Visual China.
Titanium**Note: This article is based on WeChat*** Venture Bang (ID: ichuangyebang), author |Wang Yi, editor丨Haiwaist, titanium** is authorized to publish. Prompts, which were born together with large models, have become an unavoidable concept in the AI field.
Sam Altman said in a conversation with LinkedIn co-founder Reid HoffmanIn five years' time we will no longer need prompt engineering, or just a small amount of work to do it;In the future, AI systems will not produce completely different outputs because a specific word is added, but will be able to understand natural language better. Users only need to input instructions in the form of text and voice, and the computer can complete complex tasks such as image generation, data research, and psychological counseling.
Sam Altman's statement has led to a general perception that he is not optimistic about the prospects of the prompt.
But Sam Altman's disapproval is aimed at the prompt word project, not the prompt word itself.
In the future, when the large model is becoming more and more perfect and mature, it is not necessary to let the large model better understand natural language in the form of prompt word engineering, but the prompt word itself has infinite possibilities.
Don Valentine, the first generation of Sequoias, once recalled that the people he had ever met who could ask the most questions were Steve Jobs and Michael Moritz of the second generation of Sequoias.
The essence of a prompt word is a good question, not a prompt. No matter how good a model is, a good question will always have value.
The development of prompt words is closely related to the development of generative AI and large models.
After the fire of the work "Space Opera" at the Colorado Art Fair in August last year, AI painting models such as Midjourney, Stable Diffusion, and Dalle quickly became popular all over the Internet, accompanied by various AI painting prompt words** and courses, such as the promptbase for the prompt word trading market, and the AI gallery for the prompt word generator.
The second wave is ChatGPT, especially after the release of GPT-4, users have seen the strong strength of AI chatbots and began to explore a variety of prompt word plays, and the profession of "prompt engineer" has also come into being, and there are many prompt word courses on Douyin under the banner of "becoming a prompt word engineer with an annual salary of one million", and prompt word communities and knowledge planets with threshold fees ranging from 200-1000 have also sprung up.
The third wave is after the release of GPTS in November this year, the creation process and use scenarios of prompt words have changed greatly - GPTS encapsulates some "original prompt words" in a small conversation robot in the form of dialogue, and users create their own GPTS through dialogue with GPT (the process of dialogue is also the process of entering original prompt words), which can be used or distributed for personal use, GPTS can be based on its original training data, more efficient, Solve user problems professionally.
After the release of GPTS, a number of unofficial GPTS stores such as SuperTools, Suefel, GPTS Hunter, etc. have sprung up, and users have shown unprecedented enthusiasm for "hand-rubbing large models - building their own GPTS", and have made data analysis, academic**, English learning, martial arts secrets, tarot calculations, writing poems for you, and simulating dad's ......and all kinds of GPTS. As of December 16, about a month after GPTS was opened to users, 61,818 GPTs had been gathered on GPTS Hunter, and the number of GPTs on ChatGPT as a whole far exceeded that number.
GPTS has given full play to people's imagination and creativity, and they have come up with some rather "explosive" GPTs:
For example, the "prompt pet (prompt pet)" made by AI product manager Chen Caimao allows users to get the prompt words they want by entering their needs
For example, there is a foreign god who has made a GPTS called "grimoire (magic book)", which can be regarded as a "no-** programming system", where users can create ** and applications by typing a sentence or constantly talking to grimoire.
In another example, designer and writer Jackson Greathouse Fall gave GPT-4 $100 on March 15 of this year, and then asked it to order itself to do all sorts of things. ChatGPT first asked him to make an environmental protection theme ** to sell all kinds of peripheral products;Then ChatGPT taught him how to use DALL·E 2 to design this ** logo, and even gave DALL·E 2 a prompt for him to put it directly into this model;Then, I taught him how to write a **, and gave him a part of the source**, and he quickly set up this **. Later, GPT even told him how to raise money and helped him make a PPT of a financing business plan.
A week later, Hustlegpt helped him earn $130 and 2,095 followers on Discord;Jacson has also really received an investment from an angel investor, and his company is now valued at $25,000.
These three waves of AI have not only amazed people at the speed of technological development, but also boosted the popularity of the concept of "prompt words". More and more people are paying attention to this area and are looking for more possibilities for "prompt words", and Infoark community manager Liu Ruilin is one of them.
Liu Ruilin was previously the head of content for an internet medical company and one of the first users to contact ChatGPT. After the release of ChatGPT, he found that what would otherwise take 10 hours a day could be completed in just four hours, which made him very excited, and he began to explore various AI tools while becoming a big believer in the AIGC field. In June of this year, he left his job to start his own business and launched the InfoArk community with a few friends.
The InfoArk community was formerly known as "Info.""Ark of Tomorrow" is an open-source knowledge base in the field of AIGC, which mainly provides basic knowledge in the field of AIGC, related courses, information source recommendations, academic **, tool introduction, etc. At present, InfoArk's main focus is on the field of prompts, and in the main InfoArk community documentation, there is a very detailed introduction to the principles of prompts, related tutorials, and command libraries, which can be easily read even by beginners.
At present, the InfoArk community has accumulated a large number of paying users, and Liu Ruilin's entrepreneurial team has also obtained a stable cash flow through Knowledge Planet subscriptions, high-end paid courses, and corporate business partnerships.
However, this is not what they are ultimately trying to achieve, in their view, the InfoArk community is just the first step in gathering traffic, and what they really want to do is something similar to the "uploaded intelligence" in the recently popular cartoon "Pantheon" - a digital "second brain".
We believe that there is still a large part of human intelligence that has not been developed, for example, many people are not good at mathematics because they do not have the thinking of combining numbers and shapes, and they may need to develop a product to make people understand mathematics through visualization and combination of numbers and shapesAnother example is that the human brain's ability to filter information is limited and cannot process massive amounts of information. Then we have to find a way to filter the information. This all means that you need a set of software based on the "second brain" to help you better build your mental algorithms and mental models. This set of mental algorithms and mental models is 'software', which needs to be attached to hardware, it exists on the human body and may die, but it does not exist on computer hardware, because hard disks can be copied and transferred, but this consumes a huge amount of energy. Compared with machines, the final dignity of human beings may be to use very little energy to invoke the 'cognitive model' in our body to digest knowledge and gain insights, which is the advantage of humans over machines, somewhat similar to 'intuition'.
We should strengthen this advantage, so we want to make a 'knowledge life cycle management' product, which can simulate the human mind, realize the whole process of knowledge creation-inheritance-utilization-distribution-destruction, pass on the advantages of human cognition and mind, and build a 'digital clone' similar to **atar, which may sound similar to note-taking software, but its input and thinking methods are completely different from note-taking software, and it is carried out through dialogue. We're going to design a new prompt -- a question system to talk to you, and through this question system, it can learn about your inner values, cognitive models, decision-making models, and other underlying mental algorithms, so as to build your simulated 'second brain'," says Liu.
In Liu Ruilin's view, prompt is a means of data cleaning and information processing to be more "engineering", because in the training process of large models, the continuous pursuit of better training results with lower costs has led to prompt. He thinksThe greatest value of prompt lies in how to turn tacit knowledge into explicit knowledge through process, standardization, and automationA good prompt includes understanding and thinking about the business, a good prompt process should be a systematic process, and a good prompt should be encapsulated into a workflow to solve practical problems, which is also what they will try to do in the future "second brain" products.
In addition to Liu Ruilin, Yuan Liuwei also pays for knowledge in the field of prompt.
Yuan Liuwei is the manager of the Knowledge Planet's "AI Instruction Club", and he is also the first wave of self-taught command engineers in the private sector after the release of ChatGPT. He has customized prompts for more than 30 companies such as Haier and iFLYTEK, with a prompt ranging from 5,000-20,000 yuan, and achieved a monthly income of 100,000 yuan through instruction customization.
Yuan Liuwei believes that there will be two development paths for prompt in the futureOn the side of ordinary users, as AI understands semantics more and more, prompts will become more and more simple to use and closer to natural languageOn the professional user side, prompt will develop into a "language", similar to a programming language, and there will be a dedicated prompt engineer position to use AIVertical fields such as scientific research, data analysis, technology development, and content creation require specialized command engineers to design and optimize prompts to guide AI to perform complex tasks.
In the future, AI will definitely become the infrastructure of people's lives like electricity, but the real value of AI lies in 'service'. OpenAI will definitely polish some instructions to be embedded in the large model for users to use better, but people's needs are diverse, and the official does not have the energy or ability to cover the instructions in various fields, so in the vertical field, experts from all walks of life like us are needed to create instructions. I think prompt will become a language and a discipline like the current programming language, and its market will be 100 or 1,000 times larger than that of programming languages, because it is based on natural language, and its audience and use scenarios are much larger than programming languages," Yuan Liuwei said.
If Liu Ruilin and Yuan Liuwei represent the views of the non-technical school, then in the eyes of the technical school, there is still a lot of room for improvement and imagination in the prompt words.
Yunzhong Jiangshu is the initiator of the recent hot "structured prompt word" writing paradigm, the author of the ChatGPT Chinese Guide project with 8K+ Star and the LangGPT project with 2K+Star on GitHub, and the co-founder of the Embraceagi open source community. The LangGPT project has set up a set of "templates" and "frameworks" for prompt words, and by setting hierarchical structures, identifiers, attribute words, etc., many novice users can easily write prompt words with good results by filling in the blanks.
Taking the femdom GPT production "poet prompt" as an example, the prompt written by langgpt is like this:
And when we input this set of prompts into chatgpt, the result it gives is this:
Jiang Shu Yun shared with us a case of the coolest prompt words he thought he wrote using the langgpt structured prompt word method: a super cool teacher who is good at using the simplest vocabulary and plain language to teach students with 0 basics.
The author of the above prompt, Li Jigang, is an Internet product manager, and he is also a enthusiast of prompts with a technical background. He believes that the ** generation field is now basically deprompted, and many of the previous spell writing methods are not needed anymore;But in the world of text generation, prompt words are still needed.
Li Jigang put forward the "dream weaving theory" of prompt words - the process of writing prompt words is to create a "dream" for the large model, writing prompt is weaving dreams, and prompt engineer is the dream weaver. "Prompt" is like a clue that guides ChatGPT into the depths of the dream you weave. The more skilled the dream weaver is, the more real the dreams woven through the prompt will be, and the more ChatGPT can become a "dream person".
Yunzhong Jiangshu and Li Jigang both regard prompt as itProgramming languages in the AI era, and they all take a clear stand against the "theory of the demise of prompts".
One of the obvious features of the prompt word is that it uses natural language, and each country can use it in its own language to talk to the machine. If you think of it as a programming language, then you will come to two conclusions - first, there will be more programmers, as long as anyone with an account and access to a large model can program, then there will be more different kinds of ideas;Second, programmers will be differentiated, there must be a wave of people who study more deeply (prompt word engineers), and another part is relatively shallow (ordinary users), prompt word engineers will try to make the dream weaving thing heavier, more structured, and more logically complex, while the ordinary user side is more lightweight and easier to write prompts. One goes in the direction of the front end, and the other goes in the direction of the back end, all for the purpose of making the part of the dialogue between the dream person more lightweight. Li Jigang said.
Based on the above conclusions, Li Jigang believes thatThe era of fine-tuning models will come, and companies will definitely fine-tune their own data with the ability of large models to form their own unique "small models."This kind of "small model" is closer to the application scenario and has stronger understanding ability, and it is also an inevitable evolutionary path from GPT-4 to GPT5.
Yun Zhongjiang believes that the future prompt words may be multimodal. "The prompt itself is not a transitional product, but a long-term trend in product evolution. With the development of multi-modal large model technology, in the future, we can use emojis, dynamics, etc. as part of the prompt words, and the output we get can also be the result of ** sound。The upper limit of prompt has not been lowered, but because of the emergence of this kind of thing, the upper limit of prompt has been raised. ”
Chen Caimao, the author of ChatGPT Advanced Prompt Engineering Introduction, also mentioned the idea that prompts will become "multimodal" in the future. He thinksAs AI technology advances, prompts may disappear, but "prompt engineering" will persist.
Chen Caimao divides prompts into two categories:
The first type is Pompt, which compensates for the shortcomings of AI. The role of such prompts is primarily to "compensate" for the lack of model capabilities;The second category is prompts that "help AI understand human needs". In fact, we use prompts to define a goal for AI, explain the business clearly, and help it understand human needs.
For the first type of prompt, a very classic example is "the sweeping robot does not avoid it when it encounters shit". A robot vacuum cleaner did not avoid it when it encountered dog poop, but continued to "clean". As a result, dirty stuff fills the house. In this case, the prompt engineer has to issue seemingly nonsense prompts such as "avoid dirty things when sweeping the floor, don't drag it all over the house", and "need to sweep dirty places several times".
This is actually a sign of low AI intelligence, or insufficient understanding of human needs (alignment). However, with the rapid development of technology, we don't even have to wait for the model itself to improve, and some product designs can solve some of these cases. Therefore,Such prompts may soon be retired.
For the second type of prompt, also taking the robot sweeping as an example, suppose that the AI is very smart now, not only knows to avoid dog poop when it sees it, but also knows how to take a small shovel to shovel it up and throw it away. But when you sweep the floor in your house, you must have the rules of your house, such as you must sweep the living room and then the kitchen, and you are not allowed to make a little noise when cleaning;If you meet your girlfriend, you must also "say hello to your esteemed young grandmother";To enhance the entertainment effect, the robot also does a backflip while cleaning.
In this case, no matter how smart the AI is, it is very likely that it will not be able to understand the situation and the effect we want, which is "not understanding the business". At this point, we need to list the business rules in clear language, such as what the "order" of "cleaning" looks like in a specific order. Therefore, the second type of prompt can survive for a long time.
In addition, Chen Caimao also mentioned a very important trend in the development of prompt wordsDefacilityNow the human-computer interaction centered on software and functions will evolve into AI-centric human-computer interaction, and users only need to say their own needs, without any medium, and AI can immediately realize the user's needs.
What impressed me the most during this time was an open source project called Open Interpreter, the official slogan of this project is 'a new way to use computer', which means that you can give it a command at any time, and the AI will analyze, plan, and then write ** to implement your needs. For example,In the past, we had to make a feature or softwareBehind this, a whole set of complex processes such as writing requirements documents, passing review meetings, development, testing, etc. But now, AI is getting stronger and stronger, even if you are the only one in the world who has this need, it is possible to write ** to meet you on the spot", Chen Caimao saidIn this case, the idea becomes much more valuable than the implementation
If the above scenario is too far away, what is the next trend in generative AI and prompt words in the next one to two years?
Well-known prompt word engineering document learningpromptJimmy Wong, the creator of the wiki and founder of the open-source app Polestar Chat, is convinced that it isGUI (Graphical User Interface) + CUI (Conversational User Interface) combined products.
He believes that a good example of making up for the lack of CUI prompt through GUI is the Wensheng graph and Wensheng ** workflow represented by comfyui.
Comfyui is a node-based process-based Stable Diffusion AI drawing tool, which can split the process of Stable Diffusion into nodes to achieve more accurate workflow customization and perfect reproducibility.
Comfyui's interface is intuitive and easy to useEach step of AI painting is split into a node, such as load checkpoint, sampler, prompt, etc., all exist in the form of nodes, so that users can quickly get started and draw easily.
With Comfyui, users can achieve real-time AI painting effects by adjusting points, modifying brushstrokes or hints. Its biggest advantage over the traditional Wensheng diagram large model is that it "points where to play" - we always complained that AI drawings rely on card drawing, but the emergence of Comfyui has made Wensheng diagrams controllable. At the same time, Comfyui can also make AI** and generate AI animations, and the effect is not inferior to Runway and Pika, which has recently become popular.
It's the Warring States era of the Stable Diffusion model, and there is no giant like OpenAI relative to LLMs. I think AI-native products may appear in this space, and even AI-generated movies may appear next year. For small entrepreneurs, this may be a more promising direction," says Jimmy Wong.