Making AIs work together could be a powerful force multiplier for the technology. Now, Microsoft researchers have invented a new language to help their models communicate with each other faster and more efficiently.
AI Agents are the latest buzzword in Silicon Valley. These are AI models that can autonomously perform complex, multi-step tasks. But in the longer term, some see a future where several AI agents collaborate to solve even more difficult problems.
Since these agents are powered by large language models (LLM), getting them to work together generally relies on agents speaking to each other in natural language, often in English. But despite their expressive power, human languages may not be the best means of communication for machines that fundamentally operate in ones and zeros.
This prompted Microsoft researchers to develop a new communication method that allows agents to communicate with each other in the high-dimensional mathematical language that underpins LLMs. They named the new approach Droidspeak, a reference to the beep-and-whistle-based language used by robots in Star Wars-and in a preprint article published on the arXivthe Microsoft team reports that it allowed models to communicate 2.78 times faster with little loss of accuracy.
Typically, when AI agents communicate in natural language, they not only share the outcome of the current step they are working on, but also the entire conversation history up to that point. Receiving agents must process this large portion of text to understand what the sender is talking about.
This creates considerable computational overhead, which quickly increases if agents make repeated round trips. Such exchanges can quickly become the main contributor to communication delays, researchers say, limiting the scalability and responsiveness of multi-agent systems.
To eliminate this bottleneck, the researchers designed a way for models to directly share data created in the computational steps preceding language generation. In principle, the receiving model would use it directly rather than processing the language and then creating its own high-level mathematical representations.
However, it is not simple to transfer data between models. Different models represent language in very different ways. Researchers therefore focused on the communication between versions of the same underlying LLM.
Even then, they had to be smart about what kind of data to share. Some data can be reused directly by the receiving model, while others must be recalculated. The team designed a way to solve this problem automatically in order to get the most computational benefit from this approach.
Philip Feldman, University of Maryland, Baltimore County said New scientist that the resulting communication speedups could help multi-agent systems solve larger and more complex problems than would be possible using natural language.
But researchers say there is still much to be done. For starters, it would be helpful if models of different sizes and configurations could communicate. And they could achieve even greater computational savings by compressing intermediate representations before transferring them between models.
However, it seems likely that this is only the first step towards a future in which the diversity of machine languages rivals that of human languages.
Image credit: Shawn Suttle Since Pixabay