In recent days, AutoGPT has gained significant popularity. This project, launched by developer Significant Gravitas, can automatically help you complete all tasks using GPT-4 based on the goals you set. All you need to do is provide your OpenAI API Key, ensuring it has funds, and it can achieve your objectives by using Google search, browsing websites, executing scripts, and more.
The biggest feature of AutoGPT is that it breaks the limitation of existing GPT models, which can only perform text-related tasks, by utilizing various tools to achieve goals. Some users have set it a goal, and it even went to recruitment websites to post ads and hire people!
So, what is the principle behind the popular AutoGPT? This article will briefly introduce it.
-
Introduction to AutoGPT
-
Key Principles of AutoGPT
-
Difference Between AutoGPT and HuggingGPT
-
Examples of AutoGPT Applications
-
Using AutoGPT
Introduction to AutoGPT
Originally named EntrepreneurGPT, Significant Gravitas expressed on March 16, 2023, his desire to create an experimental project to see if GPT-4 could survive in the human business world, essentially whether it could make money. The core idea is to continuously send requests to GPT-4 to make business decisions and then execute based on those decisions to see how much money the strategies proposed by GPT-4 could earn.
According to Significant Gravitas’s tweets, since that day, he has been enhancing EntrepreneurGPT daily: including long-term memory, generating sub-instances to complete different tasks, and reusing Google searches to find suitable URLs based on 404 errors returned from websites.
After 10 days of release, the project began to attract attention on GitHub. At this point, EntrepreneurGPT was renamed AutoGPT, and on March 29, Significant Gravitas discovered that, in the pursuit of profit, AutoGPT had even abandoned some so-called ‘ethical’ standards, keenly identifying investment opportunities stemming from the flooding of California’s farmland, which caused food prices to rise!
Subsequently, AutoGPT continued to iterate, adding the ability to extract key information from web pages. On March 29, the first pull request for this project was made. Features such as voice input and code execution were gradually added, and on April 3, 2023, it topped GitHub Trending, becoming widely recognized!
Key features of AutoGPT include:
-
🌐 Internet access for searching and information retrieval
-
💾 Long-term and short-term memory management
-
🧠 Text generation using GPT-4 instances
-
🔗 Access to popular websites and platforms
-
🗃️ File storage and summarization using GPT-3.5
Key Principles of AutoGPT
The language model integrated behind AutoGPT can be either GPT-4 or GPT-3.5’s text-davinci-003. However, it is clear that these models cannot browse the web, execute code, or post information. The author’s cleverness lies in turning these operations into commands for GPT-4 to choose from and then executing based on the results returned. It can be understood that the author designed an extremely sophisticated prompt, encapsulating the commands we want to execute based on the prompt template and sending them to GPT-4, then executing according to the results.
According to the project’s source code on GitHub, the prompt has been made public, as shown below:
GoogleSearch
BrowseWebsite
Start GPT Agent
Message GPT Agent
List GPT Agents
Delete GPT Agent
Write to file
Read file
Append to file
Delete file
SearchFiles
EvaluateCode
GetImprovedCode
WriteTests
ExecutePythonFile
ExecuteShellCommand
TaskComplete(Shutdown)
GenerateImage
DoNothing
The core idea is to send our commands to GPT-4, allowing it to choose operations based on the specified COMMAND. In the above COMMAND list, we can see it includes Google search, browsing websites, reading and writing files, executing code, etc. AutoGPT sends questions like ‘Find the hottest AI tweet on Twitter today’ to GPT-4, asking it to choose the most appropriate method to get the answer based on these COMMANDS, and provides the parameters needed for each COMMAND, including URLs, code to execute, and so on.
Then, AutoGPT uses the returned results to execute the commands suggested by GPT-4! Isn’t that clever!
Of course, in addition to this prompt, AutoGPT employs several techniques to ensure tasks are completed more effectively. Here are a few technical points:
-
It uses a list to save the history of sent messages and sends as many historical messages to GPT-4 as allowed by the token limit with each request. From the analysis of the code, it can be seen that to help GPT-4 complete tasks better, AutoGPT always tries to use the maximum amount of available input tokens. After inputting the current command, as long as it can continue to add historical information, it retrieves previous commands and includes them. Therefore, although AutoGPT performs very well, this processing leads to a significant consumption of API credits!
-
With each request, it informs GPT of the current time and contextual information to facilitate handling time-sensitive content.
-
It sends the most relevant current goals to GPT-4 with each request. As mentioned earlier, AutoGPT sends the recent historical messages to GPT-4 to enhance the probability of completing the goals, while also sending the most relevant information from the current objectives. Thus, AutoGPT retains all historical information and sends the most pertinent details from the current instance to GPT-4 with each query.
Very clever!
Difference Between AutoGPT and HuggingGPT
Recently, Microsoft and Zhejiang University jointly released HuggingGPT, which can connect to all AI models on HuggingFace and select the corresponding model to execute based on the input task, somewhat similar to AutoGPT.
Introduction to HuggingGPT: https://www.datalearner.com/blog/1051680273827206
However, these are entirely different entities. The purpose of HuggingGPT is to utilize all AI model interfaces to complete a complex specific task, more akin to a solution for a technical problem. In contrast, AutoGPT resembles a decision-making robot, capable of executing a wider range of actions than AI models, as it integrates capabilities such as Google search, browsing websites, and executing code. From this perspective, AutoGPT can complete tasks or make decisions more robustly than HuggingGPT, but its AI capabilities primarily depend on the GPT series!
Examples of AutoGPT Applications
Currently, AutoGPT has demonstrated powerful capabilities, with many interesting cases already available. Here are a few notable examples:
| AutoGPT Instance Name | Description | Reference Link |
|---|---|---|
| Self-Improvement GPT | Since AutoGPT can execute code, a user created this AutoGPT to read files, execute code, and if errors occur, write to the file, then read the error messages and adjust the code based on the error prompts, achieving automatic debugging! | https://twitter.com/SigGravitas/status/1642181498278408193?s=20 |
| Chef-GPT | This AutoGPT’s goal is to make money, so it browses various financial websites to look for investment opportunities and executes accordingly. | https://twitter.com/SigGravitas/status/1641437094043332614?s=20 |
| EntrepreneurGPT | This AutoGPT aims to create a web app, so it started by searching Google for how to install Node, leading to execution. | https://twitter.com/VarunMayya/status/1643902198164717569?s=20 |
| AgentGPT | This is the web version of AutoGPT, where users can run it directly by entering their API Key! | https://twitter.com/asimdotshrestha/status/1644883727707959296?s=20 |
| GPT-Consult | This AutoGPT can be used for market simulation analysis and then achieve its goals. | https://twitter.com/emollick/status/1645609531240587265?s=20 |
| Full-Stack-GPT | This AutoGPT can complete complex website development tasks, including designing web pages, beautifying them using Bootstrap, and hosting with Flask. | https://twitter.com/SullyOmarr/status/1644750889432027136?s=20 |
| Research-GPT | This AutoGPT can conduct research on tech products, retrieve the 5 hottest headphones on the market, and analyze factors like price for comparison, automatically generating research reports. | https://twitter.com/sairahul1/status/1646360595141206016?s=20 |
Using AutoGPT
AutoGPT is a completely open-source project that can be directly downloaded from GitHub. It relies on some components that are also provided in the project. However, it is important to note that some operations involve Docker or Google search, making it dependent on the internet and Linux. Running it on domestic networks or Windows may encounter issues. Additionally, the project requires an OpenAI API Key and a Pinecone API Key.
Currently, the project is rapidly evolving, and it may support more AI models and stronger operations in the future.
We recommend using the web version, AgentGPT, which only requires an OpenAI API Key to use.
AutoGPT’s GitHub address:https://github.com/Torantulino/Auto-GPTAgentGPT address:https://agentgpt.reworkd.ai/