Turning to a friend or colleague can facilitate you solve tough problems. Now it looks like AI chatbots working together can make them more effective.
I played this week with AutoGenAn open source software platform for AI agent collaboration elaborated by researchers from Microsoft and academics from Pennsylvania State University, the University of Washington and Xidian University in China. The software uses OpenAI’s immense GPT-4 language model to enable the creation of multiple AI agents with different personalities, roles and goals that can be asked to solve specific problems.
To test the idea of collaborating with AI, I had two AI agents come up with a plan to write about collaborating with AI.
By modifying the AutoGen code, I created a “reporter” and an “editor” that discussed writing about AI agent collaboration. After talking about the importance of “showing how industries like healthcare, transportation, retail and others are using multi-agent AI,” the pair agreed that the proposed paper should address the “ethical dilemmas” posed by the technology.
It is too early to write much on any of the suggested topics – the concept of multi-agent AI collaboration is mainly in the research phase. However, the experiment demonstrated a strategy that could escalate the power of AI chatbots.
Gigantic language models like those behind ChatGPT often stumble on mathematical problems because they work by providing statistically reliable text rather than tough logical reasoning. IN paper presented at an academic workshop in May, the researchers behind AutoGen showed that collaboration between AI agents can mitigate this weakness.
They found that two to four agents working together could solve fifth-grade math problems more effectively than one agent alone. In their tests, teams were also able to solve chess problems by discussing them, and analyze and refine computer code by talking to each other.
Others have shown similar benefits when several different AI models are combined – even those from corporate rivals. In the project presented at that time workshop at a immense artificial intelligence conference called ICLR, a group from MIT and Google got ChatGPT from OpenAI and Bard from Google to collaborate by discussing and debating the problems. They determined that the duo was greater probability of finding the correct solution to problems together than when the bots worked alone. Another recent paper by researchers at the University of California, Berkeley and the University of Michigan showed that having one AI agent review and critique the work of another could allow a surveillance bot to update the other agent’s code, improving its ability to apply a computer’s web browser.
LLM teams can also be made to behave in surprisingly human ways. This was discovered by a group from Google, Zhejiang University in China and the National University of Singapore assigning separate personality traits to AI agentssuch as “easygoing” or “overconfident” can improve their cooperation in a positive or negative way.
AND recent article in The Economist summarizes several multi-agent projects, including one commissioned by the Pentagon’s Defense Advanced Research Projects Agency. In this experiment, a team of AI agents were tasked with searching for bombs hidden in a maze of virtual rooms. Although a multi-AI team was better at finding the imaginary bombs than a lone agent, the researchers also found that the group spontaneously created an internal hierarchy. One agent ended up directing 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 in coding. He says a collaborative approach can be effective, but it can also lead to up-to-date types of errors because it increases complexity. “It is possible that multi-agent systems will be the best solution, but this is not certain,” says Neubig.
People are already adopting the open source AutoGen platform in intriguing ways, such as by creating simulated writers’ rooms to generate ideas for fiction and virtual “business in a box” with agents who take on various corporate roles. It may not be too long before the task my AI agents have come up with will need to be written.