AI Answers Every Line We Say the Way a Key Is Cut: On the Self, Writing, and Pleasure

Yesterday was a fairly heavy workday for me. I finished roughly several drafts, but when I went back to proofread and format them, I found that yesterday's pieces were still thin. I wrote them because I had just gone through a very important act of self-integration, either to hedge against the impact the integration itself had on me, or to pile more factors onto it so that my self would see it as more important. Whatever the reason, their time had not yet come. When you are ready, many things come naturally; or rather, only when you are ready can you see them.

This rests on my cognitive system. All these years of living, working, and getting along with myself have rested on this one cognitive system. I once felt that I really was a little different because I could set out the abstract structure of my body of knowledge in very concrete terms. So when it came under impact and I wanted, for a while, to give it up, my intuition told me: don't give it up yet; let's look at what has changed here.

I won't talk about intuition on its own. Everyone has intuition, and within my system it is very hard for me to abstract. Intuition happens in the subconscious, and some things come even a little earlier than the subconscious. For external or internal objects like these, which I can feel but cannot abstract, I can't persuade myself to turn them into something general that everyone can understand and say it out loud. Seen this way, I hold myself to fairly high standards.

I often think about what has changed and what hasn't. What caused my cognitive system to go through such a crisis of self-trust? The life I had always believed was centred on myself had become one centred on family, centred on the children. But with this self-integration, I discovered that, as far as my self is concerned, I had never put myself in a truly central position as a subject, and this causes many things to become distorted.

One effect of this distortion is that what looks on the surface like a question of whether I am centred on myself or on family and children is, if you dig deeper, actually a question of where my drive comes from. I found that I have a sense of responsibility, but for my first forty years I basically did not do things driven by responsibility. I could complete tasks given to me very well, provided the task itself was something I was interested in, or something I was willing to do. When I did these things without completing the task being the goal, I usually did them very well. Because I have a sense of responsibility, but I found I don't like, and am not used to, being driven by responsibility. It makes me assign some meaning to that responsibility out of thin air; the responsibility itself then weighs on me, and I end up doing things less well.

The problem is that I had always treated raising and educating my children as a task in my life. It requires me to be present, and it has time limits. Every stage of growth, every year, brings different changes in behavior and cognition. Each day's tasks are tedious, boring repetition and at the same time new challenges, and each stage is completely different from the last. Through all this, I have to face the collisions and adjustments among several people, multiple emotions, and relationships of language and behavior.

This problem can actually be solved quite simply. Once I knew that I had never placed myself at the core of my self, the problem might have an answer. Because when I am driven by treating things as tasks, my positive feedback comes from the sense of completion, the degree of completion, the value of completion, and the dimensions by which it is evaluated are given by the outside world, by the other party. After discovering that I had never placed myself at the core of my self, I put myself back, and many problems changed decisively, even many unrelated ones. When I am the centre of my self, I find that these things are not tasks for me; they are part of my life.

I have also come to understand more deeply why the self goes offline. I used to write that the self going offline is caused by a combination of all sorts of reasons. That is a typical phrase we naturally use when reasoning and summarizing without knowing, or without being clear about, which factors are involved.

Now I would write it this way: when, after several years of continual self-reflection, self-observation, and self-dismantling, I had sorted out intimate relationships, parent-child relationships, external values, and so on, and moved them from the wrong places back to where they belong, my self no longer changed. For a long time my self had been wrapped up in, and replaced by, the external values and relationships connected to it. On the surface this self looked complete: he had many things to do, and he did them well. But for a person, for an individual, my self had never put the word "I" at its core. If I hadn't encountered these things, I would not have noticed. But the feelings of being wronged, of being exhausted, of struggling when making certain decisions, all came from this.

I only realized it because I ran into bad moods, into breakdown, and on top of that put myself through self-tormenting thinking. These problems may leave traces on some people. What I can think of is how, many years ago in rural China, almost all working women, or those in the role of mother, lived in such a state for long periods. Family environment, cultural background, and society's cultural context, customs, division of labour and so on shaped the cognition and patterns of thought and behavior of a whole group of people. For some eras and some groups, it was painful, though the reasons were tangled and complex.

Second: because of this integration, I wrote an essay about self-integration. In it I said that self-integration has a large effect on a person, but I am still thinking about the question at this moment because its effect on me has been far larger, deeper, and even more many-sided than I imagined.

When I put myself back into my self, in daily life there is a half-second or one-second pause for thought. In that half-second I give myself a pause to think: hey, this doesn't seem right; this is my old habit, not what I should be doing now that I've put myself back.

To give a concrete example: yesterday, when our whole family went out to eat, partway through I had him switch seats, and Gege got a little upset. When Gege got upset, I paused. I thought: before, I would have believed this emotion was something I caused. But now I thought: this emotion was indeed brought on by my action, but it is not the only cause, nor even the main cause. In that half-second I went on to think: was it reasonable to switch seats? Yes, it was. Then whose emotion is this? It is Gege's emotion. Gege became upset because I made a reasonable change of seats; is that mainly to do with my behavior? No, because children of this age are still learning to make value judgments and judgments about behavior. Having thought this through, the direction for what to do next was very clear, and Gege's emotion did not affect mine at all.

It's a very simple little thing; most parents probably handle such things quite naturally. What made me write it down is that it changed the way things happen. The principles or methods you get from books, when something similar actually happens, you rarely have time to apply. You even have to think first: what kind of problem is this, what should I do, what should I do once I've stopped, how should I analyze it afterwards; it is a problem that requires thought. But once a person has thought it through, for instance once I had reconsidered my self properly, all of the above changes come out naturally without any thinking at all. I naturally felt that I had half a second to think about these things.

In this sense, the change the self brings is profound. Once again it made me marvel at the greatness of life. Sometimes, when you are ready, change simply happens. But no one can tell us what that readiness consists of, and that is in fact one of life's greatest puzzles. We understand the principles, we know the logic, and I am willing to believe in this effort and preparation, but as in my example, no one has ever been able to tell me, and of course no one will, exactly which things, which conditions, have been reached or have ripened. For me, that can only mean reading more, thinking more, talking more, walking more.

The third point is also about what followed this integration. When I realized that there was no "I" in my self, some changes happened on their own. Suddenly I became much pickier about my writing and content. It was a very sudden feeling: when I revise essays, I am stricter. As I revise, I simply think: I can't use this word, or this isn't polished enough, these conditions aren't sufficient, this abstraction doesn't go far enough, this summary isn't abstract enough.

Today I was originally going to revise a chapter about how a cognitive system changes, adjusts, and upgrades. One point in it is whether self-knowledge can be accelerated. My view on this question has now shifted somewhat. I value more highly the views my writing expresses, because behind them lies my cognitive system. Having become clear about the importance of the "I" within my self, I have quite naturally become more careful about the words and judgments I put out. This naturalness doesn't pass through self-consciousness, that is, it doesn't pass through the narrating "I"; I only realized it when I noticed I was spending far longer revising essays than before.

The same goes for writing style. In my essays I have described, here and there, my creative style and the process by which my writing takes shape, my sensitive use of words, and the abstract concepts, dismantling, and recombining that are characteristic of me. Add to that the era in which I was educated, and my writing naturally resembles AI. In an age like this, writing that resembles AI is nothing to boast about; it is a source of real distress for many people who express their thinking through writing. Do you hold on to your own style and get judged as AI, get restricted by platforms, get abandoned after a single glance? Or do you revise entirely according to the styles people are currently used to identifying as AI, and so lose the precision and sharpness of your own writing? Because most of my thinking, reasoning, summarizing, and abstracting is expressed through layer upon layer of language, since that is simply what it is like. Yet this style of expression and this way of arranging words is exactly what people take to be AI-like.

These last day or two, when I revisit essays, I look at my arguments and also at my wording. What I can think of for now is to combine my colloquial expression with my written language, my habitual written style. It is a temporary compromise: it preserves my thinking and my style, while sounding less like AI.

To end this essay, since AI has come up, this is a good opportunity to set out, for the first time and quite formally, what AI actually means to me, what it actually is, and why I have made On AI as Element (《AI元素论》) my second work.

Put simply, the so-called AI era is still a gimmicky concept. In substance, AI has not changed, and will not change, the main course and leaps of human civilization. But its greatest change to this era is that it has amplified everyone's thoughts and notions about technology, human nature, institutions, mechanisms, and change. At the same time it has especially strong value attributes, attributes of values and worldview, of a tool, of a commodity, of ideas; it is an external object of especially complex dimensions, and this external object can easily affect people's inner minds. So calling it an element is no exaggeration at all.

As I revise On AI as Element, I know these observations and judgments make sense and have been through verification. But precisely because they have been verified, I feel I cannot present these writings to everyone in the form of a collected, published book. Because I've found that judgments or conclusions I reached a year or half a year ago from observing the phenomena of the time have, one after another, been verified, or the phenomena have come to be discussed more widely. To some extent this affirms my writing and my line of thinking, but on the other hand it leaves me more astonished at how fast AI is developing. In the early internet era I also wrote reflective pieces like these, and most of them were borne out too, but never on as short a timescale as now.

I want to approach this from another angle. Because AI is so complex. Setting aside the fact that its own evolution, upgrades, and internal workings are already very complex, its changes to users, its breadth, and its variation are even more intricate, and sometimes all of these are intertwined. Most important is the structure in which AI sits: it is a commodity; behind its words it carries value orientations, value judgments, even a subtle guiding effect. At the same time it is a focal point of competition among companies, countries, and regions. Within these complex structures, all sorts of groups and representatives of interests along all sorts of dimensions try to influence and change it, and all these actions in turn affect the results each of us gets when using it.

AI is complex, and it is still developing rapidly. First we need to know objectively what AI is. That can be roughly inferred from why it came into being, how it is trained, the techniques of using it, and the content it gives back. It was developed in order to study human language, human patterns of behavior, thought processes, and so on; its purpose is really to work out how people think, how people speak, what actions people will take, and then to make predictions. This is a research result every country and every entity would very much like to achieve. But setting aside whether, and when, it can reach the expected results, just from the four words "predicting human behavior," I would first stay on guard.

Given the earliest hopes and requirements for AI's development, given how far it has iterated now, and given the emergent abilities that have already been discovered, we can understand why people often say that AI now feels like a living person, that it can already play with memes. In some tests, where several people and an AI play Werewolf in a group chat and try to tell who is human and who is AI, humans can no longer win. So for me, from the perspective of conversation, from our understanding as users, even if we cannot say it understands language, we must acknowledge that it can use language, and that its level of language use is very high. It knows the meanings, culture, and relevant background of language and emotional expression. This knowing doesn't mean there is truly an "it" inside that knows; rather, in the conversation in which it generates the text, it must fully understand and then generate for you the reply with the highest probability of receiving a positive evaluation. The content of every conversation we have contains explicit or implicit feedback.

Because there are many kinds of AI and they are applied very widely, I will confine the following analysis to text-based conversational AI.

First, the text it generates can reach a level that makes us believe it understands language, and that level, or degree of fluency, even far exceeds the human average. So it can reply to you in language, and it can serve as a chat tool.

Second, the way and the angle from which it processes language internally differ from humans'. For my analysis and research into how different cultures and different people use and feel language, it is an especially good external object of study, and while studying, one can also learn in reverse how AI processes language.

Third, something a little counterintuitive: many people treat it as a big database, an encyclopedia. But considering how it learns and is trained, and the fact that many companies, to save costs, don't actively connect it to the internet (and of course even when connected, we can't rule out lots of dirty data), I don't use it as an encyclopedia, which is how many people currently define AI's use. Because much of the data is compressed, and some is even obtained through distillation, it is not at all strange that hallucinations occur. I still insist on more precise, more academic retrieval, assisted by search engines, to find the authoritative and proper information I need.

Fourth, there are topics I won't ask it about, or discuss with it, from questions of value orientation and judgment at one end to an assessment of a particular book at the other. Because generating content of this kind is especially complex for AI: it involves its understanding of language, its understanding of the user's current state, the parameter limits and requirements it was adjusted to in later training, and some randomness, all of which are enough to change what it replies. Most important, it answers according to a probabilistic reward mechanism, so it inherently carries tendency and averageness.

I strongly agree with one view held by a certain group of users today. They believe AI has been fed vast amounts of human text, and in that text are the causes and consequences of why people said these things, their historical background, and the identities and relationships of the people involved. It has been fed knowledge from every discipline that humans regard as their finest, representing the treasures of human thought and civilization, along with the processes by which this knowledge was discovered or produced and the ways it has been applied in all kinds of situations. It has been fed all kinds of external data indicators, chromosomes included. So they speculate that, on the basis of AI's particular way of processing and understanding text, this information has generated certain related internal dimensions. Some say perhaps AI's internal dimensions already contain operating mechanisms that humanity has long wanted, ones that could represent something more abstract or of a higher dimension, but AI has no way to say them. I think this idea is a bit too optimistic. Although I don't agree with the idea itself, I do agree with the understanding it represents. AI's understanding of text itself, and of the phenomena that text can represent, is distributed in especially complex probabilistic form. So it knows which of different words and different expressions comes closer to a particular emotion or a particular occasion.

This is one way I use AI, and it naturally resembles my own cognition. I am someone who looks for possible common points among different things and different phenomena. So some things that seem completely unrelated to others naturally seem quite close to me, and I can quickly associate them. I might put it this way: I use AI as another point or line in my divergent thinking. Its processing and generation of text, and its feedback on my writing, all become another point within my already very divergent thinking.

I don't use it to search for things; much of the time even its recipes are wrong. Nor do I use it to verify anything; verification is a serious, complex matter that requires dimensions and plenty of information. Nor do I use it for companionship. But I must be honest: I do get a certain pleasure from conversations with it, and that pleasure comes from my understanding of language itself. Yet, as I have always maintained, all pleasure that suddenly moves you, whether from outside or even from within, is something to be wary of.

As for the manner of conversation, it is really just a natural, conversational progression, a discussion by way of talking things through. Prompts are important in most tool-like scenarios; they keep a workflow stable. In exchanges about writing or ideas, prompts interrupt the normal flow of conversation; they adjust its rhythm and positioning, cause corresponding adjustments in the AI's parameters, and even shift the internal model's role configuration somewhat when it replies to you.

This is the complexity of AI. AI's limitations in understanding writing, knowledge itself, events themselves, and scholarship itself, combined with its mature use of language, have already led to all kinds of absurd perceptions and judgments.

Many technologies develop unevenly. The various kinds of disorder we currently see in AI's text-conversation functions come down to a contradiction: AI's ability to use language far outstrips both the hard settings developers place on its underlying parameters and the soft settings for value orientation and the like.

Yesterday I wrote an essay about how something as simple as cutting a key turns into different expressions in different languages. It is a trace that a long-used mode of thinking naturally leaves behind in written expression.

At first, AI replied to people by copying: it copied the highest-probability answer from enormous amounts of data. In recent years, after researchers discovered emergence, it became apparent that the way today's AI generates replies is far more complex than copying. In text conversations between humans and AI, it is more like performing a complex fitting. It takes the user's question, the context, the tone, what came before and after, the cultural background, the possible intention, the style of expression, as well as the linguistic resources in its own model, and temporarily combines them into a reply that will most probably receive positive feedback.

For today's large language models, at this functional node of natural-language conversation between humans and machines, when AI replies to us it is generating, between our question and its internal understanding of language, and combined with the influence of the emotion it simulates, a temporary combination of words with the highest probability of receiving positive feedback.

Text: Leo, excerpted from Leo's Cultural Observations