AutoGPT Agents Want to Automate ChatGPT, Overrun the Internet

Now that the ChatGPT from OpenAI has caught the fancy of the masses, clever AI types have started pondering about how the chatting portion of generative AI service could be automated away, and have the AI just do stuff on your behalf.
Their early prototypes are rough, if not potentially dangerous, yet they offer enough promise that some enthusiasts are breathlessly calling them a primitive form of artificial general intelligence (AGI).
Welcome to the rapidly emerging world of AutoGPT.
On April 6, the top three trending repos on GitHub were all self-prompting projects of this sort: Microsoft’s JARVIS, Significant Gravitas‘ Auto-GPT, and, arguably the progenitor of them all, Mini Yohei‘s BabyAGI,
“Seems like a new groundbreaking update comes out every hour,” Tweeted AI observer Barsee, helpfully storming tweets of examples.
But What Is AutoGPT?
AutoGPTs “are designed to automate GPT-4 tasks, enabling the creation of agents that complete tasks for you without any intervention,” explained Nathan Lands, founder of generative AI-focused Lore.com, via Tweet.
A GPT call is a single instruction on a computer, and as such, a series of them could “be strung together into programs,” explained computer scientist Andrej Karpathy, who works on JARVIS.
BabyAGI, according to its GitHub page, is a Python script that serves as an “AI-powered task management system.” With a pre-defined objective, it creates and then executes tasks, with the result of each task determining the following one.
The script runs an infinite loop with the following steps:
- Pulls the first task from the task list.
- Sends the task to the execution agent, which uses OpenAI’s API to complete the task based on the context.
- Enriches the result and stores it in Pinecone.
- Creates new tasks and reprioritizes the task list based on the objective and the result of the previous task.
They are not hard to set up, according to Sully Omarr, who described how to create and run one using Auto-GPT, “in less than 30 minutes” no less!
The trick is to set up the right goals. Make them too lofty and the script will get hung up on the planning stages, Omarr explained. It also seems to work a lot better with GPT-4, rather than GPT-3.5, which doesn’t have the sufficient wordiness to make an AutoGPT script truly hum along.
The documentation in all the major AutoGPTs strongly discourages anyone from running these sorts of agents in “continuous mode,” because, well, they all say, who knows what would happen? (The safer method is to have the script stop after each action, and not proceed into the user hits return).
Such danger hasn’t stopped people from using them in all sorts of ways, though. Lands points out how they already have been used in myriad ways.
Someone created one for sales prospecting, in which the agent can determine the top 50 companies, identify the human resource directors of these companies, send them emails and, for those who responded, book meetings on Calendly, all on behalf of the sales agent. For news podcasters, another agent can read about recent events in the news and prepare a podcast outline. For coders, there is a Test-Driven Development agent that creates software features, just by iterating from a set of tests provided by the dev.