For those who needed to sum up what has made humans such a successful species, it’s teamwork. There’s growing evidence that getting AIs to work together could dramatically improve their capabilities too.
Despite the impressive performance of enormous language models, firms are still scrabbling for methods to place them to good use. Big tech firms are constructing AI smarts right into a wide-range of products, but none has yet found the killer application that can spur widespread adoption.
One promising use case garnering attention is the creation of AI agents to perform tasks autonomously. The fundamental problem is that LLMs remain error-prone, which makes it hard to trust them with complex, multi-step tasks.
But as with humans, it seems two heads are higher than one. A growing body of research into “multi-agent systems” shows that getting chatbots to team up might help solve most of the technology’s weaknesses and permit them to tackle tasks out of reach for individual AIs.
The sector got a big boost last October when Microsoft researchers launched a brand new software library called AutoGen designed to simplify the strategy of constructing LLM teams. The package provides all of the obligatory tools to spin up multiple instances of LLM-powered agents and permit them to speak with one another by means of natural language.
Since then, researchers have carried out a bunch of promising demonstrations.
In a recent article, Wired highlighted several papers presented at a workshop on the International Conference on Learning Representations (ICLR) last month. The research showed that getting agents to collaborate could boost performance on math tasks—something LLMs are likely to struggle with—or boost their reasoning and factual accuracy.
In one other instance, noted by The Economist, three LLM-powered agents were set the duty of defusing bombs in a series of virtual rooms. The AI team performed higher than individual agents, and one among the agents even assumed a leadership role, ordering the opposite two around in a way that improved team efficiency.
Chi Wang, the Microsoft researcher leading the AutoGen project, told The Economist that the approach takes advantage of the very fact most jobs may be split up into smaller tasks. Teams of LLMs can tackle these in parallel fairly than churning through them sequentially, as a person AI would need to do.
To this point, organising multi-agent teams has been an advanced process only really accessible to AI researchers. But earlier this month, the Microsoft team released a brand new “low-code” interface for constructing AI teams called AutoGen Studio, which is accessible to non-experts.
The platform allows users to pick from a collection of preset AI agents with different characteristics. Alternatively, they will create their very own by choosing which LLM powers the agent, giving it “skills” reminiscent of the power to fetch information from other applications, and even writing short prompts that tell the agent easy methods to behave.
To this point, users of the platform have put AI teams to work on tasks like travel planning, market research, data extraction, and video generation, say the researchers.
The approach does have its limitations though. LLMs are expensive to run, so leaving several of them to natter away to one another for long stretches can quickly turn out to be unsustainable. And it’s unclear whether groups of AIs will probably be more robust to mistakes, or whether or not they could lead on to cascading errors through the complete team.
A lot of work must be done on more prosaic challenges too, reminiscent of the very best technique to structure AI teams and easy methods to distribute responsibilities between their members. There’s also the query of easy methods to integrate these AI teams with existing human teams. Still, pooling AI resources is a promising concept that’s quickly picking up steam.
