This article was written by Rachel Metz and Saritha Rai. It appeared first on the Bloomberg Terminal.
OpenAI’s ChatGPT is simple enough to use: You type in a request and get a response that may be helpful and sounds convincingly human. Yet it has severe limitations as a practical tool. It still generally doesn’t have access to the internet (and, in fact, its training data stops sometime in 2021), isn’t designed to remember queries from conversation to conversation, isn’t all that factually reliable, and can only generate a single answer for each prompt.
Where to next? Researchers and AI enthusiasts seeking the next phase of the current chatbot hype cycle have begun to coalesce around another idea: so-called AI agents, software that connects to large language models such as the one that powers ChatGPT (and perhaps various internet services or even payment methods, too) and, after being assigned a goal by a human, comes up with a series of tasks and carries them out on its own in an effort to reach that objective.
This hints at a future where AI agents could be acting autonomously across the internet, carrying out a wide range of tasks however they see fit. Those who are most enthusiastic about AI agents see them as the closest thing yet to artificial general intelligence, or AGI, the long-sought goal of creating software that can learn and carry out any intellectual task a human can.
The most prominent work to emerge is Auto-GPT, an open-source project run by Toran Bruce Richards, a game developer. Auto-GPT users set goals for the software, which then begins executing steps to achieve them. By default, Auto-GPT asks for validation from its human user every step of the way, though it’s possible to grant it what amounts to blanket permission to keep moving ahead autonomously.
It’s become trendy among a certain brand of early adopter to post examples of the seemingly miraculous things Auto-GPT and other AI agents are already doing. One user got an AI agent to conduct product research and summarize a list of the best headphones on the market. Another created a fake shoe company and got market research for waterproof shoes. It listed the top five competitors and produced a report on their pros and cons. The chief executive officer of HyperWrite, which offers an AI writing assistant plugin for the Chrome browser, posted a demo on Twitter that purports to show how its forthcoming Personal Assistant can order a pizza online from Domino’s.
Unlike ChatGPT, Auto-GPT isn’t made for the average user. It requires some coding to set up, as well as access to OpenAI’s application programming interfaces, which allow developers to build software programs that incorporate its AI models and incur a charge each time those models are asked to do something. HustleGPT, a group on the messaging platform Discord, includes early-stage startup founders who are using the tools for “everything from designing a logo to setting up a website to marketing,” says Dave Craige, an entrepreneur who started the group.
As with many tech demos, it’s probably best to take any claims about AI agents’ accomplishments with a grain of salt—and to fact-check any information the agents produce. Bloomberg Businessweek asked AgentGPT to concoct a plan to turn $100 in Vanguard’s Total Stock Market Index Fund into $1 million in 10 years. The program assigned itself a list of tasks related to carrying out this investment goal, which included researching the historical performance of the fund over the past 10 years to estimate potential returns. It then presented several multistep plans to get there.
The plans themselves, though, were severely lacking. AgentGPT suggested that someone investing $500 per month could—if she followed several strategies to reduce fees and earned an average annual return of 7%—end up with precisely $1,031,906 after 10 years. The online investment calculator at calculator.net determined that this scenario would produce the more modest total of about $86,000 in the given timeframe.
Software such as Auto-GPT and BabyAGI, another AI agent, are “in their infancy,” says Matt Schlicht, CEO of Octane AI, which uses AI to deploy product-finder quizzes that e-commerce brands can use to help customers decide which products to buy. “So if you look at it like a standalone product, it’s still kind of a toy in a way.” That said, in May his startup is launching its own agent, Insights AI, to analyze customer reviews and help companies with strategy and content creation.
As people like Schlicht decide AI agents can be more than toys and begin testing them in real-world situations, the need for policies governing their use becomes more urgent, according to Henry Shevlin, an AI ethicist at the Leverhulme Centre for the Future of Intelligence at the University of Cambridge. “It’s unlikely that corporations will exercise the level of restraint that’s desirable for humanity,” he says.