I'm sure many of you, like me, are from the "New Interpretation of the Chinese Language"This explosive cue words know Li Jigang this "man of God". Until I saw the November 4 public"Digital Life Kazik."Interview with JigangInterview with "The God of Prompt" Li Jigang - I want to spend 20 years to leave the world a message.The first thing I learned about Jigang was that he was an idealistic, bookish, and thoughtful person. This stimulated my strong curiosity and desire to communicate, so on November 13th I traveled to Beijing to have an in-depth interview with Jigang. Uniquely, as a souvenir of the signature given to the "fan", it was also a sentence he had just generated using Claude:"The simpler you write, the more widely you use it; the more you delete, the more you save."。
Still familiar place (Wangjing-Juboyuan), still familiar topic (cue engineering), during the period Jigang thoughts fly, gabbing, completely unlike a "i man", from which can also be seen in his dedication to the field and love, so that after the meal, we are not yet finished, for me personally is also a lot of benefit, so I write this article, to enjoy you all! So I am writing this article to enjoy you all.
1. The nature of the cue
The hot pot on the table was just coming up to a vaguely steaming temperature, and after a few pleasantries between the two of us, we got straight to the point.
Jigang begins by asking, "What do you think is the nature of the cue word?" , and adds, "It took me over six months to finally figure that one out."
"It's a question I've tried to think about, but really haven't done any deep thinking about it, and it's probably hard to come to a firm conclusion anytime soon." , I said with slight embarrassment and replied truthfully.
"There's no harm in letting go of thinking and speaking openly, even if it's not right, it doesn't matter, this thinking process is important too." , Jigang comforted me, hoping that I wouldn't have extra pressure.
I pondered a little, thought carefully for a moment and said, "Maybe I can't organize it in the right language right now, but I can describe my understanding in terms such as a natural form of programming language, a decryption code for a large model, a wizard in the direction of large-model reasoning, and so on and so forth."
Seeing that I had racked my brain, Jigang decided not to "torture" me any more, and responded to me by saying, "What you said is mainly describing what the cue word is, which is still on the surface, not the essence of the cue word. It's like shooting countless spears at a target, each pointing at the bull's-eye, but the spear itself is not the bull's-eye."
"The essence of a cue word, to me, is expression."Jigang directly stated his conclusion.
I believe that most people who have not thought about the nature of the cue word may have a similar reaction to me, and remain confused in the face of this conclusion, getting the shape of the word but not knowing its meaning. For those who have some philosophical foundation, it is clear that this is what Jigang understands as the "Tao" of cue words, or what he considers as the "first principle" of cue words. However, knowing the "way" does not mean that one can practice the "way".
As stated in the Tao Te Ching:
Something mixed into, first heaven and earth born. ...... independent and not change, circumference and not dangerous, ...... I do not know its name, strong word is said to be Tao, strong for the name is said to be big. ......
The Way is there, always there, but it is not easy to realize the unity of knowledge and action. You may understand what "expression" means, but you may not understand what "the expression expressed by the cue word" means. Tao is so abstract that it needs to be "realized" to further analyze it, and Jigang gives the first step of analyzing the expression.
In short, the intended meaning is the idea in a person's mind, indicating what they want to do. Context is the symbolization (cue word) of the intended idea, indicating how the idea is described. Deconstructed meaning is for the big model to understand what the person thinks, indicating how the idea is interpreted. Through these three stages of deconstruction, the process of the cue word from the person to the big model, which is the essence of the cue word, can be meticulously restored.
The purpose of a cue word is to take the idea in a person's mind (the intended meaning) and provide it to the big model to understand (the unintended meaning) with precision and accuracy. Excellent cueing engineers are good at mitigating the gap between the intended meaning and the interpreted meaning by controlling the cueing words (context). The first and foremost thing is how to accurately depict the intended meaning and tell the big model what you really want to do. Then comes the optimization of the context, try to convey the original meaning to the big model without any loss, that is to say, optimize the cue words. The last one is the big model. Although you can't improve the understanding ability of the big model through the cue words, you can choose the big model with better understanding ability.
2. How to be clear
Jigang spent a year figuring out how to clearly describe the ideas in his head, that is, how to portray the intended meaning.
First is experience, experience is the embodiment of ideas, which is the basis of one's understanding of ideas; ideas without embodied experience are castles in the air. Second is vocabulary, which is the linguistic mapping of experience and the symbolic form of ideas. Finally, there is knowledge, which is the detailed interpretation and description of the meaning of words, and is the basis for the symbolization of ideas into precise vocabulary. With the above foundations, one can accurately express the ideas in one's mind and achieve clarity of expression.
3. How to improve results
Analyzing how to improve the effectiveness of Q&A for large models, Jigang gives this thought.
-
The first thing to recognize is that the big model is an amplifier, not a wish granulator. What you sow, what you produce, the input of the cue word directly affects the output of the big model, which corresponds to the part of the interpretation here.
-
Second, it is important to define clearly what the problem or task to be solved is, and describe the intended meaning. Here, Jigang gave a very graphic description: "People should be fiercer than AI". To put it plainly, don't be afraid of AI, have confidence to manipulate AI, there is a kind of "strategic contempt for the enemy" meaning.
-
It is finally the cue word, which expresses the idea through the text and corresponds to the context. The cue words should be precise enough to help the big model locate accurately in the Embedding semantic space. The cue words should be concise enough to help the Attention mechanism of the big model to realize the association of key meanings.
4. Cue engineering methodology
Above, is just following the summary of the cue project of the "road", then describe the cue project of the "art", that is, how to write a good cue word.
4.1 Johari Windows
How do you write a cue? Johari's Window is a theory about communication skills, but Jigang skillfully transferred it to the scene of communication between people and large models, depicting the basic communication framework of cue engineering, which is a little bit like "seeing people talking about people, seeing ghosts talking about ghosts".
Statically, quadrants one and four in Windows are described in a way that is more familiar to most people, with quadrant two requiring a higher level of ability to ask questions and requiring a person to be constantly learning (see:Problem Hammer), the third quadrant requires people and big models to collaboratively explore uncharted territory and boundaries.
Dynamically, in the future, the big model will know more and more (the X-axis keeps moving down), then it becomes important for each individual human being to enrich his or her cognitive boundaries (controlling the Y-axis).
Finally, for different quadrants, the description of cue words can be flexibly adjusted to optimize the overall state of cue words (Debug cue words) from small to large and from simple to complex, which is the basic logic of cue engineering.
4.2 Extreme compression
Jigang has summarized his style of cueing; last year he worked on the clarity of cueing, while this year he focused on the compression of cueing.
How can cue words be simplified? The easiest symbols for big models to understand are vectors, both rather tokens, words, sentences, etc., and vice versa for people. Obviously, words are the most efficient form of communication between people and big models. The LISP language (the first functional language), which treats functions as first-class citizens, has a code form that is identical to the data form, and this high degree of simplicity makes it a perfect match to act as the "glue" between words, and to build the most compressed representation of cue words!
Although this was Jigang's initial personal conjecture, after countless experiments, the big models (especially Claude) have the ability to understand this form of cue words, and really do what Zhuangzi said, "to get its meaning, forget its words", the wonderful to the tip of the iceberg!
The Tsueners so in the fish.lit. get the fish and forget the Tsuen (idiom); fig. to forget the whole country when one has the fishThe hoofers are so in the rabbit.lit. get the rabbit but forget the hooves (idiom); fig. to act as a go-between; that is why the speaker cares.beside oneself with joy。
4.3 Light up the stars
So how do you find the most appropriate word? Since the big model has the ability to understand words, the question of which words to use is critical. As the concepts of "vocabulary" and "knowledge" were mentioned earlier in the deconstruction of the original meaning, we try to use definitions of words rather than descriptions of words (cf:The Spear of Definition), so that the words in the prompts "hit the nail on the head."
This is something that is easier said than done ......
Jigang often uses the metaphor of "lighting up the stars" (cf:The bow of analogy) to describe their own search for essential vocabulary.
Imagine I'm in an unlit room (vector space) surrounded by dull stars (word vectors), I can call out the names of the stars to light them up, and when I light up the stars in order, the lines between them form a star map, and the big model can understand the meaning of the patterns in this star map and do a generalization of the output, and when I write the cue words, my mind is actually setting off fireworks.
4.5 Read in. Prompt out.
Finally, how to write a good prompt? We should be able to see that the rich accumulation of knowledge, in-depth understanding of vocabulary, mature engineering literacy, for the writing of a good prompt word is crucial, there is no shortcut, to use the words of "oil seller", can be expressed as: "no other, but reading. Read more, think more, write more, each person can realize their own cue word of the "way".
5. Cue word engineers
Looking back at the role of the Cue Word Engineer, he is a cross-disciplinary position. To borrow Jigang's exact words, "A cue word engineer is as much a cue word writer as an engineer; the answer to the riddle is in the riddle."
A technical person who loves collaboration, or a creator who loves technology, will be the best talent portrait of the cue word engineer. Wang Xiaobo is certainly a typical creator, and among the technical people, there is a category of roles that will also fit well, they are called "open source evangelists" ......
6. Postscript
After chatting with Jigang on the 13th, my brain has been in a feverish state, and in the two days of 14-15, I caught up with the Global Machine Learning Conference and started the "crazy social" mode, so I had no time to organize my thoughts. Coincidentally, on the 16th PEC 2024 (Cue Engineering Summit), Jigang gave a speech on "The Dao and Technique of Cue Words", the theme is basically the same as the content of our interview, so the article I also quoted directly from the content of his speech in the PPT as auxiliary material.
Built a small group, we talk about cue word technology, interested students can enter the group to keep attention. To the future of the excellent cue word engineers, together!
Lights outside the window, finally done with this part of the insights of the meticulous organization, but also by the family to steal the state of late-night coding ......
In addition, you can also directly follow the public number of "Structured Word AI" of Jiangshu in the cloud, and find"The Art of Prompt Design and Building AI Native Products"Watch the live recap video of the sub-forum directly (starting at 47min).
7. References
- Interview with "The God of Prompt" Li Jigang - I want to spend 20 years to leave a message to the world:/s/JT2oOG2SYw2pDYEHlEmcyQ
- Claude Prompt: A new interpretation of the Chinese language:/s/7CYRPFQxi37ONTlX0hfzRQ
- Claude Prompt: Hammer of Questions:/s/KlkomVKEYKjVAb6NEXcjSg
- Claude Prompt: the spear of definition:/s/eNcqU-_-8SMpVBXAcgeQRQ
- Claude Prompt: the bow of analogy:/s/p1viD22cPtD3iLzOIb_FMg
- All about talking:/book-detail?type=1&senderVid=4000012&v=10132d20813ab77a6g012034
- Why greatness can't be planned:/book-detail?type=1&senderVid=101531&v=0bf32020813ab7e6bg016510
- The math of deep learning:/web/bookDetail/01d327c071a122c701d71f3
- Inflection Point:/web/bookDetail/08732220811e7ef55g012f82
- GPT illustrates how large models are constructed:/web/bookDetail/e0d32f10811e7ee55g010619
- This is ChatGPT:/web/bookDetail/74332a90813ab86c4g019d98