13 spells to tap the maximum potential of GPT-4, Github Wanxing AI tutor is popular, netizens: Every line is no longer separated by a mountain

Just need a prompt word to turn GPT-4 into your "Dream Love Guide" and gain tens of thousands of stars on GitHub.

The content and difficulty of learning can be specified arbitrarily, from primary school to doctoral, and even the teaching style of the AI ​​tutor can be changed And speaking tone:

Finally, I don’t have to go to a tutor for a difficult (ruo) problem that I don’t understand!

Now, an AI can help you learn any topic, and it is still on call 24 hours a day and never loses patience.

For example, if you use it to learn a new programming language, Rust, and tell AI that you know C language, AI will skip basic knowledge such as data types and functions when teaching you.

There are even people on Github who have received great help and actively asked to sponsor the author's development funds.

#A friend who is engaged in overseas e-commerce and often needs to understand products in different industries and markets in different countries commented after the experience: I can’t do anything else anymore. No more mountains.

To realize this AI tutor Mr. Ranedeer (Mr. Reindeer), there is no need to train or fine-tune the AI ​​model, or even write code, it is completely dependent on Write prompt words to AI and it will be done.

And it has been evaluated by people in the education technology (Edtech) industry as being on the same level as the current most advanced educational technology .

Mr. Ranedeer’s version is still under rapid iteration, and the current prompt word is ridiculously complex, with a total of

7839 characters , 3721 tokens are consumed on GPT-4.

can be regarded as the "ceiling" level of prompt words for large models at present.

The author who developed this prompt

is only 17 years old and is still a high school student...

Exclusive tutor, from elementary school to PhD

The method to invite AI to be an exclusive tutor is very simple. Directly copy the complete prompt word into GPT-4, or click on the ChatGPT sharing link provided by the author and select “Continue this conversation” will do.

The next step, of course, is to ask the teacher to speak Chinese instead.

Next, the AI ​​instructor will remind you to adjust the basic settings. Just type "/congfig" and then type in your choices.

The supported options are very rich, with various colors and styles for you to choose from. There is always one that can meet your needs.

Among them, the "visual" learning style needs to open the plug-in and let GPT-4 call Wolfram and other functions to draw pictures.

If you are really not sure how to choose, you can also ask AI to explain the meaning of each option.

Here is a hidden function, except that the depth is defined in the prompt word, other style options are understood by the AI ​​itself. You If you want to choose a style that is not listed, you can actually do so.

You can even specify only the depth, and let the AI ​​randomly select another one. Each class will have a different feeling.

After you determine the topic you want to learn, enter it through "/plan", and the AI ​​tutor will tailor a study plan for you.

For example, Xuexue’s vector database was recently popularized by ChatGPT.

As you can see, the AI ​​tutor will first write a line of "self-reminder" to avoid forgetting the configuration, and then make some in-depth assumptions about you Already know the basics.

At the beginning of each subsequent class, the AI ​​tutor will repeat this "self-reminder" to ensure that it does not deviate from the lesson plan.

After the course starts, the AI ​​instructor will give an explanation first. It can be seen that the "analogy" method is used to teach according to the configuration, and the vector database is analogized to a library.

After you finish studying, you can enter "/test" to enter the quiz session.

The AI ​​tutor will judge whether your answer is correct, make comments, and judge whether you can enter the next class based on what you have learned.

And the AI ​​will praise you appropriately. If you like this, you can set the tone style to "encourage" to make the AI ​​praise you more often. Praise.

#If you have other questions, you can always "raise your hand" to ask.

After greeting everyone, enter "/continue" to enter the next lesson.

When you find that the test session of the second lesson is too difficult, you can temporarily ask the AI ​​instructor to lower the difficulty.

After lowering the difficulty level, try another in-class test. It can be seen that the difficulty level has indeed been reduced.

Unknowingly, I have used up the GPT-4 quota of 25 entries every three hours.

Yes, the biggest obstacle preventing you from learning is the GPT-4 usage limit and context length.

And the author specifically emphasized that "GPT-3.5 is not recommended", and because the prompt words are too complex, calling the GPT-4 API will be very expensive.

Other large language models are also available

The good news is that this set of AI tutor prompt words can be used in other large models besides OpenAI.

The author tested Claude, the strongest competitor of ChatGPT next door, and could successfully complete a course, but there are some subtle differences.

For example, Claude is not as good as GPT-4 when teaching poetry analysis.

However, Claude will be more effective when teaching mathematics, and will also suggest directions for continuing learning after all courses are completed, while GPT-4 is just Tell you that you have finished studying.

It seems that AI tutors are also divided into those who are good at teaching liberal arts and those who are good at teaching science?

In addition, no matter which large model is used, the situation where AI talks nonsense still exists.

The classic knowledge that exists in the training data is fine. If you want to learn something off-the-beaten-path or cutting-edge, it is best to let the AI ​​read a piece of reliable learning material before teaching you.

At this time, Claude-100k, which supports 100,000 token contexts, becomes a good choice.

In addition, Sam Altman also shared OpenAI’s short-term roadmap in a recent discussion, which mentioned that a context window of up to 1 million tokens is possible in 2023.

Before the arrival of a more powerful large model, in addition to using the AI ​​tutor first, the very complicated prompt words behind it are also worthwhile study.

Created by a 17-year-old high school student

The design ideas for this AI tutor named "Mr. Reindeer" mainly includeoverall architecture designandcourse content generationTwo parts.

On the one hand, it is the overall architectural design that carves out the entire "personality" of the AI ​​tutor.

From the content point of view, it includes name, version information, author, course level, instructor personality type (storytelling/humor/serious, etc.), reasoning framework, instructions, rules, user preferences, configuration, course settings , initialization and other contents.

You can download the full version of the prompt word directly on GitHub:

The prompt word also supports multiple formats, including JSON and YAML and Markdown.

So, how does this long "mantra" given by Mr. Ranedeer in GitHub work?

We directly asked GPT-4:

##GPT-4 quickly helped us analyze this The

operating principle of a large string of codes mainly includes six parts: characteristics, commands, rules, student preferences, formats and initialization:

# #Using GPT-4 to summarize, this prompt word usually has the following characteristics:

Set specific rules and formats

• Personalized preference settings
• （ Optional) Use icons to increase readability
• Command interface
• Structured output

Among them, the core method for the AI ​​tutor to maintain his personality lies in the writing of the rules

part of this long series of spells.

The format of this part is relatively close to the "Microsoft TrainingGitHub Copilot

prompt word" that was previously leaked online.

Although the "GitHub Copilot prompt word was leaked" is not true, judging from the content of this prompt word, the rules are written in the same way Including specific scene rules, clear role identities, behavioral norm guidance and interaction mode settings

, etc.:

So, overall After the structure is designed, it is time to design the prompt words for specific courses.

On the other hand, in course content generation, the AI ​​tutor draws on the currently very popular large model "thinking tree" principle.

From the picture, the core principle of this prompt word is "stepped iteration", that is, the purpose is to let AI repeatedly generate and evaluate Course quality, and ultimately output the best course content to users.

Both course content and course quality assessment are all completed by AI. Taking a high school physics class as an example, the prompt words need to include Generation Goal (High School Photoelectric Effect Course), Thinking Tree Steps and Rule Limitations:

Give me a high school lesson on the photoelectric effect.
First, create 5 plans to introduce how you plan this course.
Then, create a table with all proposed plans and rate them from 0 to 100, including rating them on their interactive consistency and usefulness (based on your language model capabilities).
Subsequently, a complete lesson is generated using only the best grading plan and iterated 3 times so that we can analyze the best individual lessons we should give students (using the selected plan).
After you finish the lesson, rate the plan again just as you did for the plan (engagement, coherence, usefulness) on the form. Finally, output what you think is the best course and why.
NOTE: This course is generated solely by you. That is, you will use your large language modeling capabilities to teach this topic. Don't initiate activities because you can't see what students are doing. It can only be text interaction.

In fact, the technique of writing prompt words in the form of a thinking tree (chain) is now "standard" in various cutting-edge tutorials.

For example, in the new ChatGPT prompt word engineering course jointly opened by Andrew Ng and OpenAI researcher Isa Fulford, a similar Thinking Chain# is also introduced ##How to write prompt words.

Taking the design of an AI computer customer service as an example, we can use the thinking chain method to let it answer the customer's questions according to the following logic Question:

First, determine the type and quantity of computer products that the user wants to ask;

Then, determine whether the type of product the user wants to ask is within the scope of the database;
Next, answer according to the content of the database Customers have questions about computer products, and use a friendly tone to correct their unreasonable questions;
Notes include answer methods and formats, etc.

To sum up, first design the overall framework, and then specific to the course content, it can be generated in the form of a thinking tree (chain).

What’s interesting is that the author who wrote the prompt words for this AI-specific tutor plug-in turned out to be a high school student who is only

17 years old.

According to the Financial Times, the high school student’s name is Justin and he is from Australia.

Like everyone, Justin also faced the situation of "cramming" one week before the exam. For this reason, he created a GPT-4 personal trainer to give him extra lessons and found that the effect was very good.

After the exam, he continued to teach himself more advanced prompt word engineering and developed a more complete AI tutor program that can be customized according to different course levels, learning styles, reasoning abilities, tone and interaction methods.

Subsequently, he named it the Mr. Ranedeer project and uploaded it to GitHub. Unexpectedly, it became popular unexpectedly.

Justin said that when he was 14 years old, he had already had the idea of ​​letting AI be a teacher.

It is worth mentioning that judging from Twitter, this guy is also a fan of Musk and has not left:

In the contact interface, he also gave a string of Dogecoin codes so that everyone can reward him with Dogecoin support.

One More Thing

After reading so much, are you a little tempted and want to "copy" the homework? Similar prompt words to develop your own AI plug-in?

As always, just ask GPT-4 and you will be able to extract a template very quickly:

Truth is everything All can be GPT-4.

Get a tutor:https://github.com/JushBJJ/Mr.-Ranedeer-AI-Tutor

The above is the detailed content of 13 spells to tap the maximum potential of GPT-4, Github Wanxing AI tutor is popular, netizens: Every line is no longer separated by a mountain. For more information, please follow other related articles on the PHP Chinese website!