, indexed, and retrieved. The entire knowledge-management toolchain of the past decades, from Evernote and Notion to Roam, Obsidian, and Logseq, is built on this assumption.
But real cognition does not work that way. Knowledge in the mind is not statically stored information; it is a relationship network that is constantly being rebuilt. Understanding is not the same as remembering, and being able to use something is not the same as being able to find it. This mismatch means knowledge management has been doing a mismatched job from the start.
Several mechanisms have happened at the same time over the past few years, turning this mismatched work completely into a pseudo-problem:
The retrieval problem has been solved by AI. One of the core purposes of organizing notes used to be finding things later. Now AI knows almost all textual information in the world, and retrieval cost has dropped to near zero. The traditional goal of organizing for retrieval is already 80 percent invalid.
The trap of externalized memory. Psychology has a phenomenon called the Google effect: knowing that information has been saved makes the brain less likely to truly remember it. This effect is especially obvious among heavy note users. Having it in notes is not the same as having it in the mind. Usable knowledge is the active model in your head, not the tags in your notes. For many heavy note users, active knowledge in the mind is actually degrading, because they treat the brain as an index rather than a workbench.
Organizing notes is the highest-grade anxiety management. It scores extremely high on the anxiety-relief scale: it looks like work, gives a sense of achievement, such as 12 new backlinks today, requires no judgment risk, and provides a feeling of growth without any output. This is more dangerous than scrolling ruminated content, because it has a stronger disguise of doing serious work.
Organizing notes and forming judgment are different activities. Luhmann's Zettelkasten generated the core material for more than 70 of his books not because of the system he used, but because every card contained his own thinking. Each card was a small judgment, a calibration of existing knowledge. The essence of the system was not managing knowledge, but forcing thinking. Most modern Obsidian use is managing other people's thinking: highlights, excerpts, citations. It has nothing in common with what Luhmann did except borrowing the same visual form.
▌ Karpathy's Wiki Approach: What It Really Solves
Against this background, Andrej Karpathy proposed an LLM Wiki approach in early April 2026: put raw materials into a raw/ folder, let an LLM automatically compile them into a structured markdown wiki, and let the human curate only the input. His own wiki for one research topic has grown to about 100 articles and 400,000 words, with almost no manual editing from him.
This approach mainly solves three things: context loss between sessions, the unsustainability of note maintenance because LLMs are better at bookkeeping than humans, and the lack of compounding knowledge because each new source automatically updates multiple existing pages and creates cross-references.
It sounds like a rebuttal to knowledge management is a pseudo-problem. After all, he is doing a form of knowledge management, and it truly works for him.
But one foundational fact determines the real reason it works for him: Karpathy is a scientist.
His raw sources are arXiv papers, experimental results, code, and his own unpublished work. These are things absent from or already stale in large model training data. Model training data has a cutoff; models often do not know the latest developments at the research frontier, and scientists live precisely in the information space after that cutoff. He is not managing knowledge; he is maintaining a delta for model knowledge, adding the frontier on top of the existing base.
This is completely different from the knowledge-management community's idea that I need to manage everything I have read.
▌ Why Ordinary People Fail When They Copy It
Once you understand Karpathy's premise, you can understand why ordinary followers turn his approach into a new form of anxiety management when they copy it.
The content ordinary people care about, such as business news, AI updates, popular science, management methods, and industry analysis, is mostly known better by the model than by them. The knowledge they try to manage is actually a ruminated version of things already present in the model's training data.
What happens when Karpathy's approach is copied in this situation?
AI generates a wiki that looks structured from content that has already been ruminated. This adds another layer to the rumination chain. The output looks more like knowledge, more like a system, more like research, but it is farther from the original information and emptier.
And because a wiki has a stronger visual sense of structure than notes, it creates an even stronger illusion of learning. Watching your wiki grow every day and seeing cross-references increase can feel more like I am growing than scrolling Xiaohongshu. But what is really happening is essentially the same as with Xiaohongshu readers: the person is consuming information they cannot truly call upon.
The only difference is that one consumes on a public product, while the other consumes inside a self-built system. The latter is more cope, not healthier.
▌ What Reasonable Personal Information Infrastructure Looks Like
To avoid this trap, you need to ask a fundamental question: what does the model lack relative to me?
Most knowledge-management practices never ask this question. They only ask what should I record, and then record a pile of things the model already knows more accurately and completely than they do.
There are only two reasonable answers to this question:
First: the delta at the frontier of a discipline. You work in a fast-moving frontier field: research, cutting-edge engineering, or emerging practice not covered by training data. In this case, maintaining a continuously updated knowledge base has real value, because you are accumulating things the model does not know.
Karpathy's wiki approach is suitable for this kind of scenario. The premise is that you are really doing frontier work, your raw sources are first-hand, and you have the judgment to check whether the wiki has drifted.
Second: the delta of your unique personal cognition. You are not working at the frontier, but you have your own judgments, preferences, non-consensus models, and lessons from personal experience about specific things. These are things a general model cannot generate from training data.
In this case, the reasonable practice is a minimal alignment layer: record only what the model does not already know, so that the next time you collaborate with it, it can start from your previous endpoint instead of from zero.
The two approaches look very different, but the underlying logic is exactly the same. Both answer the same question: what does the model lack relative to me? The difference is only where the lack sits in the information distribution for each type of person.
The second approach has a counterintuitive but important property: it cannot be performed. Its entry threshold is the ability to identify what the model does not know. People without that judgment cannot use the method, and will discover that they actually have very little worth recording. People with that judgment will record content that is necessarily useful. This self-selection mechanism is even more valuable than the method itself, because structurally it prevents the method from turning into another rumination ritual.
For most people, the second approach is the reasonable starting point. If, during use, you discover that the model lacks enough knowledge in your field, then consider adding a Karpathy-style disciplinary base. Reversing the order easily leads into the ritualized knowledge-management trap. This is the biggest mistake of the whole knowledge-management community: they advise newcomers to build the base first, so newcomers spend months organizing other people's judgments while cognitive activity is replaced by the act of organizing itself.
▌ A More General Principle
Push the specific topic of knowledge management to the most general level, and you get a principle that can evaluate almost every future new approach:
In the AI era, any methodology that does not require the user to have independent judgment has already become, or is becoming, a variant of ruminative consumption.
The difference is not the methodology's content. It may improve efficiency, promote learning, organize knowledge, or strengthen creativity. The difference is whether the user has outsourced judgment to the methodology.
This also explains a common phenomenon: every popular methodology is, in essence, a collective ritual of cognitive outsourcing. The more popular a methodology is, the more it shows that what it satisfies is not a specific tool need, but the need to avoid judging for oneself. Truly effective methodologies are often not popular, because they depend too heavily on the user's specific context and cannot be batch-copied.
Karpathy's wiki approach itself is effective for him. But after it became popular, the wave of Karpathy Wiki tutorials, Second Brain 2.0 guides, AI knowledge-management workflows, and so on has already started turning into a new round of performative solutions. The test is simple: how long can you use this approach before it starts making you feel, I do not need to judge personally? That moment is when it begins to rot.
To be continued...
Chapter 5 · What They Got Right: Generalists, Taste, and the Refolding of Intelligence
Chapter 6 · The Four-Part Formula, Explained Item by Item
Chapter 7 · The End
