ChatGPT: AI artifact that more than 90% of people can’t use well, through this article, you can exceed 90% of users
Since the birth of ChatGPT on November 30, 2022, it has been popular in China for more than two months. However, there is no need to worry about this, because humans will always be more powerful than AI . Different from what most bloggers hype, the application logic of GPT is very simple, the interactive experience is in line with human nature , and it can be easily used only through dialogue. While everyone can use GPT, not everyone can use it well.
Today, I will use an article to lead you to master the usage of ChatGPT. The full text is practical, even a pure novice can understand it. As long as you read carefully , I guarantee that your ability to apply ChatGPT will surpass 90% of people . Before I officially start, I need to clarify:
This article will not explain too much about the basic concepts of GPT. If you need to know more, you can pay attention to the AI module in the menu bar of this official account, or wait for my subsequent updates. In addition, how to register and use GPT requires special environment and conditions, which will not be covered in this article. In order to take care of some students who do not have a special environment, I will provide an entry in this article.
So, let's start exploring the application of ChatGPT!
One: How to make AI work for you? ——Practical application skills of ChatGPT
In the past two months, I created a WeChat group and integrated GPT into the group. Through continuous observation, I found that most students have the following two problems when using GPT :
1. Why can't AI work for you?
Many people will face two pain points after getting GPT: not using it well and having nowhere to use it. This caused many people to give up on using GPT after the novelty wore off. But my situation is different. I have integrated GPT into my work, study and life, which has increased the efficiency by more than three times . It has become my personal assistant, covering all aspects of me, even the typesetting, polishing and error correction of this article are all done by AI.
If you're using AI and not creating real value, it's probably because you're lacking two things:
One is not familiar with how to use AI ;
The second is that AI is not associated with its own application scenarios .
So we will explain around these two issues to help you thoroughly grasp the application of GPT. First, we will introduce how to use AI to help you master the skills of using "Dragon Slaying Knife".
2. How to use AI?
The core method of using AI can be summed up in one sentence: The quality of the answers generated by GPT depends entirely on the way you "ask it" and the method of "guiding it" . As long as you can ask well and guide well, it can generate surprising answers for you; on the contrary, it is worthless to pretend to be empty.
The "way to ask it" here refers to the language used to communicate with AI , while the " way to guide it" refers to the method of training AI . Once you have mastered these two points, you will be able to master the ability to wield the dragon-slaying knife of AI.
So, what does this sentence mean? It's very simple: the current human technology is not perfect for natural language processing, resulting in the quality of content generated by AI is very dependent on prompt words. The so-called "prompt word" is to allow AI to accurately understand your intentions , or the language way to communicate with AI.
If the prompt words you give to AI are of poor quality or not in place, then the result returned by AI is often a pile of correct nonsense . This kind of result is not fundamentally different from the patchwork saliva you find on search engines, and it is not inspiring to you.
Therefore, if you want to obtain high-quality AI answers, the first step is to learn the language to communicate with AI, that is, to learn to write prompt words . OpenAI's CEO, Sam Altman, also known as the father of ChatGPT, once emphasized on Twitter: It is a very high-leverage skill to write prompt words for AI .
In order to understand the importance of being able to write prompt words more clearly, we can look at an example and compare the impact of being able to write and not writing prompt words on the quality of AI answers.
How to write high-quality prompt words?
Now that you have realized the importance of writing good prompt words, how to write high-quality prompt words? Here is a general and tried-and-true prompt word template: establish a role + describe the problem + set the goal + supplement the requirements . A good cue word needs to be composed of these four parts!
-
Create a role : Guide the AI into a specific scene, and give the AI an expert status.
-
Stated Questions : Tell the AI your confusion, your question, and the background information you need to supplement the AI's questions.
-
Set goals : Tell AI what you need and what you want it to do for you.
-
Supplementary requirements : Tell the AI what it needs to pay attention to in its answer, or what form you want it to reply to you.
Although this set of templates looks a bit complicated, it is actually very simple to operate . Let's take the travel guide again as an example, apply this template to the actual scene, and you can get the following effect...
Of course, you may think that the parts of "Questions", "Goals" and "Supplementary Requirements" in this template are easy to understand, because this is the language mode we usually ask questions. But, why add the action of "standing up a character" ? Is this action redundant? Wouldn't it be easier to ask directly? In fact, this action is very important! Let's take the travel guide again as an example to see the difference in the answers returned by GPT for questions with and without characters...
Through the comparison chart, it can be clearly seen that after adding the nine words "You are a professional tour guide" , the answer returned by GPT is completely different from the previous one! This means that questions with expert roles can be more specific, more practical, and the tone of the answers is more humane.
The reasons for this difference are well understood. Imagine Wang Yuyan in Mr. Jin Yong's martial arts novel "Tian Long Ba Bu". Although she has written down the knowledge of martial arts in the world, the knowledge of martial arts she has mastered is very broad and interferes with each other . Although she can give advice, she cannot provide it. targeted guidance.
However, when we add an expert role to AI, it will no longer be Wang Yuyan, who can only study hard , but will truly become a practical expert in the field, able to give more targeted answers. Expert roles can help AI specify the scene, clarify the scope of the problem, and provide the background information needed for the problem .
Because an expert in a field represents the knowledge system and the highest industry standard in that field , adding an expert status to AI is equivalent to adding a buff to Wang Yuyan. On the basis of mastering the world's martial arts knowledge, after superimposing the identity buff representing the highest achievement in this field, the AI's answer will naturally be more professional and targeted, which is quite different from the search engine-style answer.
Therefore, if you need more in-depth and professional answers instead of search engine-style answers, please remember: the first step to ask AI questions is to superimpose expert Buff on it, and then tell it your specific needs and supplements request .
Now that you understand what it means to apply the Expert buff to GPT, let's see how to apply the Buff to it . In fact, the method is very simple. In practice, I found that the following prompt words are very effective:
-
You are now [xx]
-
Please play [XX]
-
If you are [XX]
-
Please use [XX]'s angle/identity/tone.....
The above prompt words can be understood and accepted by AI, you just need to choose the expression that best suits your language habits. Once you've mastered these cue words, you can compare them to how you've asked questions in the past and gain an unprecedented level of quality in your responses.
Example 1: Using the ''prompt word template'' to design the course outline
Prompt words:
Effects generated by AI:
Example 2: Use the ''prompt word template'' to achieve the effect of a mock interview
Prompt words:
Effects generated by AI:
Example 3: Use "prompt word template" to assist the work
Prompt words:
Effects generated by AI:
The idea of this set of prompt words is universal and applicable to almost all scenarios . For example, you can use this method to design a course syllabus, to achieve the effect of a mock interview, or to assist in a job. In addition, this set of ideas can also be used across scenarios, such as using prompt word templates to realize AI painting. In short, as long as you write the prompt words according to this set of SOP templates , the answers given to you by GPT will not be too bad .
Of course, for some simple questions, you may not need to design prompt words exactly according to this set of templates. Therefore, remember to be flexible and change according to your actual scene needs . Through the above content, you have mastered the method of writing prompt words. But it should be noted that although high-quality prompt words can be designed to obtain high-quality answers through this set of ideas, the current AI has not yet evolved to the extent that it is against the sky . Therefore, for some slightly complicated questions, AI's one-time answer is often not accurate enough. If you want to get better, deeper, and more valuable answers, then you need to tune it .
Two: Tuning method
The method and principle of tuning GPT are actually well understood. The reason why GPT is powerful is because it has the ability of Chain of Thought technology (Chain of Thought), that is, it can understand the context and conduct multiple rounds of dialogue . With the support of this technology, GPT can remember the content of the previous conversation, and based on this, answer the following questions in a targeted manner, achieving an effect similar to a real-life conversation. Based on this mechanism, we can train GPT by continuously "feeding data" and "casting instructions" to help it generate more specific, in-depth and valuable answers , or achieve other effects. So, how to train GPT to achieve the effect you want? Here we need to use the following two commands.
The first command is called the "continue command" . Its essential role is to help you break through the output limit set by the AI vendor , so that the AI answer can reach its full potential . Because the training cost of AI large models is very high, in order to control the cost of computing power, major AI manufacturers including OpenAI will try their best to control the length of AI generation , and summarize the text content as much as possible to make it concise .
Taking ChatGPT as an example, its single maximum output characters are 2048 characters , exceeding this number of characters will cause the AI's answer to be truncated and stopped.
Therefore, under the two conditions of space limitation and space generalization , the answer of AI may not feel detailed or deep enough to us. Using the continue command can help us break through these two limitations. It lets the AI continue to answer beyond 2048 characters, or go into further detail if the first answer wasn't sufficient. For example:
In our AI development article example above, if the AI's answer exceeds 2048 characters, it will be truncated .
We can use the continue command to make it continue to answer the unfinished words .
Likewise, even if it answers the question, we can use the continue directive to make it go further or to specify what the answer was about. Of course, the continuation instruction mentioned here is only its most basic usage. In addition, it has further uses, for example, in the example of curriculum design we mentioned before, we can use further continuation instructions to ask AI .
When continuing to ask questions, we can use the prompt words of [supplementary requirements] mentioned earlier, such as:
Please explain with examples that children can understand,
Please provide no less than 5 examples,
Please select an example from the field of XX,
Please answer in a lively and colloquial way,
Please expand...
Please summarize...
Theoretically, we can continue to follow the routine of "continuing" to inquire deeply. For example, for the [lecture case] mentioned above, we can inquire in accordance with the original outline framework, continue to dolls to dig deeper into each point, and will inquire The results are filled into the initial big frame, and finally a courseware content that is basically completely generated by AI can be generated .
Finally, we only need to replace the language style generated by AI with our own language style, carry out logic splicing and polishing, and then we can start lecturing, which is very powerful.
It should be noted that there are two points to be aware of when using the continuation directive and the extension usage:
Note the ambiguity of the instructions . If the question is too long or there are too many layers of dolls , continuing instructions may cause ambiguity in the AI, resulting in irrelevant answers. Therefore, it is necessary to clarify the target , for example, replace "please introduce the second point in detail" with "please introduce the second point in the outline in detail", which is more specific.
Pay attention to the relevance of context . Due to the powerful ability of AI to have multiple rounds of dialogue and link context, when multiple different topic scenarios are interspersed in the same dialog box, the AI's answer may be affected by the previous content and cause random answers .
Therefore, when we want to ask multiple different topics in a dialog box, it is best to initialize the AI when a new topic is opened, and clear the previous dialogue , so as to avoid the interference of the AI's answer by the previous content. The specific operation method of resetting the ChatGPT prompt words is as follows:
The above is the [Continue] instruction for training AI. Although this instruction can make AI's answers more colorful, due to the limitations of language transmission information , each answer of AI may not always be as we wish, and may even deviate. Therefore, when we encounter this situation, we can use the second instruction of training AI, which can help us design tasks with "routine attributes" and "template" characteristics, and produce magical effects.
Before using this command, you need to define the type of routine you want to train AI, such as: copywriting, sales promotion, customer service answering, etc. Then, according to the characteristics and needs of the type, construct the corresponding routine template.
For example, in copywriting, we need to consider many factors, such as brand positioning, audience portraits, style of writing, and so on. Therefore, we can combine these factors to form a copywriting template for subsequent use.
The second instruction is called "reward and punishment instruction" to train ChatGPT through reward and punishment instructions, making it obedient like educating children
In fact, training AI based on neural networks is similar to educating children. Just like teaching children, we need to use rewards and punishments to help AI understand our expected behavior standards .
In my training, I used a reward and punishment command to train my AI assistant to generate thinking questions. First, I let the AI learn the format of my thinking questions, and then let it try to generate answers. If the AI's answer meets my requirements, I will praise it with positive words, and if it does not meet the requirements, I will correct it with negative words. In this way, the AI will gradually learn the format of my expected answer, and then generate an answer that meets the requirements.
Through repeated feeding and training, my little AI assistant finally became the assistant's right-hand man. Like educating children, reward and punishment instructions can help us train AI to become smarter and more obedient.
In practice, AI can do some patterned tasks very well. You can try to find these tasks in work scenarios, and then train the AI to generate content that meets your requirements through rewards and punishments. This way of training could make AI your long-term work assistant and help you be more productive when you need it.
However, we must not forget that the ultimate goal of AI is to help us create actual productivity. This goal cannot be achieved if we do not connect AI to the application scenarios we need. Therefore, we need to find suitable application scenarios for AI and integrate it into specific scenarios in order to maximize its value. On the basis of mastering the operational AI methodology, we need to have a deep understanding of how to combine AI with ourselves in order to truly improve productivity.
In fact, the idea of integrating artificial intelligence with ourselves is simple and involves only two steps.
The first step is "combing" , the concept is very simple and easy to understand. When e-commerce was on the rise, Jack Ma once said: "All business is worth doing again with the Internet." Similarly, in the era of artificial intelligence, we can also adopt the same idea: almost all working methods involving knowledge can be reconstructed with artificial intelligence.
Therefore, we can reflect on our work scenarios, sort out those parts that may be replaced or assisted by artificial intelligence, and find ways to combine them with artificial intelligence . Then, according to the ideas described above, develop standardized tools or processes.
In other words, sort out your daily work trajectory, find out all the tasks that artificial intelligence can handle, and then hand over all these tasks to artificial intelligence or let artificial intelligence assist you to complete them. That way, you can free yourself from these tasks to do something more rewarding and creative .
The idea is actually very simple: according to our behavioral needs, sort out each scene one by one with three basic aspects (study, work, and life) as dimensions.
Use artificial intelligence to realize dialogue learning with masters
Of course, in these scenarios, you can also replace other great figures you are interested in, such as Confucius, Zhuangzi, Mencius, Mao Zedong, Napoleon, Socrates and so on.
GPT can also play a variety of roles, such as pitting multiple bosses against each other, and you can learn from the conversation to achieve a Socratic learning effect.
The following introduces several cases of learning scenarios : The first one is to use AI to assist reading to improve comprehension efficiency . It is very practical for those students who are not strong in comprehension or want to improve comprehension efficiency. For example... The second one is to use AI to implement introductory coaches and mentors in various fields . It is very helpful for students who want to build a system in a certain field or systematically study a certain field. For example... Other than that, the same thing applies to studies in philosophy, sociology, product management, operations, etc. In addition, there are many other learning scenarios, such as debate coaches, learning effect testers, etc.
The same approach is also very useful in work scenarios . For example, using AI as a work assistant , you can let it help you write recruitment information, scripts, work copywriting, self-media copywriting, codes, plans, etc. In addition, AI can also be used for data analysis and writing meeting invitations . It can also help you if you want to spark creative work. For example, part of our article is inspired by GPT. There are many other usages in work scenarios.
There are also many examples of using AI in our daily life . For example, use AI as a fitness coach, private nutritionist, private lawyer, private doctor, private tour guide, etc. Whether it is work, study or life, AI can be used to achieve higher efficiency and quality .
Therefore, for some valuable scenarios, we need to save the trained scenario data in order to provide us with services for a long time, rather than a one-off sale. For example, the email assistant we trained can automatically read the previous data, including the signature, title, format, and writing style of the email, without telling it repeatedly, just by telling it the content of the email. Similarly, for fitness coaches, nutritionists or other scenes that require repeated interaction, the trained scenes can also be saved for direct recall next time.
After we save the trained scenes according to certain rules to form a scene library , we can reuse them, saving a lot of time and energy. The specific operation method will be introduced in the next step.
At this point, the entire AI guide is basically over. In short, when you sort out the system according to the three-dimensional idea we proposed, and train your own AI model according to the writing prompt word skills and training AI method introduced at the beginning of this article, you can classify it and fix it to your own AI scene library , you have successfully owned an AI assistant. It will help you share a lot of trivial work, life and study affairs, and greatly improve your efficiency. As long as you implement it carefully and arrange it in place, your efficiency can not only be increased by 2-3 times, but may even be increased by ten or eight times.
Of course, the last thing that needs to be reminded is that AI has not yet reached a level that is completely against the sky . Therefore, in many scenarios, it still cannot achieve the effect comparable to real people, and the answers it generates cannot guarantee 100% correctness. In some important occasions, we still need to manually correct, modify, and verify and confirm the information it provides.
In short, even with the assistance of AI, we cannot lose the ability to think independently. In the present and future, only those with independent thinking ability can control AI, not be controlled by AI .