AI Intelligent Body Explodes, Why Does It Always Feel "Almost There"?

Recently, AI Agent products have exploded. A variety of design and automated task processing (e.g., Manus, Coze Space, etc.) products are emerging. The way we use AI is no longer just a simple conversation; AI is now able to accept tasks, split them up, and call on different capabilities to execute them, ultimately delivering results directly.

However, after experiencing all kinds of Agent products, I always feel that it is still a little short of the mark. For example, you have given the outline, as well as clear notes, but the output of the AI will always deviate from what you envisioned in your mind or the results of your own hands. That feeling of "missing a little flavor" may be the missing part of "you" - your individuality, your unique style, your ineffable experience and intuition.

These "characteristics" seem difficult to describe or summarize in words. If we could, we would have put them in our prompts long ago. But the reality is that there are many subtleties that cannot be fully summarized in literal requirements or rules. This, perhaps, is a central bottleneck facing AI development at this stage.

AI seems to have "can do things", but from the real "understand things", there is still a critical step. This step is precisely full of human nature and the subtleties of human tacit knowledge. Even if the results given by AI seem to be usable, in fact, there may be a kind of "temperament cut". We often say that "AI flavor is too heavy", or feel that the generated content is "good, but not what I want", which is exactly the source of this. That's why we wanted to dive deeper into this article.

The Evolution of Know-How

Let's go back to the original discussion of Know-How. Back in 2023, when we were all excited about prompt prompts, we realized a key issue: a good prompt often requires the incorporation of domain-specific expert Know-How.

• For example, to generate effective Little Red Book copy, the prompt word needs to include an understanding of the characteristics of the Little Red Book platform, such as the suggested use of emoji, more colloquial expressions.
• The prompt words of the design category often emphasize the need to first determine the style and think about the use scenarios of the design, rather than directly requesting a diagram right off the bat.
• The same goes for retrieval tasks, where you may need to explicitly tell the AI that you want to retrieve English-language websites, or that, for travel information, sources such as Xiaohongshu may be of higher quality.

All of these are essentially Know-Hows, and when these Know-Hows can be effectively textualized and placed into the cue words, the AI output (under the same model) will be much better indeed**. **

We can roughly define these "Know-Hows" as the so-called "Workflow", i.e., the methods, actions or steps to accomplish a certain task. In the past stages of writing prompts, these "Know-How" can be further divided into:

1. Background Knowledge: Define the role of the AI, e.g., tell it what industry it is an expert in.
2. Analysis & Understanding: guides the AI on how to analyze and understand tasks.
3. Creation & Refinement: guides the AI on how to better complete and optimize tasks.

And now, with the advent of Agents, models are getting smarter. Many AI products (e.g., Manus) can even automatically disassemble tasks and generate Markdown-like To-Do lists , which is essentially the process by which the AI tries to understand and build "Know-How" based on your goals and background. As we have seen, the result is much better than simply making a general request in the past, as the AI can break down the steps, execute them separately, and even add a summary and check node at the end. The "thinking mode" or "reasoning mode" of models such as DeepSeek shows a similar ability.

The unspeakable "craft" and "craftsmanship"

While AI has made significant progress in understanding and performing Know-How, we still find that something is missing. Is this missing piece simply better rules?

I don't think so. I don't think so. Because if it's a rule, we can always find a way to write it.

At this point, we might as well think of the Japanese "craftsmanship" or what we often call "craftsmanship" or "skillful work". These are things that can only be realized, but not conveyed in words. They are hidden in the "deepest part of kung fu", and may even be called "XX Immortals" to describe those who have reached the state of perfection in their professional fields, and whose skills cannot be described in words alone.

These may be the key elements that still exist beyond the obvious "Know-How" we talked about before. Especially in some highly empirical and practical fields of specialization, such as painting and cooking, "Know-How" is far more than a simple list of rules or logic. It is more like an implicit, tacit control of processes, judgments, or values. It is often ambiguous, highly context-dependent, and often relies on intuition, empathy, aesthetics, and an ineffable sense of detail.

This part seems to be an insurmountable gap for current AI. Because we can't put all these "feelings" into cues, AI inherently relies on clear inputs to complete its outputs, and deeper information that is poorly processed, linguistically unspecified, or semi-explicitly implied may be harder for AI to capture and understand.

This highly specialized or implicit "Know-How" often relies on subtle value trade-offs, boundary judgments, and sensitivity to situational characteristics that are difficult to describe in a structured way. For example, when a master teaches his apprentice to cook vegetables, what he teaches is not only the details that can be recorded, such as "how much salt to put in and how many minutes to cook", but also the perception and judgment of the changes in the fire, the smell, and the state of the ingredients (e.g., the color). This is a kind of "Tacit Knowledge" that cannot be simply regularized.

Is Reinforcement Learning the Answer?

So, is there a technical solution to this ineffable "Tacit Know-How", or "Taste", or "Sense"? For example, is Reinforcement Learning (RL) possible?

As we know, after AI enters the second half, reinforcement learning is considered an important way to improve the ability of models. The learning process of human beings is also in constant interaction with the environment. As we mentioned earlier, although there are no written rules, the master will guide the apprentice through constant feedback (incentives) - good work will be recognized, and bad work will be corrected.

So, the missing part of our discussion, whether we call it "implicit Know-How", or "taste", or those "unspeakable things", is that AI has the potential to learn through reinforcement learning, but it is also that AI has the potential to learn through reinforcement learning. Is it possible for an AI to acquire it through reinforcement learning?

Frankly, I don't have a definitive answer, and I don't even know how to do it. The central difficulty here is this:

1. What exactly is the Environment? How do you build an environment that allows AI to effectively learn these subtleties?
2. How do we define incentives (rewards)? If we don't even have rules, how do we define "good" and "bad" incentives? The incentive itself may need to be provided by a master who already "knows the ropes", or a model who already has the ability to do so.

This leads to the dilemma of "the chicken lays the egg, the egg lays the chicken". There is no "master", it seems that the training of the "apprentice"; but if there is already a "master", then we still need the "apprentice"! What do we need this "apprentice" to do?

How to be able to translate human body nuances, tastes, and feelings, really migrate to the AI, I think it is a very ambitious, but also may be an extremely difficult subject. But if it is really done, then we may not be far from the so-called AGI (generalized artificial intelligence), and that will be a very bright future indeed.

What is Irreplaceable?

After talking about technological possibilities, let's go back to a more realistic question: when AI can accomplish more and more things, what is left for us humans? As someone discussed earlier, if "intellectual equalization" really comes, what can humans still do?

The answer may be, as we analyzed earlier, that what is left may be the "hidden know-how" that cannot be easily summarized and transferred, and the unique "experience" in all things and processes.

AI may be able to deliver results efficiently, but the value of the process itself, as well as your feelings and insights during the process, may become more and more important. The result, on the contrary, may just be an accessory. Our goal may no longer be just to pursue the result, but to shift to the experience in the middle, as well as your "taste" and unique judgment in the process.

For example, AI can give you a general sense of "good" results, a good based on the average value judgment of the public. But this "good" may not be what you think is "good" . Your personalized "good" may not be covered by the AI's output. That is to say, the "goal" it ultimately realizes does not fully include "you".

Then, the final result we hope to achieve may be "you" become an integral part of your desired result. This may be the space left for us in the age of AI.

What AI brings is more of an equalization of the so-called "explicit knowledge". And those difficult to copy, difficult to name the expression of aesthetic judgment, feel, taste and other abilities, on the contrary, will become a real premium scarcity, become each of us a unique existence.

In other words, what AI replaces is more like the part of us that has been "alienated" since the Industrial Revolution - the part that turns people into tools and machines . What is left behind is the part of us that is the most authentic, the least alienated, the least standardized part of us as human beings. It is those things that we once considered "minor" and less important that may become more and more valuable in the future. Those standardized things may not be the most important.

What is our relationship with AI?

This leads to the thought that the relationship between AI and us may not be a simple "replacement", but rather it can help us to realize a "better self" process.

As Marx revealed in Capital, to a certain extent, people have been alienated into tools, and our goals and pursuits have been highly standardized: study well, find a good job, buy a house and a car, and even consumerism has also coerced us to travel and go to the food card ...... It seems that everything has been put into a standardized track.

When AI can better complete that standardized delivery, those can not be replaced by the content, is precisely that part of our body "has not been alienated". Based on this "not alienated" part, combined with the powerful ability of AI, we can become a "better us".

Therefore, it is more important for us to cultivate those abilities that cannot be easily quantified:

• Advanced judgment
• Unique personal taste
• Stylized self-expression
• Deep empathy
• Extreme physical and artistic experience
• Ability to adapt to gray areas and ambiguities (often from a flash of insight, not from pre-determined rules).

Our relationship with AI is not a passive "lying down" and enjoying the benefits. This symbiotic relationship requires us to actively reposition ourselves. If we just passively wait, we may only continue to be "alienated" and fail to recognize and value the valuable qualities that are truly ours and have not been alienated.

The stronger the AI, the more dangerous our human inertia and inertia will become. In the past, we may have relied on good memory, fast retrieval speed, and familiarity with routines to gain certain advantages and dividends. But now, almost all of these have been "packed away" by AI.

This means that you must stand taller, see farther, need stronger generalization ability, more originality, as well as irreplaceable expertise in an extreme niche. As many people are mentioning now, "the biggest lever in the age of AI is to be yourself. " But the problem is that many people may not know what they really need and how they want to be themselves. If they can't find this position, they are likely to be marginalized.

AI is like a "camera of truth", which will strip away all the parts of a person that can be "outsourced", leaving only the true personal color and those with transcendent orientation. Therefore, we need to take the initiative to define ourselves, and to deeply cultivate in the value areas that AI cannot reach for the time being. Sublimate and transform your experiences and emotions into higher levels of empathy and creativity. Or, simply change the track, take AI as a powerful "physical" assistance, and bravely explore those non-standard answers, the richest "human flavor" of the uncharted territory.

In the age of AI, the vast majority of people may face two core problems:

1. The first problem is that they don't know what they really want. We often say that "being yourself" is the greatest leverage, but unfortunately, many people do not have a clear self-awareness, it is difficult to really "be yourself".
2. The second problem is: there is no "problem". We know that the development of the AI era can not be separated from human thinking and questioning. Raising valuable questions is even more important than solving problems. However, the reality is that for the vast majority of people, they may not have a "problem" that really bothers them and drives them to explore.

This is perhaps the most fundamental challenge we, and most of us, face in this ever-changing AI era. How we find ourselves and how we ask questions will be the key to whether we can dance with AI and ultimately become our "better selves".