Developers Experiment With Autonomous AI Agents
- By Paul Mah
- April 19, 2023
Even as the world is enraptured in recent months by instruction-following AI models such as ChatGPT, groups of developers are working to further stretch the limits of what is possible with the creation of autonomous AI agents.
The objective of an autonomous AI agent is to perform complex, multistep tasks with as little human intervention as possible. This is achieved by feeding various prompts to models such as GPT-4 or GPT3.5, and then iterating the output back into the same session – or creating new instances as necessary.
Autonomous AI agents
There is a handful of them at this point, with two of the most popular being Auto-GPT, created by Toran Bruce Richards, and BabyAGI, created by Yohei Nakajima.
A report on Ars Technica described Auto-GPT this way: “Auto-GPT takes the output from GPT-4 and feeds it back into itself with an improvised external memory so that it can further iterate on a task, correct mistakes, or suggest improvements. Ideally, such a script could serve as an AI assistant that could perform any digital task by itself.”
Users will need an OpenAI API key, and if they want the agent to perform a search, a Google API key. Under the hood, the agent architecture is designed to deal with issues such as memory, reflection, and planning, as noted by a separate report on Towards Data Science.
To see it in action, Ars Technica reporter Benj Edwards ran the Auto-GPT Python script on a Windows machine. According to Edwards, a prompt to purchase a pair of sports shoes, for instance, would see Auto-GPT generate a multistep plan and attempt to execute it.
It is unclear if these autonomous AI agents could do anything that cannot be done on ChatGPT itself. Indeed, the creators, when prompted, could not come up with substantive examples.
Part of this is due to the limitations of GPT-4 itself, which despite its impressive ability to transform and analyze text, is limited to a narrow range of interpretive intelligence.
Moreover, the propensity for current AI models to hallucinate untruths also makes it risky to rely on their outputs, which would only surely grow across multiple runs.
For instance, it is easy to imagine a request to run an analysis of competitors to generate fake firms which the autonomous AI agent will then attempt to research and compare – rendering the entire task meaningless.
For now, it is an area worth watching for future developments. You can check out a web-based version of Auto-GPT here.
Image credit: iStockphoto/Dmitrii_Guzhanin
Paul Mah
Paul Mah is the editor of DSAITrends, where he report on the latest developments in data science and AI. A former system administrator, programmer, and IT lecturer, he enjoys writing both code and prose.