1 Introduction

ChatGPT has become all the rage at the moment. As a leader in natural language processing models, ChatGPT's advantage lies in its ability to generate smooth and coherent conversations, while also being able to understand the context and answer according to the context. It can be quickly customized for different application scenarios. For example, in customer service, education, entertainment and other fields, ChatGPT can be used as an intelligent assistant to provide users with convenient service and entertainment experience.

From GPT-3 to GPT-4, we can see that ChatGPT has made remarkable progress in continuous optimization training in terms of advanced reasoning, input setting, fine-tuning behavior and understanding longer contextual information.

But for this interactive GPT model, we hope to see that it can be like the artificial intelligence in the movie, have self-awareness, and can learn and evolve by itself. Without frequent guidance and interaction, we only need to simply propose a goal, and AI can complete all execution planning and logic processing autonomously, and continue to self-justify and optimize, and finally complete the goal we proposed.

Well today it comes, it is the evolution version of ChatGPT: AutoGPT .

2. What is AutoGPT?

AutoGPT is an experimental open source application that leverages OpenAI's GPT-4 language model to create fully autonomous and customizable AI agents. It was released on March 30, 2023 by Toran Bruce Richards.

Toran is a game developer and founded a game company called Significant Gravitas.

In just over a month, AutoGPT's Git project has gained more than 120,000 stars. As one of the first examples of GPT-4 fully autonomous operation, AutoGPT has broken through the boundaries of AI possibilities.

AutoGPT is unique compared to other AI tools in that it operates independently, which means you no longer need to manipulate the model to suit your needs. Instead, you just need to write down your goals, and the AI will do the rest for you. Thus, AutoGPT fundamentally changes the way AI interacts with humans, who no longer need to take an active role, while still maintaining the same or better quality of results than other AI applications such as ChatGPT.

3. How does AutoGPT work?

AutoGPT works based on an autonomous AI mechanism, where the AI system creates different AI agents to fulfill specific tasks, including:

Mission Creation Agent: The first AI agent to interact with a mission creation agent when you enter a goal on AutoGPT. Based on your goals, it will create a task list with steps to achieve those goals and send it to the priority agent.
Task Prioritization Agent: After receiving a list of tasks, the Priority AI Agent ensures that the order is correct and logical before sending it to the Execution Agent.
Task Execution Agent: After prioritization is done, the execution agent will complete the tasks one by one. This involves utilizing GPT-4, the Internet, and other sources to obtain results.

The aforementioned agents communicate with each other. So when the execution agent completes all tasks and the result is not satisfactory, it can communicate with the task creation agent to create a new task list. An iterative loop between the three agents until all user-defined goals are accomplished.

The behavior of the AI agent is also displayed on the user interface, which divides them into four groups: thinking, reasoning, planning, and judging .

THOUGHTS : The AI agent shares its thoughts about the target.
REASONING : The AI agent reasoning about how to develop and implement its ideas.
Plan (PLAN) : The AI agent enumerates the plan of the task to be completed through analysis.
CRITICISM : AI judges itself, corrects mistakes and overcomes any limiting issues.

By sharing this computational process, AutoGPT can perform trial-and-error demonstrations and perform targeted optimizations that can overcome all problems encountered without any user intervention.

3. Comparison between AutoGPT and ChatGPT?

Although the underlying LLM model is the same, there are quite a few differences between AutoGPT and ChatGPT.

3.1 Real-time insights

The latest GPT-4 model used by ChatGPT is trained on the same data as GPT-3.5, which is only available until September 2021, you can't get real-time data information using ChatGPT because you can't access websites and online platforms to get information and extract information.

In contrast, AutoGPT has access to the internet. Not only can it surf the web, but it can also verify that the source is legitimate. Furthermore, AutoGPT can access any platform to perform tasks. For example, if you ask an AI to research prospects for selling a product and send an outreach email, it will draft and send the email directly using your Gmail account.

3.2 Memory Management

The context window is very important for the language model to give accurate answers. But in LLM like GPT-4, the window has a limit of 4000 to 8000 tokens. Therefore, if the requirements exceed the limits, the model may not follow all instructions correctly, or it may deviate from tangent and provide unreliable output.

In contrast, AutoGPT is good at short-term and long-term memory management. By using database, local Cache and Redis for memory management, a large amount of contextual information or previous experience can be stored, allowing the AI model to make better decisions.

3.3 Image generation

AutoGPT is able to generate images because it can use a variety of image generation engines, and DALL-E is used by default. If you want to enable image generation for your AI agent, you need access to DALL-E's API. Despite being a multimodal input method, this feature is currently not available in ChatGPT-4.

3.4 Text-to-Speech

python -m autogpt --speakYou can enable text-to-speech on AutoGPT by typing on the command line . But you must enter commands every time you interact with AutoGPT. You can also add different voices to your speech by connecting AutoGPT to the versatile AI voice software Eleven Labs.

4. Limitations of AutoGPT

There is no doubt that autonomy adds a new dimension to AI systems. At the same time, we cannot ignore the limitations and risks of AutoGPT. Listed below are some key limitations that you must be aware of.

4.1 High cost

While the features are amazing, the usefulness of AutoGPT might disappoint you. Since AutoGPT uses the expensive GPT-4 model, even small tasks can be costly to complete per task. This is mainly because AutoGPT uses GPT-4 multiple times in the steps of a specific task.

4.2 Often stuck in a loop

The most common problem users face while using AutoGPT is that it gets stuck in a loop. If this happens for more than a few minutes, it probably means you have to restart the process. This happens because AutoGPT relies on GPT-4 to properly define and decompose tasks. Thus, there is a trial-and-error problem if the underlying LLM returns insufficient results for AutoGPT to take any action.

4.3 Data Security

Since AutoGPT is fully authorized to run autonomously and access your system and the Internet, such as using your twitter account, logging in to github, using search engines, etc., your data may be leaked . AutoGPT does not have a safe proxy, so you have to be careful when using AutoGPT, you can't let the model continue to run without giving the correct instructions and safety guidelines. **

5. How to install AutoGPT?

Unlike other AI tools, AutoGPT does not have an easy registration process to access its platform and features. Before you can start using AutoGPT, you have to download various software to meet the requirements. The following are the detailed steps required:

Step One: Download the Prerequisites

First of all, you need to have a Git account, and you need to install Python 3.1.0 or later. In addition, you must be able to use common shell commands proficiently or have a Docker container for project startup and configuration.

Step Two: Set Up Your OpenAI API Key

If you haven't, please create an OpenAI account (of course, if you want to create an account in China is not a simple matter, you can refer to other articles on the Internet to apply for an account). After opening the OpenAI account, open _USER - API keys_ and go to the API keys tab. You will see an option to create a key. Click on it, and copy the key.

Step Three: Clone the latest version of AutoGPT

(1) clone project

Open the command line tool git clone https://github.com/Torantulino/Auto-GPT.git to clone the project to the local through the command

(2) Execute the installation

After entering the directory through the command cd Auto-GPT && ls -al, you can see that there are many files, one of which is requirements.txt. In this file, you will see the modules needed to run AutoGPT.

To install these modules, you can use the command pip install -r requirements.txt to download and install.

(3) Modify the configuration

vim .env.template Configure the open-api-key through the command (modify and replace your-openai-api-key), execute after the configuration is complete mv .env.template .env to make the configuration take effect

Other related configurations can refer to the table as needed

LLM PROVIDER	You can configure OPENAI_API_KEY, whether to use AZURE
LLM MODEL SETTINGS	You can configure the token limit provided by openAI to avoid excessive call cost waste, the default is 4000-8000
LLM MODELS	LLM underlying language model, GPT-4 or gpt-3.5-turbo can be selected by default
MEMORY	Memory management, you can configure local, redis, PINECONE, MILVUS, etc.
IMAGE GENERATION PROVIDER	Image generation, image size and image generation engine can be configured: dalle, HUGGINGFACE, STABLE DIFFUSION WEBUI
AUDIO TO TEXT PROVIDER	Speech to text, HUGGINGFACE can be configured
GIT Provider for repository actions	github configuration, by configuring github api key to access and manage github
WEB BROWSING	Search engine management, you can configure different browsers: firefox, chrome, safari, search engine: google, etc. Authorize open api to access the Internet to obtain information and manage access depth
TTS PROVIDER	Text-to-speech, MAC OS, STREAMELEMENTS, ELEVENLABS can be configured for text-to-speech
TWITTER API	twitter account management, manage and configure your twitter account, and configure token to access the corresponding api
AUTO-GPT - GENERAL SETTINGS	Some default configurations of AutoGPT, such as storage directory, switch, user Agent, AI settings, etc.

(4) Start using

After completing the above configuration, the basic configuration of AutoGPT has been completed. At this time, you can python -m autogpt start your AutoGPT journey through commands!

As can be seen from the above figure, AutoGPT requires you to give the AI a name [Name], a role positioning [Role], and you can set goals for it [Goals] (up to 5 goals, if you only have one goal, you can directly Enter).

After you set your goal, AutoGPT will think independently and analyze your goal [THOUGHTS], after thinking, you will start to understand and reason about how to accomplish this goal [REASONING], and then start to disassemble it into a specific plan [PLAN], Finally, a judgment [CRITICISM] will be proposed to ensure that the AI agent corrects errors and makes correct decisions.

After completing the above behavior planning, AutoGPT will prompt it to make instructions and actions [NEXT ACTION], which contains specific execution commands [COMMAND] and parameters [ARGUMENTS]. Users can identify risky commands at this time to avoid In the event of unexpected risks such as data leakage, you can yeither nauthorize or reject the next command action of AutoGPT.

AutoGPT will go through the above steps for multiple cycles. Since AutoGPT can store context and historical experience, each time it will conduct more in-depth thinking based on the feedback results, formulate a better plan, and finally list the plans he wants to implement, and try again and again and supplement until you achieve your desired goal.

(5) Docker use

Of course, you can also use docker to run:

// 最简单的方式就是通过docker-compose
docker-compose build auto-gpt
docker-compose run --rm auto-gpt

// 使用docker命令构建
docker build -t auto-gpt .
docker run -it --env-file=.env -v $PWD:/app auto-gpt

You can pass additional parameters, for example, run as --gpt3onlyand --continuousmode:

// docker-compose
docker-compose run --rm auto-gpt --gpt3only --continuous

// docker 
docker run -it --env-file=.env -v $PWD:/app --rm auto-gpt --gpt3only --continuous

6. Summary

Compared with traditional text generation technology, we found that the evolution of AutoGPT’s ability is astonishing. It can automatically disassemble it into the tasks it needs to perform by analyzing your goals, and based on existing experience and decisions in the process of execution Continuously optimize, improve and summarize. At the same time, AutoGPT has rich means of obtaining information. It can download and refine the required information through search engine search, github, web tools and other channels, and through local caching, voice conversion, image generation and other plug-in capabilities, Ultimately accomplish the goals you set. This form of self-awareness, self-iteration and renewal is very close to the artificial intelligence moss in the movie "The Wandering Earth"!

Finally, the question is, is this article automatically generated by AutoGPT?

Text: Leo

Recommended offline activities:

Time: June 10, 2023 (Saturday) 14:00-18:00 Theme: Dewu Technology Salon No. 18 - Wireless Technology No. 4 Venue: Dewu Hangzhou R&D Center, No. 77 Xueyuan Road, Xihu District, Hangzhou Training classroom on the 12th floor (Exit G of Wensan Road Station of Metro Line 10 & Line 19)

Highlights of the event: This wireless salon focuses on the latest technology trends and practices, and will bring you four exciting speech topics in Hangzhou/online, including: "Douyin Creation Tool-iOS Power Consumption Monitoring and Optimization" , "Dewu Privacy Compliance Platform Construction Practice", "Netease Cloud Music - Daily Guarantee Program Practice for Client Mass Traffic Activities", "Dewu Android Compilation and Optimization". I believe these topics will be helpful to your work and study, and we look forward to discussing these exciting technical contents with you!

Click to sign up: Wireless Technology Salon

This article belongs to Dewu technology original, source: Dewu technology official website

It is strictly forbidden to reprint without the permission of Dewu Technology, otherwise legal responsibility will be investigated according to law!

AI with self-awareness: AutoGPT ｜ Dewu Technology