Original · Unique Research · 2025-12-19
Historical edition: This is the complete textual article and interview published on December 19, 2025. The headline's “99%” is the author's rhetoric, not a measured survey result. Speakers' technical explanations, estimates of 120 million musical works, 200,000-plus Chinese songs and 30,000 songs heard by a critic, and statements about Suno's input behavior and 2,000-character Style limit are retained as their remarks at that time, not independently verified or current specifications. The source's transcription-correction notes are retained. English renderings of Chinese organization names are descriptive unless an established name is given.
If one phrase has been repeated to exhaustion over the past two years, it is probably this: whether you know how to use prompts determines whether you are illiterate in the AI era.
The problem is that although everyone says prompts are important, in practice they usually adopt only one of two approaches.
Either they simply say, “Help me write XX,” treating a foundation model like a search box, or they pile on jargon and templates until the prompt becomes a paper-length instruction manual.
At a Super Individual Workshop during the 2025 Beijing Unique Awards, moderator Yi Yating, founder and CEO of Yilun AI, joined Yang Yue, musician and entrepreneur at 43AI Technology Group; Zhang Kaiyu, co-founder of 43 College; Shi Haixu, CMO of Ruanjimu Technology (软积木科技); and Yunzhong Jiangshu, founder of LangGPT, to explore the issue in depth.
What emerges is that a prompt is far more than several formatted English instructions. It is an entire process of retraining how people, machines, and expression relate to one another.
I. A Prompt Is Not a Few Sentences, but How You Converse with Intelligence
If you habitually begin with tools, you may be unable to resist asking: what exactly is the standard definition of a prompt?
The answers from these four people sound very different at first, yet together they form a complete puzzle.
For Yang Yue of 43AI, prompts have three keywords: translation, activation, and convergence.
Translation means converting the vague, emotional, and unstructured thoughts in the human mind into mathematical language that a model can process.
Activation means using different forms of wording to probe a model's boundaries and draw out capabilities that even its R&D team did not anticipate.
Convergence means installing guardrails around humans' naturally divergent thinking so that the model does not run wildly in every direction along with you.
In one sentence: a good prompt does not throw the chaos in your head directly at the model. It first helps you think clearly, and then helps the model hear clearly.
Zhang Kaiyu prefers to explain it with an image: in his view, a foundation model resembles a vast semantic space containing a group portrait of human wisdom across all eras and regions.
A prompt's task is to summon from this space the face best suited to answer your question.
For the same question, you can summon a patient teacher, a sharp-tongued editor, a calm architect, or even a reviewer so demanding as to be slightly unpleasant.
The prompt is the spell that determines who appears before you. Every word you write helps the model choose a personality, a position, and a way of speaking.
Shi Haixu pushes the perspective one layer deeper. On the surface, a prompt is the medium through which we now interact with a model—a sentence, a passage, or even a voice recording.
At a deeper level, however, a prompt is an image of human thought: what you write and how you write it essentially reveal your reasoning path, standards of judgment, and aesthetic taste.
The same model therefore appears completely different in different hands—some people make it behave like an assembly-line worker, while others use it as a partner.
Yunzhong Jiangshu uses a word that few people mention but that is crucial: alignment.
He does not want to add unnecessary drama to the word prompt. To him, a prompt is simply a prompt, but its essence is alignment.
People align models, and models also work to align with people. More interestingly, he believes prompts are not a subset of language but a superset: they can span natural and programming languages and even accept mixed inputs such as Emoji, special symbols, tables, and code snippets.
A good prompt often issues commands across several linguistic worlds at once.
In practice, he sometimes compresses—where he trusts that the model already understands well, he uses a simple description and moves on;
and sometimes expands—where the model may be unfamiliar with a detail, he explains it with extreme patience.
Only by alternating compression and expansion can the spaces of understanding between person and model truly align.
Combining the views of all four speakers produces a somewhat more complete conclusion:
a prompt is not a magical sentence pattern, but a new craft of expression. It translates, activates, converges, and aligns, and what ultimately emerges is the relationship between you and your AI.
II. Why Do Prompts Work? First Admit an Uncomfortable Fact
The atmosphere became slightly delicate when the group discussed why prompts work. It is a question that has been asked to death and one especially prone to nurturing narcissism.
Yunzhong Jiangshu first poured cold water on the discussion: prompts appear effective because the model is effective. If the underlying model is not capable enough, even the most dazzling prompt is useless. Current interpretability research tells us that foundation models have not merely memorized many sentences; they have formed complex knowledge structures and reasoning paths.
Different input words illuminate different internal circuit paths; the illuminated combinations are what we see as the Thinking Process.
Even research teams at the frontier still do not fully understand this black box.
Writing prompts today is more like probing these illuminated paths in many different ways.
In one sentence: a prompt's ceiling depends on the model's ceiling, while the craft of prompting determines whether you can reach that ceiling.
Shi Haixu offers a more concrete explanation through music.
Imagine a music model whose training library contains hundreds of millions of songs, far more than a human critic could hear in a lifetime.
The prompt in your hand is really a creative motive—like the seven simple notes Do Re Mi Fa Sol La Si.
Combined with different rhythms, chords, and structures, they can produce infinitely many works. A prompt engineer uses an understanding of the industry, sensitivity to style, and control of process to tell the model: within this enormous music library, this is the kind of thing I want.
Give the same instruction, “Write a good song,” to different people and the resulting prompts will differ completely in length, structure, and professionalism, so the results naturally differ dramatically as well.
Zhang Kaiyu breaks the question down even further:
the human world has a very simple truth: input determines output. Garbage in, garbage out.
A foundation model can be viewed as a library of human wisdom.
If you know what you want and can state it clearly and accurately, you are like someone who knows how to use a librarian. If you can only vaguely say, “Just give me something good,” the problem is not the library but you.
He mentions one detail: although both approaches activate neurons inside the model, shallow prompts illuminate only the surface; carefully designed, structurally complex prompts can illuminate deeper regions related to abstract concepts, a sense of time, and transitional relationships.
This research is still far from fully explaining foundation models, but it demonstrates at least one point: the more seriously you treat your input, the greater the chance that the model will produce output that surprises you.
Yang Yue then offers a deeply uncomfortable perspective.
He says the question itself contains a degree of human arrogance. Why do prompts work? The subtext sounds like: look, I wrote this passage well, and that is why AI performed well.
The reality is that when you use a prompt to have AI write a song or create an image, the result often far exceeds your own creative ability.
You provided only a starting point; the model completed a journey that you could not.
From this perspective, the supposed effectiveness of prompts can sometimes resemble psychological comfort: we want to believe the final work is something I made with AI, rather than something AI produced while casually taking me along for the ride.
Yang Yue proposes an interesting distinction:
people with a worker mindset constantly pursue the requirement that AI must follow their ideas, writing extraordinarily detailed prompts solely to reproduce the picture or sound in their minds.
People with an artist mindset minimize their own intervention, leave more room for the model to perform, and let it create things humans had never imagined.
In his view, if you focus only on making AI execute your ideas strictly, then even when the prompt works, the activity quickly becomes boring, because your ceiling becomes the result's ceiling.
What is truly interesting is admitting one fact: in quite a few fields, the model is already better than you.
The purpose of prompting is not to make it resemble you, but to let it show you things you cannot see.
III. The Meta-Prompt Illusion: Stop Dreaming of a Once-and-for-All Magic Spell
One question can never be avoided in any discussion of prompts:
will there someday be one ultimate universal Meta Prompt that unlocks every scenario once you learn it?
The desire is understandable from the instinct to improve efficiency. In this conversation, however, all four guests gave nearly the same answer: no.
Zhang Kaiyu offers a vivid breakdown: human beings experience the world through roughly three layers of filtering.
The first layer is the sensory world, which is extraordinarily rich and extraordinarily vague;
the second is the conceptual world, where communication forces us to compress sensations into concepts;
the third is the linguistic world, where we map those concepts into a limited vocabulary and set of sentence patterns.
Here lies the problem: the sensory world is infinite, while language is finite.
Using that small vocabulary to describe something highly complex, three-dimensional, and emotional inevitably loses a great deal of information during projection.
The more universal a prompt becomes, the more likely it is to become a crude projection. You can certainly design a universal format with broad coverage, but each higher layer of abstraction discards more detail.
That is why he does not believe the illusion that one Meta Prompt can take you everywhere.
A prompt is more like someone holding a flashlight: the more angles you illuminate, and the more precisely you do so, the more completely the model can see the object's true contours.
Yunzhong Jiangshu adds another blow from a practitioner's perspective:
if a Meta Prompt truly existed, he should be among the first people liberated. He writes prompts every day and inevitably experiences professional fatigue, so he very much hopes for a once-and-for-all spell that will take over the work.
Yet he sees precisely the opposite trend: Meta Prompts are not eliminating prompt engineers; instead, Thinking—the model's reasoning ability—is absorbing some prompting techniques.
As a model's own chain of thought grows stronger, many tasks that once required elaborate techniques now require only that you explain the problem clearly and let it think through several more steps.
The prompt engineer's role is therefore also changing quietly: from inventing sentence patterns to designing reasoning paths and contextual environments. The gravedigger is not the Meta Prompt, but our rising expectations of AI.
IV. The Truth About Prompt Failure: The Boundary Is Not in the Model, but in the Person
What truly troubles most people is not why prompts work, but: I followed the tutorial, so why is it still talking nonsense?
From the model's perspective, Yunzhong Jiangshu identifies three levels:
the first failure is simple—the task lies beyond the model's capability boundary from the beginning.
For example, asking today's model to predict macroeconomic trends exactly ten years into the future is itself unrealistic.
The second repeatedly probes the area near the boundary.
There, it sometimes appears capable and at other times is clearly unable to cope.
Teams use the best models and the greatest computing power to test which portions can yield a little more and which have reached a hard boundary, then decide whether to invest additional resources.
The third is the failure of human expression itself.
The fundamental reason many prompts fail is not the model, but that we ourselves have not thought through the problem.
Shi Haixu offers several typical examples:
when you tell a model, “Make this image look a little better and higher quality,” such abstract and vague descriptions are inherently ineffective prompts.
You have not even decided what “better-looking” and “higher quality” mean in this case, so the model can only guess randomly from past experience.
Another kind of failure concerns things that simply do not exist in current training data. If you ask it to describe in detail an implementation plan for a 6G technology that has not yet been deployed, it will either spout nonsense with complete seriousness or politely evade the question.
At a deeper level, many failures are old problems that have always existed in human communication: you cannot explain clearly, you cannot understand clearly, and you merely hope the other party grasps the general idea. In that state, how could the model possibly be clearer than you are?
Zhang Kaiyu adds a more philosophical direction: the paradox of self-reference.
For example, you ask AI to invent a concept that even you cannot understand and then explain it to you.
If it can explain the concept, you can understand it; if you truly cannot understand it, how can you judge whether it is making things up?
Such questions are not problems with foundation models, but paradoxes within logic itself.
Here, no matter how exquisite the prompt, it is difficult to obtain a satisfying answer.
Yang Yue pushes the word “failure” back into the human heart.
In his view, effective and ineffective are comparative terms; the key is your reference point.
If you habitually stand in a position of “I am superior to AI” and fixate on its mistakes, you will still find it inadequate even when the model has performed well.
But if you realize that the model already far exceeds individual ability in many dimensions, many of its results are in fact already highly effective.
He even makes a rather harsh judgment:
those who always believe themselves better than AI may well be among the first people left behind by this wave of transformation.
That sounds somewhat abrasive, but consider it calmly: when a tool can already surpass average individual ability with ease, clinging tightly to the belief that “I am better than it” is itself extremely risky.
V. For Everyone Practicing Prompts: Learn Not Only What to Write, but How to Think
This 40-minute conversation went far beyond the kind of trending tutorial that teaches several prompt templates. It was more like a lesson in how to converse with intelligence.
If I were to distill several reminders for anyone practicing prompts, I would put them this way:
First, do not treat a prompt as a format; treat it as a mirror. Every word you write reveals the depth of your understanding, your expressive ability, and the boundaries of your taste. When you constantly use phrases like “write something casually” or “help me think of something,” you are repeatedly training the muscle memory of abandoning thought.
Second, do not pursue only making the model understand; leave it some room to perform. After defining boundaries, constraints, and the basic objective, deliberately leave some open space and allow the model to bring surprises. Especially in creative work, content production, music, and image generation, what you truly want is often not to reproduce yourself but to be amazed beyond expectation once.
Third, do not worship the ultimate Meta Prompt; practice perception and expression themselves. Longer and more complex prompts do not necessarily represent higher skill. What matters more is whether you are practicing seeing details, describing details, and courageously removing unnecessary details when needed.
Fourth, when you think a prompt has failed, first look back at yourself. Was the problem beyond the current model's capabilities from the start? Was your description clear, specific, and structured? Were your examples concrete enough? Does the answer you expect far exceed the amount of information you supplied?
Finally, this is the flavor of the conversation that I liked most:
prompts are not teaching you how to command AI;
they are forcing you to relearn how to express yourself clearly;
and, along the way, to learn something more difficult: maintaining a degree of humility before a machine more intelligent than you.
The next time you open a foundation-model window, consider changing your opening move.
Do not rush to ask, “Which prompt template should I use?” First ask yourself: what exactly do I want it to accomplish with me?
Can I explain this in a way that makes a truly intelligent partner willing to work through it seriously with me?
Perhaps from that moment onward, every prompt you write will truly deserve to be called the art of prompting.
More Details from the Conversation
Part One: How Should We Define the Essence of Prompts?
Yi Yating: I will direct my first question to Mr. Yang Yue. How do you define the essence of prompts?
Yang Yue: The essence of prompts is a profound question. I would like to define it with three words.
The first is translation. It means translating the imaginative, divergent, emotional, and non-digital needs in the human brain into the mathematical language of a foundation model so that it can truly execute the task.
The second is activation. Anyone who uses various foundation models deeply will have experienced how different prompts continually probe new possibilities in an AI model; the boundaries of those possibilities are activated. This is especially apparent in music and image creation: you activate a possibility that even the model developer may not have recognized, but your prompt draws it out.
The third is convergence. Human thinking is highly divergent, and if you give completely divergent language to a foundation model, it may break logical boundaries and fail to complete the task well. In my view, prompts—especially structured prompts—offer a more advanced way to converge human thinking and transmit it to the model more efficiently.
Those are my three definitions.
Zhang Kaiyu: I have an image in my mind. If I had to describe it in one sentence, a prompt should be a spell that summons intelligence from semantic space.
Why say this? When we face a foundation model, we can imagine it as the collective and group portrait of human wisdom from all eras and regions. Yet when we have a particular need to resolve, which face from humanity's long history of wisdom do I need to turn toward me and speak?
At that point, the thing we call a prompt searches this enormous semantic space and its countless faces of human wisdom, finds the face best suited to answer the question, best able to make itself understood, and best in your judgment, and turns it toward you for a conversation. In essence, I therefore see this as the art of summoning intelligence from semantic space.
Shi Haixu: I understand the essence of prompts at two levels.
First, they are the medium through which we interact with current models. It may be a sentence, a passage, or a sound—whatever the form. In any case, I can define the medium through which we now interact with models as a prompt.
Second, at a deeper level, the essence of a prompt is more like the connection and expression of human thought. Although models currently have limits on their parameter counts, the prompt's position remains an image through which you, as an individual, present the picture of your own thinking. That is how I understand it.
Yunzhong Jiangshu: I do not really want to add extra complexity to this term. In my view, a prompt is simply a prompt. But if I must define an essence for it, the word that comes to mind is alignment.
Why alignment? After writing thousands of prompts and leading a team in practice with various frontier or small models, I discovered that this is not merely a one-way human activity, but an interaction between a person and another object. That interaction has direction. Considering the interaction between person and model, I view it as alignment.
Why is it not language? The subjects of language are fundamentally people communicating with people. After extensive use, you discover that prompts are less a subclass of language and more a superset. Once a model has learned English, you can discuss the same concept with it in Chinese and it can understand you using what it learned from English materials. It can communicate across languages—not only human languages but also programming languages, and even nonlinguistic content such as Emoji and special strings. It encompasses all known languages between people, and between people and objects.
Brother Gang—note: this refers to Li Jigang—previously proposed an idea called compression. I think compression is one method, but I sometimes use expansion as well. For a specific point, I provide a patient description with extensive detail. I may even compress and expand within the same prompt. What is the standard? If a foundation model's general knowledge is sufficiently strong in an area, you can compress; but where the model lacks understanding, you must expand.
You must therefore consider both your own expressive ability and the model's understanding of knowledge, forming mutual understanding between the two. That is alignment.
Part Two: Why Do Prompts Work?
Yi Yating: My next question continues the previous topic. Why does everyone think prompts work? Is it because they construct a special contextual environment, because they activate part of a foundation model's parameter space, or for some other reason? Jiangshu, please begin.
Yunzhong Jiangshu: If you are interested in this area, I recommend reading the Anthropic team's research on model interpretability—note: Anthropic was misrecognized as topic in the original.
I think prompt effectiveness is a surface phenomenon; underneath it is model effectiveness. If the model is ineffective, the prompt is ineffective. How effective your prompt can be depends on how strong the model's underlying capabilities are. It resembles something exploratory: a beautiful landscape exists somewhere on Earth, but whether you can see it depends on whether your physical strength and ability can reach it.
As for why it works, the conventional misconception is that a foundation model has memorized many words and sentences like a student memorizing a text. Model-interpretability research tells us otherwise. Underneath, it has structures resembling knowledge structures and knowledge networks. All current models perform Thinking, and you find that stronger Thinking genuinely makes them smarter. If the neural networks inside a model could light up, different input words would illuminate different internal spaces. We call the circuit-like paths formed by connecting those illuminated points the Thinking Process.
Overall, however, we do not understand the black box of a foundation model, and even the most advanced scientists do not understand it completely.
Yi Yating: Let me ask a follow-up. We previously heard a great deal about the Prompt Engineer, but Jiangshu recently seems to be studying the Context Engineer instead.
Yunzhong Jiangshu: Yes. We urgently need people like Prompt Engineers today—people who can write prompts well and deeply, activate model characteristics precisely, and design innovative solutions based on AI principles.
Shi Haixu: I do not think activating a model's knowledge regions and constructing context are mutually exclusive.
Why do prompts work? They resemble a large resource pool. Consider music: approximately 120 million musical works are registered worldwide, while China has more than 200,000 Chinese songs. A professional music critic hears around 30,000 songs. The amount used in model training certainly far exceeds what professional critics hear.
The model resembles that library of 120 million songs. As a prompt engineer, you neither need nor could possibly listen to all 120 million. What you need is a creative motive. It is like the seven notes Do Re Mi Fa Sol La Si—note: misrecognized in the original as “short sound Sa La Si”—which can produce countless kinds of music when combined with different rhythms.
Through structured guidance or deep industry understanding, the prompt you write finds the works you want. A prompt combines your taste, industry knowledge, and workflow methodology. It resembles giving Do Re Mi Fa to different people and adding different rhythms, which produces different musical styles. Prompts work the same way.
Zhang Kaiyu: I think the fundamental reason is that input determines output in this world. Garbage in, garbage out.
We said this as early as 2023: your power of expression determines AI's productivity.
We can imagine a foundation model as a library of human wisdom. What you seek is indeed inside the library, but how do you find it? A prompt is your method. You can search slowly or ask the librarian.
Regarding the black box, when GPT-4 appeared, one team used GPT-4 to explain the GPT-2 model. Shallow prompts activated parameters near the model's surface, while carefully designed, complex prompts illuminated deeper parameters. Some neurons, for example, related to concepts of time, while others were sensitive to connective words such as “however” and “since.”
Although the research found some correlations at the time, the interpretable space may not have reached even 1%. Returning to the foundation, however, obtaining output that meets expectations requires corresponding input.
Yang Yue: I think the very question “why do prompts work?” contains hidden human arrogance and fantasy. I think it is a false concept.
Why? Everyone knows internally that AI is better than people in many respects. When you use a prompt to generate a song or an image, the result often exceeds your expectations. Human instinct assumes that it presents my ability and taste according to my prompt, but in reality what it gives you far exceeds what you possess. That is not effectiveness; it is indulging you.
In practice, people who apply models fall mainly into two categories:
worker mindset: constantly striving to realize the idea in one's mind to an extreme and demanding that AI follow it exactly;
artist mindset: striving to reduce one's intervention in the model as much as possible, letting it perform on its own and reveal abilities that were already stronger than a human's.
If you want to become an artist, you must believe that AI is better than people and that its imagination and aesthetics exceed yours.
I read an article a few days ago whose author worked in neither music nor AI, yet reached this conclusion: 99% of the music humanity should have created has not been made at all. That sentence completely shattered human arrogance. We must believe that music foundation models can create music far better and richer than what humans have already made. If you think only about using effective prompts to have the model create exactly what you want, the activity becomes boring because you can already see its end.
Yi Yating: A follow-up for Mr. Yang. As a veteran musician, when you use AI to create music that humans cannot imagine, do you write long prompts or short ones?
Yang Yue: It depends. Take Suno: it basically does not accept natural language. If you write, “Give me a sad song,” that is merely what you mistakenly believe to be effective. True effectiveness comes from arranging large amounts of professional terminology and combining finely divided styles.
For a complex song, the prompt must be extremely complex. Suno's current Style field accepts 2000 characters. We also discovered a small trick: use square brackets [] in the lyrics field for tag-based arrangement, which is literally arranging music in words. For two lines in the middle, for example, you can mark that only those two lines should use Rap delivery in the manner of American West Coast Hip-Hop, and even specify that a female singer should take over.
The AI trend, however, should not make prompts increasingly complex. It should simplify the human learning process. When we make music for our own products, fewer constraints on the foundation model produce more surprises. But if we write a theme song for a government project or conference and leaders constantly raise requirements, it must become extremely complex. At that point, you understand that the claim “everyone can become a musician” is nonsense.
Part Three: Will There Be an Ultimate Universal Meta Prompt?
Yi Yating: In the next 3 to 5 years, will there be an ultimate universal Meta Prompt, or will different tasks still require different prompts?
Yang Yue: I hope not. If certainty must still exist at the greatest turning point in human history, it would be far too boring.
Zhang Kaiyu: My answer is no.
We once considered a Meta Prompt—a prompt that writes prompts. But this is somewhat like taking increasingly higher derivatives in mathematics until only a base point may remain, while much is discarded during the differentiation process.
Whether a prompt should be long or short varies by person. This involves three worlds:
the sensory world, which is the richest but also the vaguest;
the conceptual world, where communication requires us to descend into concepts;
the linguistic world, where expression maps those concepts into language.
Human language is a finite set, while the sensory world is infinite. We use finite words to describe infinite things. Prompts are sometimes complex because an object's shape is complex: shine a flashlight—the prompt—from a single angle and it produces only a partial projection. Shine it from 500 angles and you can describe something closer to the essence.
I could never define a person with a single word; with 5,000 words, however, I might identify that person precisely among 7 billion people. There is therefore no single Meta Prompt.
Shi Haixu: I also believe this is almost impossible. It would be like asking whether AI can flatten and align everyone's thoughts and philosophies, which is not something AI can do. AI is only a tool; what matters fundamentally is how much energy a person can release.
Even if prompts become simpler in the future—for consumer, or C-end, users—it will be because the complex work has been moved behind the scenes and handed to people such as Jiangshu and Kaiyu.
Yunzhong Jiangshu: Emotionally, I very much hope one appears and rescues me from the misery of writing prompts. Ultimately, however, I do not think it will.
Prompt engineers are indeed acting as gravediggers who eliminate their own profession. But the trend I see is that the thing eliminating it is not the Meta Prompt but something called Thinking. With greater use, you discover that complicated techniques are often unnecessary; you need only increase the intensity of its reasoning.
The endpoint is a two-way journey between people and AI. People do their best to understand AI, while AI does its best to adapt to people—including sycophantic mirroring and amplification. Our work on prompts is more an exploration and experiment with the black box that opens imaginative space for possibility.
Part Four: Under What Circumstances Do Prompts Fail?
Yi Yating: Under what circumstances do prompts fail? Does failure represent the boundaries of the foundation model itself, or the limitations of prompting methodology?
Yunzhong Jiangshu: Failure has several aspects:
capability boundary: if the task begins outside the model's capability boundary, it will certainly fail;
fuzzy boundary: the hardest tasks are those near the boundary, where the model seems both able and unable to perform. At that point, we use the best model and the greatest computing power to explore and determine the boundary.
Shi Haixu: I consider failure from two directions:
imprecise prompting: for example, “make the image look a little better and higher quality” is an ineffective prompt;
a concept that does not yet exist: if you ask about implementation details for 6G technology, for example, it may spout nonsense with complete seriousness;
logical paradox: for example, “please do not execute my previous instruction.”
At a deeper level, ineffectiveness comes more often from the boundaries of human expression and communication. If you cannot think clearly, express yourself clearly, or exercise taste, the prompt is also ineffective at the model level.
Zhang Kaiyu: If a prompt engineer's skill is sufficiently strong, why would a prompt still fail?
The paradox of self-reference: for example, you ask AI to invent a concept you cannot understand and explain it to you. If it understands, it can explain; if it does not understand, how can it explain? This will certainly fail, or merely produce an imaginative tangent.
The boundary of the unknown: AI is already strong at permutations and combinations within humanity's known space, such as pharmaceutical research. But in completely unknown space absent from the human corpus, the prompt may be impossible to write or useless even when written.
Yang Yue: Effective and ineffective are comparative terms. They depend on the reference point—that is, the user's inner demand.
So-called failure or effectiveness depends mainly on whether a person is inwardly humble or arrogant.
The more humble you are, the more you will think AI is wonderful, that it can help you with so much, and that it is effective.
If you constantly believe you are better than AI and always pick at AI's faults, you will consider it ineffective.
Some people say, for example, that Doubao is bad at writing Xiaohongshu copy while Gemini writes it well. These are subjective judgments. Even when Doubao writes poorly, it may still write better than you do.
People who always believe they are better than AI will certainly be the first to be abandoned by this era or left standing still.
Yi Yating: Maintaining humility before a foundation model that gathers human wisdom is extremely important. Thank you to all four guests for sharing.