An “experimental open-source attempt to make GPT-4 fully autonomous”, Auto-GPT attempts to minimise human intervention, unleashing the full power of conversational AI. And some early results have been truly impressive.
Here’s what Auto-GPT is and how to try it for yourself.
What is Auto-GPT?
Anybody who has used ChatGPT will know the process requires a back-and-forth to get good results if you ask it something remotely complicated. After your initial prompt, ChatGPT will spit something out, and you then respond with a new prompt until the result is refined into what you’re looking for.
Auto-GPT attempts to automate this process and cut out the middle man — you.
All you need to do is feed Auto-GPT an initial prompt and — in theory — Auto-GPT will generate a task list of what it believes is required to reach that goal and feed those instructions to ChatGPT on your behalf.
What can Auto-GPT do that ChatGPT can’t?
It’s not that Auto-GPT can do anything that someone with enough time couldn’t with ChatGPT. It’s more that, for more complex requests, ChatGPT users need to keep adding to their original prompt with corrections and follow-up instructions.
The idea of Auto-GPT is that you can leave all of this to the computer to automate and carry on with your day, just enjoy the finished product at the end of the process.
It’s also significant that Auto-GPT may think of requests or phrasing that you would never consider yourself, potentially producing better results than a human alone.
What can you use Auto-GPT for?
Search Twitter for Auto-GPT and you’ll find plenty of exciting examples of how it’s being used. Like this one, where Auto-GPT is prompted to do four things:
Do market research for different headphones on the market today.
Get the top five headphones, and list their pros and cons.
Include the price for each one and save the analysis.
Once you are done, terminate.
Just when I thought AI couldn't get any more impressive, AutoGPT and AI Agents blows my mind 🤯
Here, AI agent performs product research and creates a summary on the top headphones.
This is insane! And the craziest part is that it's powered by GPT-4. pic.twitter.com/47AwQZfaIC
— Rahul (@sairahul1) April 13, 2023
The AI almost immediately begins to search the internet for headphone reviews and specs, and then outputs a plain-text file ranking the headphones, along with prices and advantages of each.
In another example, someone else pretended to be a shoe company and asked Auto-GPT to research the top-five competitors in the waterproof market.
It continued to analyze the various sites, with a combination of googling, updating its queries, until it was happy with the results.
Here's an example of when it thought "critically".
It knew that some reviews could be biased to fake, so it had to validate the reviewer. pic.twitter.com/UyfiGpx1aP
— Sully (@SullyOmarr) April 9, 2023
Impressively, the software had the reasoning to note that not all reviews are legitimate and weighted them accordingly.
This is just scratching the surface, and others have used Auto-GPT for everything from debugging code and writing emails to compiling business plans for a new venture.
What are the drawbacks of Auto-GPT?
The problems with Auto-GPT are similar to those with ChatGPT — it’s prone to errors. Without a human providing follow-ups, you run the risk of compounding these errors by outsourcing the sense checking to something without sense.
There are other issues, too. Auto-GPT often doesn’t remember how to replicate a completed task for later use, and it struggles with more complex tasks, sometimes failing to grasp how connected goals overlap with each other in a way that might seem obvious to a human.
As Nvidia AI scientist Jim Fan said on Twitter, it’s a “fun experiment… but nothing more”. He encourages people not to get carried away and to question the “heavily cherry-picked” success stories.
In my experiments, AutoGPT can solve certain simple & well-defined knowledge tasks well, but is unreliable *most of the time* for harder tasks that are truly useful. I also worry a lot whenever I give it python execution and disk access.
— Jim Fan (@DrJimFan) April 16, 2023
“In my experiments, Auto-GPT can solve certain simple & well-defined knowledge tasks well, but is unreliable *most of the time* for harder tasks that are truly useful,” he added.
Is Auto-GPT easy to use?
While you don’t necessarily need knowledge of code to get it up and running, if terms like “Python”, “API key”, and “command line arguments” bring you out in a cold sweat, you may want to try browser-based demos, like AgentGPT and GodMode instead.
That said, if you want to have full control and install it yourself, there are extensive instructions available on the GitHub site.