The big guys' pursuit of automating everything is endless. When we lamented the power of ChatGPT, they launched Auto-GPT[1].
Auto-GPT leverages OpenAI's GPT-3.5 and GPT-4 model capabilities to interact with online software and services, allowing it to perform our tasks "autonomously".
But can it really achieve complete automation, and can it really free your hands? We will wait and see.
The version experienced today is 0.3. It can be seen that the new version has added the Plugins function, and third-party developers can extend the functions of autogpt by writing plug-ins. This way we can add more interactivity to it.
Note: Because the demo needs to use the API key of OpenAI, in order to avoid the account being banned, it should be used with caution in the domestic environment. This time I will continue to use the Google-colab[2] service for free to demonstrate the capabilities of Auto-GPT. If you don't have a way to access colab, you can also use the one provided by github codespacesto (it's a little slow in China, but it can be used).
Install
gitWe clonechoose the stable branch through the direct Auto-GPT project.
!git clone -b stable https://github.com/Significant-Gravitas/Auto-GPT.git
%cd Auto-GPT
复制代码Change .env.templatethe filename to .env. And fill in your own OpenAI API key correctly.
OPENAI_API_KEY=your-openai-api-key
复制代码start up
!sh run.sh
复制代码Because he will install the dependent files first, the first startup will be slower. When you execute it locally, there may be various dependency installation failure problems. It is recommended that you replace the domestic mirror source.
to experience
Our goal: Let autogpt write a 1000-word article, introduce autogpt and write a tutorial .
We follow the prompt, he needs to enter Enter to continue, we will enter in the input box later. Then it asks what autogpt needs to do, enter our purpose here. It is recommended that you input English, because autogpt will be searched through Google in many cases, and it is not very friendly to Chinese support.
我们输入目的后可以看到，新版本的 autogpt 已经可以自己设定自己的角色了，甚至连上一版本需要我们输入的目标步骤也可以自己定了。他似乎又进化了：
AUTOGPTGPT has been created with the following details:
Name: AUTOGPTGPT
Role: an AI writing assistant that helps users create high-quality content by providing suggestions, generating ideas, and improving the overall writing process.
Goals:
- Assist users in creating a 1000-word article introducing Autogpt and providing a tutorial on how to use it.
- Provide suggestions for improving the article's structure, tone, and clarity.
- Generate ideas for the article's content and help users expand on them.
- Offer guidance on how to effectively use Autogpt to enhance the writing process.
- Ensure that the final article is well-written, informative, and engaging.
复制代码大家可以自行翻译下，autogpt 给自己设定了角色并制定了每一步的目标。可以看到还是很合理的，至少比我自己输入的要好很多。接着 autogpt 开始列出它的想法：
THOUGHTS: To begin, let's create a file to store the article. We can then use the 'write_to_file' command to write the initial draft of the article. Once we have a draft, we can use the 'analyze_code' command to check for any grammatical errors or areas where we can improve the structure and clarity of the article. Additionally, we can use the 'google' command to research more information about Autogpt and gather ideas for the article's content.
复制代码THOUGHTS：
- 想先创建文件，然后执行
write_to_file命令编写文章的初稿 - 一旦有了草稿，就可以使用
analyze_code命令来检查是否有语法错误或可以改进文章结构和清晰度的地方 - 此外，可以使用
google命令来研究有关 autogpt 的更多信息
然后它根据想法开始进行推理，并列出计划：
REASONING: Creating a file to store the article will allow us to easily edit and update the article as needed. Writing the initial draft will give us a starting point to work from, and using the 'analyze_code' command will help us improve the quality of the article. Researching more information about Autogpt will help us create a more informative and engaging article.
PLAN:
- Create a file to store the article
- Use the 'write_to_file' command to write the initial draft of the article
- Use the 'analyze_code' command to check for any grammatical errors or areas where we can improve the structure and clarity of the article
- Use the 'google' command to research more information about Autogpt and gather ideas for the article's content
复制代码PLAN:
- 创建一个文件来存储文章
- 使用
write_to_file命令编写文章的初稿 - 使用
analyze_code命令检查是否有语法错误或可以改进文章结构和清晰度的地方 - 使用
google命令研究有关 autogpt 的更多信息，并收集文章内容
接着它开始对自己的计划进行很人性化的思考和批判：
CRITICISM: We should be careful not to rely too heavily on the 'analyze_code' command and instead use our own judgment to determine the best way to improve the article. Additionally, we should make sure to properly cite any sources we use for the article.
复制代码这一系列的准备工作做完后，autogpt 开始干活了。它列出它下一步的行动：使用write_to_file 命令创建一个文件，文件名是autogpt_article.txt，然后内容是它生成的初稿。
NEXT ACTION: COMMAND = write_to_file ARGUMENTS = {'filename': '/content/Auto-GPT/autogpt/auto_gpt_workspace/autogpt_article.txt', 'text': "Autogpt is an AI writing assistant that helps users create high-quality content by providing suggestions, generating ideas, and improving the overall writing process. In this article, we will introduce Autogpt and provide a tutorial on how to use it. We will also provide suggestions for improving the article's structure, tone, and clarity, as well as generate ideas for the article's content. Finally, we will offer guidance on how to effectively use Autogpt to enhance the writing process. By the end of this article, you will have a better understanding of how Autogpt works and how it can help you create better content."}
Enter 'y' to authorise command, 'y -N' to run N continuous commands, 's' to run self-feedback commands'n' to exit program, or enter feedback for …
复制代码然后它开始征求我们的意见，问我们是否执行write_to_file，我们可以输入以下信息去批准它还是否决它：
y：允许执行命令y -N：允许执行连续 N 个命令n：退出程序s：运行自反馈命令- 其他反馈指令
这里我直接输入y -10，autogpt 创建了文件后，自己使用analyze_code命令对初稿进行检查，然后通过google 命令去搜索更多关于 autogpt 的信息。
Because the key I demonstrate here does not have the api permission of 4.0, so it is not unexpected that the 4096 tokens limit of 3.5 is exceeded. It is a pity that I did not get a satisfactory article.
Summarize
In this article, I saw that autogpt has been updated to a new version, and then I will introduce you to it, and then demonstrate its functions.
Although autogpt has brought us many surprises, in actual use, more of them are unsatisfactory. On the whole, I think autogpt can't free our hands at present. Everyone can experience it for themselves.
Many implementations in this project are worth learning from. You can learn more about it by reading its source code.
References
[1]Auto-GPT: github.com/Significant…
[2]Google-colab: colab.research.google.com/
This article is reproduced from the personal public account: AI does not understand the essence of life