Microsoft has created an artificial language that allows AI models to talk to each other faster and more efficiently than in English, with the hope that groups of models will be able to team up without having to resort to clumsy and sprawling human words.
Many researchers believe that using several artificial intelligence models, each with different specialisms and abilities, to solve problems collectively holds promise for tackling thorny problems that individual ones can’t solve. Although large language models like ChatGPT have been shown to be capable of communicating at high speed, even reaching consensus in groups of up to 1000, designing these models to talk to each other in a language like English – as humans would talk to them or each other – creates a significant bottleneck.
Now, at Microsoft and his colleaguesĚýhave created a language called Droidspeak to streamline that communication. Named after the beeping dialect “spoken” by R2-D2 in the Star WarsĚýfilms, the language is essentially a mathematical shorthand for evoking certain words, concepts or instructions.Ěý
Advertisement
Most current AI models take English prompts and split them into a series of tokens, which can be anything from individual characters to words or even whole phrases.ĚýThis string of tokens is then turned into a mathematical description of the complex path taken by each token as it is processed by the AI’s neural network. These paths embed conceptual and factual information about language.
For instance, the individual paths for the words “king” and “queen” will be more similar to each other than “king” and “ball” are. The paths can also be manipulated mathematically, such as by taking “king”, subtracting “man”, adding “woman” and coming up with a path for “queen”.
“It’s not working in words, it’s not working in tokens, it’s working in high-dimensional space,” says at the University of Maryland, Baltimore County, who wasn’t involved in the work. “These are very complicated curves.”
Droidspeak allows the AI models to remain working in this high-dimensional space, rather than converting into text and back, which the researchers say is 2.78 times faster and has negligible accuracy loss. But there is a catch: currently, it only works with multiple copies of the same AI model. The researchers say they are hoping to investigate how to apply the concept to models that differ from each other.
Feldman says the approach will speed up communication between AI models and could also make collections of models more powerful and able to tackle larger and more complex problems.
He points out that similar ideas have been used in the past to attack AI and get it to do things that the makers want to prevent. For instance, if a model checks inputs for dangerous keywords like “bomb recipe”, you can’t get it to tell you how to make a bomb. But by working out the path for a sequence of words that conveys the same meaning, you may be able to input that to a model and sidestep the safety mechanisms.
arXiv
