Get started with AutoGPT, the stunning GPT-4 experiment that controls itself

AutoGPT is a GitHub repository that has created a buzz in the tech community, gaining more than 100,000 stars in just a few days. That’s the equivalent of tens of millions of likes or shares on social media platforms. At first glance, it looks like a simple Python script running in a terminal window, powered by GPT-3.5 or GPT-4 and DALLE-2. It takes plain text input and produces plain text output, which makes it seem nerdy and unimpressive. But it’s much more than that.

The fuck is AutoGPT?

AutoGPT is like a to-do list that extends and executes itself. It creates ways to solve given tasks and goes beyond what you ask it to do. It tries to solve problems and suggests alternative approaches by making assumptions about missing information from your prompt. For example, if you ask it to mail pictures of Bánh mì’s to you (with your name and some context), it will probably start searching the web for your email address. It can parse the web through terminal, or even open the visible web through a Selenium browser instance. It can open your blog, your LinkedIn profile, or your employer-related websites and even suggest sending an email to someone who might know you to ask them for your email address. Once it has your email, it can generate pictures of delicious Bánh mì’s using DALLE-2 or search for some pictures on Google and send them to you along with a nice message. This is just a simple and dumb example that barely scratches the surface of what AutoGPT can do, but it gives you an idea.

So what’s the brainfuck behind AutoGPT? Well, until now, it has been clear how computers work: they process instructions, nothing more, nothing less. The instructions must be crisp and clear, leaving no questions unanswered. With AutoGPT, you simply state what you want to accomplish and what the end result should be. How the computer gets there is up to it. That’s a big change. Instead of asking questions back and waiting for your response, AutoGPT prompts itself to fulfill your request.

The hype is real

AutoGPT is one of the most fascinating pieces of software I have ever seen. It clearly shows where the AI journey is going. These systems allow you to program without programming knowledge and could soon become our personal assistants. You could say something like, »Hey, I need a case for my Google Pixel 6 Pro. Please order me one that is black, has good reviews, and costs less than 100,000 VNĐ.« Or, »Please make me an appointment with a Thai massage salon in HCMC District 1, preferably on Saturday afternoon, but Monday morning is possible if necessary.« Sure, it has similarities to Siri, Alexa, Google Assistant, and so on, but these tools feel like limited medieval technology compared to what AutoGPT could soon become. And it’s under your control… at least sort of.

AutoGPT is an impressive tool that has the potential to navigate the web, understand user interfaces and images, solve problems, and even communicate with other systems and humans. It can submit forms, perform checkout, download and run software, create text and images, navigate and interact autonomously, it can even create a web app project, install npm packages and build a small website – and much more. Even more with the help of plugins. It can do almost anything you can do on the Internet, and it’s clear that these capabilities will soon be further autonomous. Whether this development is good, bad, or dangerous is not the issue; the fact remains that these systems are coming, and people and businesses will use them to save time and money. And companies WILL use it, even if the results are not brilliant; but passable results are already enough to automate entire jobs and put economic pressure on large numbers of people.

99 little bugs in the code

However, please keep in mind that AutoGPT is not fully functional yet. It’s still an experiment, and it has bugs and clunkiness that need to be worked out. In some cases it can even be dangerous, running system tasks and doing other silly things that can damage your machine. Therefore, I strongly advise you not to run AutoGPT on a production system, but rather in a virtual machine without any important data or sensitive systems.

Another problem with AutoGPT is that it’s not very smart yet. It often fails on longer quests, and can get stuck in a loop or start over on some already solved tasks. But AutoGPT always communicates clearly what it is doing, so you can always answer its questions or adjust its way to solve your problem. I’m confident that in a few months or even just weeks AutoGPT will work better and we will see the first serious attempts to use it in production out there. A lot will change and improve, leading to exponential development. The future is now.

How to run AutoGPT on your machine

Requirements

If you want to try AutoGPT and play around with it, you need four things:

1. A computer running Linux, MacOS or Windows (nobrainer).
2. A local installation of Python. You can load and install it from here: https://www.python.org/downloads/.
3. An OpenAI API key. For this you need an OpenAI account, either a free or a paid one. Free accounts are limited to GPT-3.5 and have only a handful of dollars available for API usage. Paid accounts have unlimited access to both GPT-3.5 and GPT-4. You can generate and copy your free API key here: https://platform.openai.com/account/api-keys.
4. And the latest version of AutoGPT. You can download and unpack AutoGPT from its GitHub repo: https://github.com/Significant-Gravitas/Auto-GPT.

Setup and start

Follow these simple steps to set up and run a basic instance of AutoGPT locally. This was successfully tested with AutoGPT v0.2.2 from March 20, 2023, so some details may have changed in the meantime. Always check the official README for the latest information and further instructions, such as how to run AutoGPT on an Azure instance, in a Docker container, and how to use custom plugins to greatly extend the capabilities of AutoGPT.

1. Unzip the loaded AutoGPT file to a folder of your choice.
2. Open a terminal/command line. On Windows, you may need administrator privileges (press the Start button, type cmd, then right-click and select run as administrator).
3. Navigate to the folder containing the unzipped AutoGPT script.
4. Run pip install -r requirements.txt to load and install all dependencies AutoGPT needs to work.
5. The installation will create a .env.template file in the AutoGPT folder. Open it with a text editor and enter your OpenAI API key after OPENAI_API_KEY= by replacing your-openai-api-key with the actual key.
6. Save this file as .env, without the .template part.
7. Run AutoGPT. There are two ways to do this, depending on whether you have a free or a paid account:
   1. If you have a free account, type ./run.sh -m autogpt –gpt3only.
   2. If you have a paid account, type ./run.sh -m autogpt.
   3. Hint: If you are using a Windows machine, type .\run.bat instead of ./run.sh.
8. You will now be asked to name your AI to personalize it and describe its role. Be creative. I named mine GLaDOS_Indev, which is pretty uncreative though.
9. And that’s it. You can now play around with it. Give it some tasks or send it on a quest. By default, you have to confirm each step AutoGPT wants to do with y, but you can also type y -20, for example, and then the next 20 steps will run autonomously. However, this is risky and considered a bad idea.

Ready for action

Enjoy a taste of the near future with AutoGPT. Try new things and think outside the box. Browse YouTube, Twitter, or tech blogs for inspiration and ideas about what you can achieve with it. And finally, share your findings, experiences, and failures with others, because it’s fun after all.

Hero image: MATRIX CODE by pixelz.cc.