AI bots develop their own language

Kramer Phillips By Kramer Phillips | 1 year ago

Facebook engineers training automated “bots” to interface with human Facebook users ran into an unexpected problem, according to a recently released company report: The bots had formulated their own symbols and codes to communicate with each other, and the human engineers could no longer follow along. The engineers said it was a rare instance of robots creating and using their own language.

This bot improvisation happened when the engineers activated a machine-learning software program to facilitate bot-to-bot communication. In the end, the engineers had to modify the program, as the bots’ communications were diverging too far from their original programmed lingo.

The engineers do not consider this phenomenon to be true artificial intelligence, per se, but they do consider it to be a step forward for “machine learning”—i.e., the ability of machines to learn new information without having it programmed into them. These bots’ behavior offers new clues as to how machine learning works and where research on it might go next, they noted in their report.

“There remains much potential for future work,” the engineers wrote, “particularly in exploring other reasoning strategies, and in improving the diversity of utterances without diverging from human language.”

This would not be the first time that computers created their own way of speaking. Last year, Google discovered that its Neural Machine Translation system, which translates any one human language into another, was using a unique internal language that it had created on its own to understand the human languages and translate them more accurately. And in March of this year, a joint project by Igor Mordatch and Pieter Abbeel of OpenAI and the University of California, Berkeley, made bots come up with a new language to communicate with each other.