AI Does Not Think for Me. It Makes My Intuition Refutable.

2026-09-04 AI · thinking · cognition

Beginning with a Series of “No”s

This essay grew out of a series of “no”s.

After looking at my projects, AI initially described me as someone who programs with the help of models. No, I said. I cannot really write code; I am simply good at getting from a model the capabilities needed to get something done.

It treated the ad system’s 64,000 QPS as a performance test. No, I said. That was real production traffic; but it does not mean the entire business chain ran steadily at that number over time.

It tried to place both the Shang Han Lun mini-program and the cloud follow-up consultation tool under a single line of “meta-research.” Again, no. The former was made to structure and query material; the latter was for learning and accumulating material. Neither began with a grander mission.

It suggested that the novelty of the MRI project might lie in combining a knowledge graph, formal proof, and engineering experiments. That still did not feel right. Most of those components were not new. What I cared about was starting from a category-theory-like intuition, trying to transfer structures from other fields, and then decomposing the resulting questions into logical and engineering units, each routed toward its own endpoint: formalization or reproducibility.

Each correction added only a little. After several rounds, however, the shape of the question had changed completely.

Strangely, before saying “no,” I could not necessarily give the correct version directly.

Compression

I describe my own thinking as fuzzy thinking, parallel thinking, or thinking without symbols. I also describe myself as having aphantasia. But these are descriptions of subjective experience, not conclusions from neuroscience; I cannot establish any causal relationship among them.

All I can describe is the process.

This pattern existed before AI.

Since childhood, I have liked thinking about metaphysical questions, but I have never been especially at home in situations that require long chains of reasoning. Education often asked us to unfold a problem layer by layer according to prescribed steps, so that each one could be repeated and checked. My own thinking seems to lean more toward pattern recognition: first sensing that several things share a certain shape, then trying to compress many differences into a smaller structure.

This is only my summary of long experience, not a conclusion from cognitive testing. But it accounts for a strong impulse of mine: I always want to remove the vocabulary of a field and its intermediate details, and see what remains.

Compression can sometimes reveal a hidden relation. It can also skip over an intermediate layer that truly matters. When I compress two problems into the same shape, I may have found a shared structure; I may also simply be bad at preserving their differences.

AI happens to complement this way of thinking, while amplifying its risks. I offer a highly compressed direction, and the model can unfold the intermediate steps for me. But if the initial compression has already discarded a crucial condition, it may also supply a fluent, complete, and wrong line of reasoning.

Fermentation

Compression describes the structure of how I process a problem. Fermentation describes how an idea forms over time. They often occur together, but they are not the same thing.

Once a subject enters my mind, it often leaves conscious awareness. I also find it hard to advance every day along one continuous path that I could report. Most of the time, I do not even feel that I am still working on it.

After some time, it may suddenly appear as a relatively complete direction, accompanied by a strong urge to make it real. I call this process “fermentation.” The name does not prove that the unconscious has been continuously calculating. It describes only this: the middle of the process is invisible to me, while the result seems to arrive all at once.

For me, it is not merely an occasional flash of inspiration, but a recurring rhythm of thought:

A problem enters
→ temporarily leaves awareness
→ fermentation
→ a whole emerges
→ a strong urge to implement it
→ AI catches it and helps unfold it
→ correction through a sense of drift

At that point, what I usually have is not a set of sentences, but a sense of direction.

I do not know how to state it all at once. But when an external version appears, I can feel where it has been pulled off course: it turns a part into the whole, writes a later explanation as the original purpose, or replaces “what I want to study” with “what sounds more like research.”

So what I often know first is not the answer, but why this answer is not yet it.

AI Gives Intuition a Surface to Collide With

Without AI, this kind of intuition can easily remain internal.

It cannot be searched because it has no keywords; it cannot be discussed because it has no sentences; it cannot be implemented because it has no steps. It is even difficult for me to refute: a feeling that has not been clearly expressed can always retreat into vagueness when challenged.

What AI changes is not intuition itself, but the cost of externalizing it.

It can first offer five explanations that are not quite the same, split an overall direction into a series of questions, or translate a judgment into a model, code, or experiment. It does not need to be right on the first attempt. As long as the candidates are concrete enough, I have something to compare: which sentence comes close, which level has been silently substituted, which example is beautiful but does not belong to the question.

The process is roughly this:

A direction cannot yet be fully expressed
→ AI generates candidates
→ I sense a drift in the facts, intent, or center of gravity
→ use counterexamples to explain “why not”
→ AI generates new candidates
→ implicit constraints gradually become visible

The most important product here is not any single answer. It is that an intuition once private finally has a surface to collide with.

It can be rewritten, compared, and shown not to be what I actually meant.

But AI Is Not a Faithful Printer

It would be equally inaccurate to stop here and say: “The thought already existed, fully formed, in my mind. AI merely printed it for me.”

Models offer words I had not thought of, suggest new classifications, connect material that had been separate, and sometimes provide a counterexample that forces me to abandon my original judgment. At times I think I am only looking for an expression, and later discover that the question itself has changed.

The dual-track framework is one example. It was explicitly articulated in the MRI project: logical problems terminate in formalization, engineering problems terminate in reproducibility, and questions that enter neither class are not adjudicated by this method. In retrospect, Minimax’s use of random baselines and cross-validation to rein in optimistic results can be understood as an experience on the engineering track. I cannot prove that it directly caused the later framework. That connection was not a plan I wrote down before the project began. It emerged only later, while discussing the project’s history.

There are at least three possibilities.

First, I may already have had a relatively complete structure in mind but been unable to report it in language; AI helped externalize it.

Second, I may initially have had only a direction, an attraction, and a sense of drift; the concrete structure may have gradually formed through repeated comparison.

Third, the model and I may have constructed it together: I supplied the pull of the question and the standards of choice; the model supplied concepts, language, counterexamples, and connections. Beforehand, the final view did not belong fully to either of us.

At present, I have no way to establish which explanation is closer to the truth. To read a fully formed later essay back into the past and say that I had “already thought it through” from the beginning would likely be a retrospective story.

So AI is not merely a tool, but neither has it simply thought in my place.

It participates in the environment in which thinking happens.

“This Is What I Mean” Does Not Mean “This Is True”

Being sensitive to drift can easily be made to sound like a mysterious ability: as though, even if I cannot explain myself, I can always recognize the right answer in a model’s output.

That is wrong too.

What I can often judge quickly is whether an expression departs from my current direction. Was the ad system a production system? Why was a tool originally made? What did I actually care about in the MRI project? These questions concern my experience and intent, and I am entitled to correct them.

But “this is what I mean” only accomplishes alignment of intent.

It cannot prove that what I mean corresponds to reality. A cross-domain structure may be only a habit of my own thinking. An explanation that strongly resonates with me may have no novelty. A Lean proof and a set of programs may both rest on a mistaken definition.

At a minimum, we need to distinguish two kinds of correctness:

I can sometimes discover the first kind of error through a sense of drift. The second requires literature, counterexamples, formal checking, independent reproduction, and feedback from reality. Neither can substitute for the other.

My intuition can examine whether it has been replaced. It cannot issue itself a certificate of truth.

From Fluent Answers to Objects That Can Fail

The most dangerous thing about AI is that it can complete a thought that has not yet formed.

It can fill vagueness with clarity, gaps with derivations, and accidental connections with a unified framework. The more fluent the prose, the more formal the mathematics, and the more runnable the code, the easier it is for me to forget that completeness may be only the model’s style, not a property of the object.

So I now prefer to keep a few points of friction.

First, I keep the shortest record of an idea when it first appears, so later explanations do not completely cover it. I ask AI to produce candidates that conflict with one another, rather than merely polishing one version until it becomes prettier and prettier. Each time I say “no,” I try to retain the counterexample and the constraint that was violated. Finally, I give “is this what I mean?” and “is it true?” to different forms of checking.

An essay should accept criticism from readers. A program should run in reality. A mathematical proposition should be allowed to be rejected by a formal system. A research project should face random baselines and independent reproduction.

Once a vague intuition acquires a form, it loses the right to hide.

This is AI’s most important role for me: it does not guarantee an answer. It finally lets me be wrong in a sufficiently specific way.

Whose Thought Is This?

Including this essay, the prose is mainly expanded by AI.

But without my repeated indications of drift, it would write another essay that was just as fluent and not mine. Conversely, without its candidates, concepts, and rebuttals, I could not honestly say that these distinctions had already existed fully in my mind.

So “whose thought is this?” may not be a simple question of ownership.

The pull of the direction, the feeling of error, and the things I have cared about over time come more from me. The expansion of expression, unexpected connections, and part of the conceptual structure come from the model. The final version forms through continued choice, rejection, and rewriting.

But once I decide to publish it, responsibility cannot remain suspended between us.

AI can participate in generating it. It cannot take responsibility for its consequences in my place.

In the past, many vague intuitions could remain inside me. They may have been profound; they may only have been illusions. No one knew, including me.

Now they can leave me and become a sentence, a piece of code, a system, or a research project—then be refuted by others, by formal systems, or by reality.

That does not necessarily bring me closer to being right.

But it finally gives me a chance to know where I was wrong.