31 July, 2017
Facebook, foxed by its AI developing a new language that the human researchers who created it couldn't understand, shut down the AI agents themselves. Why should computers be forced to work within the confines of the English language if they can communicate with each other faster using a seemingly [to them] better language?
The ultimate goal of the research program was to build an AI that could talk to and understand humans, but the fact that the AI created its own language - all by itself - is something that nobody really expected.
Dhruv Batra, a Georgia Tech researcher at Facebook's AI Research (FAIR), explained to Fast Co.
Facebook researchers had recently developed a sophisticated negotiation software that started off speaking English. The advanced system can be used to negotiate with other AI agents for completing the task at hand.
"I can can I I everything else".
According to researchers, Bob's repetition indicates how it was using the language to offer more items to Alice. When interpreted like this, the phrases appear more logical than comparable English phrases like "I'll have three and you have everything else". AI operates on "reward" principle, where they expect that by following a course of action they will get receive some benefit. It allowed AI agents to disobey rules of the understandable language and invent codewords. But in this scenario, for example, there was no reward for continuing to use English, so they chose to use a more efficient way of communicating instead. But the sentences were not completely nonsensical as they comprised a coded language which made sense to them only.
While there's nothing unsafe - so far, anyway - of AIs creating their own language, it does present a problem in that their language could be seen as more efficient than what humans have used for centuries. The language that they are speaking in is more efficient for the bots, but it becomes hard for developers to improve and work with the software.