Editor’s Note:
This is Tabbit’s new column, “Tinsight.” We share thoughtful long-form pieces that may be lengthy but ask sharp questions and offer deep thinking. This article, from offbook.press in April 2026, describes a dilemma many people feel but struggle to name: they consume huge amounts of information every day yet feel more anxious; they appear to be learning, but are often only managing emotions. It offers an actionable framework for seeing what is worth training, what should be abandoned, and how to find one’s position in an era of divergence. The full piece is about 20,000 Chinese characters and will be published in four parts. Today is part one.
▼
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 cognition is.
Chapter 1 · Four Phenomena
Looking back at information consumption over the past year, several phenomena appear together. Each one alone is not surprising, but together they reveal an under-discussed structural fact.
First, the four phenomena.
1 · The Regurgitation Chain
Much AI-related content on Xiaohongshu, WeChat, and X is produced through the same chain.
Original material such as a paper, code snippet, experiment result, or first-hand observation is first summarized by one account using AI, then selected and packaged by another marketing account, rewritten again by the next account, and finally pushed to readers.
You have seen titles like “Everything changed again,” “Gemini went wild,” “Mass unemployment countdown,” or “OpenAI just revealed…” Each creates urgency, but the article is often empty.
This is the regurgitation chain.
At every layer, three things happen at once:
Information density falls as details, conditions, and uncertainty are eaten by summaries.
Errors accumulate as every summary adds small deviations that reinforce one another.
Emotional intensity rises because every layer must add drama to survive algorithmic distribution.
What reaches the reader is processed, less nutritious, and easier to digest: prepackaged junk food in content form.
In the past this chain was human and slower. Today, massive amounts of regurgitated content are being reinvented and manufactured quickly.
When people think they are “getting information,” they often touch the end of a chain already three or four layers removed from the original source. This sets up the next phenomena.
2 · Tooling the Consumption of Regurgitation
The more important thing is not the chain itself, but consumers’ response to it.
Faced with endless regurgitated content, some people do not consume less or seek primary sources. They ask for better tools to consume more.
OpenClaw has many abilities, but a widely shared use case is not coding or data analysis. It is “summarize what happened in AI today.” Users proudly show workflows where AI reads dozens of sources, creates summaries, and pushes key updates.
This looks like efficiency, but it is a strange reinforcing loop: one AI tool helps me consume another AI’s regurgitated output more efficiently.
Readers shift from passively receiving regurgitation to actively optimizing the pipeline for consuming it. It looks more active, but mainly expands volume.
3 · The Producer Divide
Meanwhile, another group is doing something entirely different.
Andrej Karpathy described this divide precisely on X.
One group mainly uses free or old ChatGPT and sees AI as hallucination and low-quality content, because that is what their feeds show them.
Another group spends hundreds of dollars a month using Claude Code or Codex for real technical work and sees systems refactor codebases or find vulnerabilities in an hour.
These two groups live in very different realities.
The first judges AI by summaries at the end of the regurgitation chain. The second judges AI through real-time collaboration with models.
They use the same word, AI, but mean different things. “AI is not that impressive” and “AI is exploding” can both be true. The difference is what the user is doing.
4 · Seeing the Three Layers Together
Each layer alone is a local observation. Together they expose a sharper question.
People consuming regurgitation are not lacking information.
They encounter more information than almost anyone in history. They are not “unable to learn” either, because the products are not designed for learning. What they really do is manage the fear of falling behind by consuming information.
People doing real cognition are not consuming more; they are doing something different. They push, test, verify, and form their own judgments. Regurgitation consumers read finished products, remember keywords, close the tab, and move on.
Together, these phenomena point not to information overload or AI making people lazy, but to a sharper fact: most people mistake knowledge, experience, and fluency for cognitive ability.
What is real cognitive activity?
First, exclude common misidentifications:
It is not merely the brain running.
It is not reading and remembering a lot.
It is not long experience in a field.
It is not fluent expression.
These were mistaken for cognition because they often appeared alongside cognition in older environments.
Real cognitive activity means:
Producing structural judgments that can be calibrated against reality.
Forming specific judgments that later facts can verify or refute.
Identifying deep structure in an unfamiliar problem.
Treating the distance between your current judgment and reality as something you can operate on.
This is what the AI era isolates and prices separately. Most people never noticed the distinction, so they never had the chance to practice it.
Older evaluation systems did not require it. Experience, patterns, fluency, memory, expression, and social sensitivity could all be rewarded without real cognition. The AI era is the first time cognition itself is separated from those abilities and priced on its own.
Once knowledge is flattened by models, repetitive tasks are absorbed, and polished judgments cost almost nothing to produce, what remains is structuring reality, spotting errors, and abstracting unfamiliar problems. Many people simply do not have that capacity, even though past success made them feel capable.
This is not a new problem. It is an old fact hidden by old evaluation systems, now exposed by new technology.
Chapter 2 · AI Is Not Strong Evenly
Before the main topic, one common misunderstanding needs to be addressed.
The misunderstanding is that AI has surpassed humans across the board, so it is natural for most people to feel small and consume regurgitated content passively.
That judgment is half right and half wrong.
Where AI is strong: on structured, measurable tasks, AI has surpassed median humans and many experts, including benchmarks such as AIME, GPQA, SWE-Bench, and language understanding tests.
In knowledge retrieval, fluent expression, and instruction following, AI is far beyond most people. The best use is as a lever that multiplies capable people’s output by more than tenfold.
Where AI is weak: evidence also shows AI strength is sharp and uneven. Change a problem’s surface while preserving its structure and models can collapse, as benchmarks such as SimpleBench and BrainBench show.
AI also struggles with tasks that lack clear right answers: judging design and decisions, handling messy reality, and abstracting from completely unfamiliar structures.
Why is it uneven?
There is a technical reason. Frontier models improve largely through reinforcement learning, which calibrates them with right-or-wrong signals. This works extremely well where rewards are verifiable, such as code, math, and tests.
By contrast, areas without clear right answers, such as judgment, nuance, and aesthetics, cannot provide reliable reward signals, so progress is slower.
Commercial incentives compound this: code and structured tasks bring B2B revenue, so compute and research resources concentrate there. AI grows exponentially in verifiable, commercially valuable areas and more slowly elsewhere.
The real conclusion: AI is an amplifier, not an equalizer. It amplifies what structure the user can provide. For someone who can decompose problems and judge outputs, it multiplies production. For someone who only asks it to “write something,” it amplifies anxiety management.
Most important: AI has not replaced problem construction.
A reasonable objection is that AI already shows structural abstraction in reinforcement learning and systems such as reasoning models or AlphaGo.
That observation is correct, but a crucial distinction matters.
AI’s structural abstraction happens under a given reward function. In closed spaces with clear goals, such as Go, math, or runnable code, AI can find better structures than humans. This is instrumental decoupling inside a defined game.
But another layer is deciding what is worth pursuing and which questions deserve to be asked. What hidden human need does a product solve? What backlash will a policy trigger? What does messy user feedback really say? This layer has no ready-made reward function.
AI can generate ten thousand logical structures, but judging which one corresponds to a real-world pain point must be done by an embodied subject inside reality who bears consequences.
So AI takes over finding optimal structures inside defined games. It does not take over deciding which game to play.
At the level of division of labor, what question is worth asking, which direction is right, and whether an output solves the problem still depend on humans. These are exactly the things that require real cognition.
So “AI is too strong and I cannot keep up” is a lazy story.
AI has surpassed you in things you should give up, such as memory, fluent expression, and instruction execution. But it has not surpassed you in judgment, construction, aesthetics, and structural abstraction. Those who give up where they should participate are the ones truly left behind by AI.
To be continued...
Chapter 3 · The Human Position in the Regurgitation Chain: Learning or Anxiety Management?
Chapter 4 · A More Hidden Trap: Knowledge Management
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 · Closing
Chapter 7 · Closing
