Turning to a friend or coworker can make complicated problems easier to address. Now it seems that making AI chatbots team up with each other can make them more effective.
I’ve been playing this week with Self-generationan open source software framework for AI agent collaboration developed by researchers from Microsoft and academics from Pennsylvania State University, the University of Washington and Xidian University in China. The software leverages OpenAI’s GPT-4 large language model to allow you to create multiple AI agents with different personas, roles, and goals that can be asked to solve specific problems.
To test the idea of AI collaboration, I had two AI agents work together on a plan for how to write about AI collaboration.
By modifying the AutoGen code, I created a “reporter” and an “editor” who discussed writing about AI agent collaboration. After speaking about the importance of “showing how industries like healthcare, transportation, retail, and more are using multi-agent AI,” the pair agreed that the proposed article should delve into the “ethical dilemmas” it raises. the technology.
It’s too early to write much about any of the suggested topics – the concept of multi-agent AI collaboration is mostly in the research phase. But the experiment demonstrated a strategy that can amplify the power of AI chatbots.
Large language models like those behind ChatGPT often run into mathematical problems because they work by providing statistically plausible text rather than rigorous logical reasoning. In a paper Presented at an academic workshop in May, the researchers behind AutoGen show that AI agent collaboration can mitigate that weakness.
They found that two or four agents working together could solve fifth-grade math problems more reliably than one agent alone. In their tests, the teams were also able to reason through chess problems by talking about them, and they were able to analyze and refine computer code by talking to each other.
Others have shown similar benefits when combining several different AI models, including those offered by corporate rivals. In a project presented to the same workshop At a major AI conference called ICLR, a group from MIT and Google got OpenAI’s ChatGPT and Google’s Bard to work together discussing and debating problems. They discovered that the duo was more likely to converge on a correct solution to problems together than when the robots worked alone. Another recent paper from researchers at UC Berkeley and the University of Michigan showed that having one AI agent review and critique another’s work could allow the supervising robot to update the other agent’s code, improving its ability to use the agent’s web browser. a computer.
LLM teams can also be encouraged to behave in surprisingly human ways. A group from Google, Zhejiang University in China, and the National University of Singapore found that Assign AI agents different personality traits.such as “calm” or “overconfident,” can fine-tune your collaborative performance, either positively or negatively.
and a recent article Several multi-agent projects are summarized in The Economist, including one commissioned by the Pentagon’s Defense Advanced Research Projects Agency. In that experiment, a team of AI agents was tasked with searching for bombs hidden within a maze of virtual rooms. While the multi-AI team was better at finding imaginary bombs than a lone agent, the researchers also found that the group spontaneously developed an internal hierarchy. One agent ended up giving orders to the others as they carried out their mission.
Graham Neubig, an associate professor at Carnegie Mellon University, who organized the ICRL workshop, is experimenting with multi-agent collaboration for coding. He says the collaborative approach can be powerful but can also lead to new types of errors because it adds more complexity. “Multi-agent systems may be the way to go, but it’s not a foregone conclusion,” Neubig says.
People are already adapting the open source AutoGen framework in interesting ways, for example by creating mock writers rooms to generate fictional ideas, and a virtual “business in a box” with agents assuming different corporate roles. It may not be long until the task my AI agents came up with needs to be written down.