Getting AIs to work together may very well be a strong force multiplier for the technology. Now, Microsoft researchers have invented a brand new language to assist their models seek advice from one another faster and more efficiently.
AI agents are the newest buzzword in Silicon Valley. These are AI models that may perform complex, multi-step tasks autonomously. But looking further ahead, some see a future where multiple AI agents collaborate to resolve even more difficult problems.
Provided that these agents are powered by large language models (LLMs), getting them to work together often relies on agents talking to one another in natural language, often English. But despite their expressive power, human languages may not be the very best medium of communication for machines that fundamentally operate in ones and zeros.
This prompted researchers from Microsoft to develop a brand new approach to communication that enables agents to seek advice from one another within the high-dimensional mathematical language underpinning LLMs. They’ve named the brand new approach Droidspeak—a reference to the beep and whistle-based language utilized by robots in Star Wars—and in a preprint paper published on the arXiv, the Microsoft team reports it enabled models to speak 2.78 times faster with little accuracy lost.
Typically, when AI agents communicate using natural language, they not only share the output of the present step they’re working on, but in addition your entire conversation history leading as much as that time. Receiving agents must process this big chunk of text to know what the sender is talking about.
This creates considerable computational overhead, which grows rapidly if agents engage in a repeated back-and-forth. Such exchanges can quickly develop into the most important contributor to communication delays, say the researchers, limiting the scalability and responsiveness of multi-agent systems.
To interrupt the bottleneck, the researchers devised a way for models to directly share the info created within the computational steps preceding language generation. In principle, the receiving model would use this directly slightly than processing language after which creating its own high-level mathematical representations.
Nonetheless, it’s not easy transferring the info between models. Different models represent language in very other ways, so the researchers focused on communication between versions of the identical underlying LLM.
Even then, that they had to be smart about what kind of knowledge to share. Some data may be reused directly by the receiving model, while other data must be recomputed. The team devised a way of working this out robotically to squeeze the most important computational savings from the approach.
Philip Feldman on the University of Maryland, Baltimore County told Latest Scientist that the resulting communication speed-ups could help multi-agent systems tackle greater, more complex problems than possible using natural language.
However the researchers say there’s still loads of room for improvement. For a start, it could be helpful if models of various sizes and configurations could communicate. And so they could squeeze out even greater computational savings by compressing the intermediate representations before transferring them between models.
Nonetheless, it seems likely that is just step one towards a future by which the range of machine languages rivals that of human ones.
