Still don't know how to use the popular ChatGPT? This time I will teach you step by step!

This article brings you relevant knowledge about the recently popular ChatGPT. It mainly introduces how the boss teaches you how to use ChatGPT step by step. Friends who are interested can take a look below. I hope it will be helpful to everyone. help.

The text is as follows:

Recently OpenAI has launched a new work, ChatGPT, a dialogue model with amazing effects, which is already fast The netizens who were so imaginative were spoiled! Chatting, Q&A, writing code, you name it! Some foreign netizens even said that Google is done!

Here I have prepared a ChatGPT practice guide for everyone, teaching you step by step Chat with ChatGPT!

Let me talk about the conclusion first, personal feelingsGenerating model RLHF may bring us the dawn of practical text generation, let us wait and see!

The context of this article: first briefly introduce the principle, and then teach you how to register an OpenAI account and start communicating with ChatGPT. Then, the author will provide examples of communicating with ChatGPT in ten fields, and finally summarize and look forward to it. Let’s welcome the era of AIGC together!

Note: The article is long, it is recommended to read selectively. If you just want to see the effect of ChatGPT, please jump directly to Chapter 3. If you just want to see the registration tutorial, please jump directly to Chapter 2. You just want to experience ChatGPT directly. If you want to know the effect, please jump to Chapter 2: The Almighty Taobao.

1. Generating model RLHF

Two days ago, I was still discussing under the post about the diffusion model whether the idea of ​​​​the diffusion model can be used in the NLG field (it may have to be modified or combined with other technology) to improve its level of practicality?

https://www.zhihu.com/question/568791838/answer/2781856167

Unexpectedly, on the 2nd, OpenAI gave everyone a new job, ChatGPT, a dialogue generation model. The effect looks amazing, but it has been ruined by everyone.

Generative Model Reinforcement Learning From Human Feedback (RLHF), this approach feels very reasonable, and human dialogue is basically this approach (conversation based on the other party’s feedback). I have always felt that text generation not only requires new and powerful generative models, but also requires human feedback to participate (especially in interactive scenarios such as dialogue). Because I personally feel that the biggest difference between text generation and image generation is that text is exclusive to humans and does not have a natural comparison. It is more difficult to evaluate the generation effect without the participation of human feedback, whether it is GAN, VAE, or the current diffusion model. Or a combination of various diffusion models. No matter how powerful the generative model is, it will not be adaptable to the text field and can only be rootless duckweed.

So last night my friends and I lamented that this year may really be the first year of AIGC. AIGC has begun to move from the laboratory level to the practical level. The landmark event is that the diffusion model has brought AI painting to a new practical level. The text field has also begun to see the dawn of things like ChatGPT. I am even more looking forward to the upcoming GPT4, and I am even more looking forward to the practical use of text generation!

The road to text generation is long and difficult. Guys, hurry up and roll it up QAQ!

20230203 Update: When I first wrote this article, I may have overestimated the role of RLHF. Currently, according to some discussions on the Internet, the amazing effect of ChatGPT should mainly come from the powerful GPT3.5 behind it. A series of powerful language models (basic generative capabilities), followed by SFT (), and then RLHF, and RLHF may play more of a role in learning human preferences and constraining the behavior of the model.

2. Teach you how to use ChatGPT step by step

I can’t wait to see how ChatGPT performs first, but

openAI’s service does not support China, GAN (a model)!

Official website:

https://chat.openai.com/auth/login

The third step,

Bypass the problem that the OpenAI service is not supported.

The problem that OpenAI's services are not available in your country may be one of the biggest obstacles for everyone to use ChatGPT. Here are three ways to get around this problem:

Question template:

Here is a great question template for you to help you easily use ChatGPT.

##The following will give you a detailed experience of the power of ChatGPT dialogue and generative capabilities.

After successful registration, we return to the chatgpt test page:

https://chat.openai.com/chat

##Before the official trial, let’s take a look at the features of ChatGPT:

Free Research Preview:

ChatGPT is optimized for dialogue.

Our goal is to make AI systems more natural to interact with, and your feedback will help us improve our systems and make them safer.
So, try it out and provide feedback for chatgpt, so that reinforcement learning based on human feedback can help the GPT series become stronger and stronger!

Philosophy

: Let’s start with some philosophical questions:

You can see that the effect is quite good. For things it is not good at, it can also Will cleverly avoid:

##World Cup:

A few more World Cup questions:

Unfortunately, ChatGPT was trained based on data from 21 years ago and cannot answer , but just this kind of refusal to answer is already a very big improvement.

love:

Literature:

Awesome! The effect is really beyond my expectation! At least the writing assistance is useful!

#Political and Military:

economy:

educate:

Technology:

Occasionally, content generation stops abruptly.

But the effect is amazing again, can even explain the principles of GPT and transformer! I have a bold idea. If you don’t know how to interview, maybe you can ask about it!

programming:

Shocked me!

Take the classic square root calculation as an example:

It’s too 6, you can also copy the code directly!

#I made it happen from scratch, and it was not only written out, it even had comments and explanations! ! ! too strong!

I have to praise:

Entertainment:

I admit it, I'm just looking for trouble with QAQ.

The entertainment industry is too deep, you can’t control ChatGPT!

Okay, the author has done about ten aspects of ChatGPT's high-intensity test, high-intensity refers to myself, it will not tire me, I tire myself~

4. Summary and outlook

Summary ：

In fact, the summary of the Internet is quite comprehensive. The capabilities and limitations of ChatGPT:

Advantages:

• The overall effect is stunning, and it feels like it brings the dawn of practical text generation
• The fluency is very good: dialogue, continuation, and coding~
• Multiple rounds of dialogue The ability has been greatly improved
• Learned to refuse
• Allow users to provide subsequent corrections

##Disadvantages:

Generating incorrect or harmful information: a long-standing problem in text generation
• Limited knowledge: Data training before 2021 is used, and subsequent situations cannot be handled well; The performance in some fields is average, especially in niche fields
• Generating results occasionally stops abruptly
• No aggregate search engine: This is not a shortcoming in fact, but I feel it can be combined

Outlook:

Starting from the diffusion model that brought AI painting to a practical level, to various diffusion models blooming everywhere, to the sudden emergence in the last month of this year ChatGPT has brought us the dawn of text practicality, which makes people lament the rapid development of technology. Perhaps this year is really the first year of AIGC! let us wait and see!

Regarding the future direction of text generation:

Personal feeling, mainly: stronger generation model, better human feedback mechanism, more and better data, better evaluation indicators, and also need to focus on supplementing Short board.

Stronger generative models:

• New generative models: VAE/GAN/diffusion models are more suitable for images. The NLG field seems to be the only GPT series at present. Can it be better combined to create some composite models or brand-new text generation models?
• Human feedback: Dialogue scenes are naturally suitable for human feedback, but how to better collect user feedback in scenes such as novels and summaries? Annotations and barrages may be good choices, but they feel far from enough.
• More and better data: OpenAI’s data has always been plentiful and good, making others envious.
• Better evaluation indicators: At present, many evaluation indicators in the field of text generation are still character-based. Indicators such as fluency, rationality, correctness, logic, etc. are difficult to quantify, and the evaluation indicators are basically It is still completely in the hands of humans, which is why human feedback is needed.
• Fix for shortcomings: How to reduce factual errors and harmful information?

In addition, of course there are many things that can be done. One very imaginative thing is to combine it with traditional search engines to generate model search engines! When we tried it earlier, ChatGPT also said that it is not a search engine and does not have the ability to search online, but this is really possible!

These two are not contradictory, but can be combined with each other to promote each other. Huawei recently released the world's first multilingual authoritative knowledge retrieval generation model - WebBrain.

https://www.sohu.com/a/602953833_121431682

Generating a model search engine may be able to achieve what you search for.

When the practicality of AIGC begins to improve, its potential commercial value will naturally increase. Perhaps we will soon see AIGC applied to various fields and generate huge commercial value!

Let's play with AIGC!

Let's chat with ChatGPT!

Finally, let's let ChatGPT itself look into the future of text generation!

##5.ChatGPT It was mentioned that the ChatGPT search engine would be a very imaginative direction, and as expected, related issues quickly aroused heated discussions across the Internet.

https://www.

zhihu.com/question/5700

62224/answer/2787787130

As for whether it will replace search engine? The answer is of course not. The two are not competitive, but they can be combined. ChatGPT vs Search Engine: no!

ChatGPT Search Engine: yes!

The extension of

ChatGPT Google

has now appeared:

https://

github.com/wong2/chat -g

pt-google-extension

##https://github.com/ ZohaibAhmed/ChatGPT-Google

Let’s take the first one as an example to show the usage guide of this type of extension: Install the extension first:

At this time, open Google search and search for anything, such as what chatgpt is. A chatgpt-related window will appear on the right to remind you. You need to log in first. After successful login, you can see the results returned from chatgpt!

ChatGPT Google, this is so cool! ! !

There have been many extensions about ChatGPT X, including vscode extensions, mac extensions, chat robots, etc. In the future, ChatGPT

Note: AIGC really ignited my long-lost passion. I played for two days on the weekend, and it was early morning again before I knew it. Coding is not easy. If you find it useful, please support . Finally, this is a family opinion, please refer to it carefully~

20221217 Update: As reminded by the comment area, China Academy of Information and Communications Technology has made a relatively comprehensive evaluation of ChatGPT~

##20230130 Update: The new version of ChatGPT improves math and factual ability.

https://mp.weixin.qq.com/s/wYhxr5_-LF0UXhBUEIDO-A

20230203 Update: The registration process in Chapter 2 has been reorganized and is now clearer. You can choose the appropriate method according to your needs to experience the powerful capabilities of ChatGPT.

20230206 Update: Added a question template.

Original address: https://zhuanlan.zhihu.com/p/589201080