The capabilities of AI tools are progressing rapidly, with Google, Microsoft, OpenAI, and many others racing to stay ahead of the competition. It feels like advances and apps are arriving on a weekly basis, with the bar constantly being raised in terms of what AI can do for us.
Auto-GPT is the latest evidence for this: It leverages the power of ChatGPT to create an autonomous AI assistant, capable of taking on tasks and projects on its own and working through multiple steps in a job without you having to prompt it every time. In other words, it does a lot of the hard work for you, without you having to come up with your own follow-up responses or ideas.
Auto-GPT can be run locally on your computer.
Auto-GPT via David Nield
Think about everything you can do with ChatGPT, then imagine rolling that into a system that can supply its own feedback and make its own choices. Take coding for example: ChatGPT can output blocks of code based on your specifications, but Auto-GPT can manage an entire software project for you.
We asked Auto-GPT to research the differences between iOS and Android, and it split the research into sub-tasks: looking at the user interfaces, the available apps, the security and privacy features, and so on. These sub-tasks were run and analyzed separately, and the agent decided to save its findings to a text file in each category so they’d be easier to refer to later.
Another way to use Auto-GPT is to specify a particular role for it, whether that’s as a website builder or a podcast researcher. It will then act accordingly and keep you up to date on its progress. There is something of a learning curve with the software, but it pays to experiment with what it’s capable of.
Auto-GPT is able to access current information from the web, and it can remember its previous actions. The code for the tool has been released for free, though you need to pay for access to the OpenAI API. If you want to get at the latest ChatGPT-4 version of the large language model then you also need to have a paid OpenAI account (otherwise you’ll be on ChatGPT-3.5).
Auto-GPT is powered by ChatGPT from OpenAI.
Auto-GPT via David Nield
Because Auto-GPT is working on its own for much of the time, the chances of it going off the rails are increased: We know that ChatGPT can still produce inaccuracies and make mistakes, and with Auto-GPT those problems can be compounded, because you’re not always checking that its working every step of the way. It’s nowhere near the stage yet where you can trust it with anything of real importance or consequence.
However, as a developing experiment in AI that shows the possibilities of the technology both right now and in the near future, it’s a fascinating tool to play around with—and it’s a tool that’s improving all the time. Have a quick check online to see what other users are up to, attempting everything from ordering pizza to running a marketing campaign.
Installing and Using Auto-GPT or AgentGPT
If you know how to install GitHub projects on your computer, then you can go ahead and download Auto-GPT. If you’re in need of step-by-step instructions, there are numerous installation approaches you can take across different operating systems and using different tools (such as Docker and Python). We have room here to cover only one installation route, using Python on Windows, but there’s plenty of help online if you need another option. Check out the official Auto-GPT documentation, for example.
When it comes to the Python on Windows approach, you first need to install both Git and Python on your PC. During the setup process, be sure to enable the Add python.exe to PATH option during setup. The process is straightforward, but if you get into any difficulties, there’s plenty of support documentation on the Python website.
Auto-GPT will tell you what it’s doing as it works.
Auto-GPT via David Nield
Next, you need to download Auto-GPT from GitHub: Follow the link to the latest release and download the zip file you find there. When you’ve unzipped the archive, duplicate the file called .env.template, rename it as .env, and open it in a text editor. Look for the line “OPENAI_API_KEY=your-openai-api-key” and replace “your-openai-api-key” with an actual API key from your OpenAI account.
You can get an API key from here. You only get a few free API calls with ChatGPT, so you’ll want to set up a payment method—you can see the latest pricing at this page (at the time of writing, it’s a few cents per thousand tokens or bits of information processed). When you set up your payment method, you can also configure payment limits to make sure AutoGPT doesn’t cost you more than you’re comfortable with.
After you’ve saved and closed your .env file, you can launch Auto-GPT by right-clicking on its folder and choosing Open in Terminal. Type “pip install -r requirements.txt” and hit Enter to install the necessary libraries, then when that’s finished, type “python -m autogpt” and press Enter to launch Auto-GPT itself. You’re then free to start getting Auto-GPT working anonymously for you: Simply tell it what you want it to do, and it will get to work.
Its thinking will be shown on screen, and you’ll be asked to confirm each action in the process Auto-GPT works out—if you’d rather approve several actions in a row, type “y-x”, where x is the number of actions you’re happy for the bot to get on with without your approval (remember all of these ChatGPT requests are costing you a small amount of money). If you need to end a task before it’s completed, use Ctrl+C.
AgentGPT is an AI assistant available on the web.
Auto-GPT via David Nield
If this seems like a lot of work and you want something simpler to play around with, AgentGPT is a similar project based on Auto-GPT that works in the browser, no installation required. It’s not quite as advanced as Auto-GPT, but it is more straightforward for beginners, and it gives you an idea of what’s possible with these autonomous AI agents.
As with Auto-GPT, you need to get an API key for ChatGPT for AgentGPT to run properly (and again you’ll be charged based on usage). You can plug this API key into the tool via Settings on the left. Give your agent a name and a goal using the input on screen, then choose Deploy Agent to set it off—tasks are shown on the right as they are generated and completed.
