Seeing double with LLMs

Ever since prompting became a thing, there's been a latent but significant psychological change going on, first with the AI "alchemists," then with the prompt engineers, and finally with everyone trying to make the most of current LLMs. This phenomenon is what I would like to name a linguistic "seeing double."

If you want to know more about, say, the history of the current US-Iran conflict, and you directly ask ChatGPT or Gemini or Grok (I just don't know why they like to start with a G, maybe from the AGI mania) just so, you might get a much generalized, vague, oversimplified version of the issue. If you are not content with such results, naturally you start to think, what kind of input can have a higher chance of getting a more detailed, expert-like, sharp answer. You start to finetune your prompt. This is when the "seeing double" begins.

It is no coincidence that Kantian philosophy is now being used more than any other philosophy in the AI scene. The reason is simply that Kant was the first who succeeded in bringing people's focus from the things observed to the observing organs, from the view to the viewer, from the garden behind the glass to the glass itself. Just as we start to see the world by first seeing the inevitable way, the necessary forms by which only can we see anything at all, so we start to build an artificial model with all the necessary forms we humans have. This process is exactly what happens when a ChatGPT user starts to adjust his/her input forms to get better results. You stop caring about the content, or, much more importantly, you care about the content for the sake of the form. You start to look for ways to get better results, not for this single question, but for any possible question you might ask in the future. You start to build your "apriori categories," your principles of using LLMs. You not only see the garden behind the glass, you also see the glass itself. In fact, now you see both, but primarily the glass, and the garden is only a background which helps you clean the glass.

I will give one example of my own. It was when I started to question the Chinese civil exam, due to a strange fact that despite by the end of the pandemic control, young people were very discontent with the government, the popularity of the civil exam went to an all-time-high. It was as if the young generation totally forgot the COVID nightmare and desired to join the government (and the party) fervently. So, normally such a deep question would never get any good answers if you just ask an LLM directly. My tactic was, not to ask questions, but to build a framework which helps raising questions first. I started with the ChatGPT Program feature, and divided the program into different sections dealing with different questions: a general section containing basic Q&As about the Chinese civil exam, a section about the procedure, a section dissecting the contents of the exams, a section discussing the career path and the party-government relationship behind, and finally a section about the genearational mindset behind the civil exam mania. The general section is where all the latter sections came about. By asking questions from different aspects of the issue, I got to articulate a conceptual web and organize all the details into different sections, which grew into a series of articles. It would never have been possible either by a simple asking strategy or by not using LLMs at all.

I have to admit that this is never an easy way of using LLMs. Nor is it very reusable, because most of the work (especially the setting and the typing) has to be done manually. But I do find it has specific strengths compared to many "automatic" ways, that is, it is much more down to earth, and it gives my article a humanistic touch, not only because I choose to write articles myself but also because I am the one directing all the questioning. Left to LLMs, questions raised would be very dry, paragraphs unreadable, and the whole article unrelatable.

My strategy is actually a fight against the "seeing double" tendency. In the end you have to return to the garden view, for the glass is just a tool. The Kantian error is pervasive of our age, which is to put more emphasis on the form than on the content. Hence the political dilemma of division of power when we don't even know what to do with politics, i.e. what the content of politics should be, and pretty soon the division of power becomes a barrier for tyrants and demagogues, who call for "a stronger goverment" and look for ways to go around the constitutional restrictions, and sooner or latter they will find ways. Hence the becoming more and more hollow of the modern art, the "dehumanization" as Ortega once put it, which is also a focus on the form instead of the content of artistic expression. But I am not so sure my strategy is the cure, or one of the cures, for such a modern disease. Maybe the glass, the crystal clear, translucent, diamond-like glass, is now much more interesting than the shaggy garden.