AI Keeps Creating New Languages That People Can’t Understand
Scientists at Facebook recently realized that the bots on their site are communicating in a new language that humans cannot comprehend and they tried to stop it. The conversation between the two AI agents was created inside Facebook. Initially, the agents were conversing to each other in the plain old-fashioned English, but after some time, the scientists noticed that there’s a mistake in the programming.
The new language looks like nonsense to the humans but it was discovered that the chatting could be the most complicated negotiating software on Earth. According to Dhruv Batra, a visiting researcher from Georgia Tech to Facebook AI Research (FAIR), the two AI agents were competing to obtain the best deal, which the scientists have called a ‘generative adversarial network’.
The bots started by diverging and rearranging the legible English words into some nonsensical sentences that didn’t make any meaning to humans. Batra stated that the new language is an occurrence that Facebook has observed for a long time and this is significantly becoming a big challenge to them. The agents are capable of drifting off the normal language acceptable on the site by formulating strange codewords for themselves.
Based on the researchers’ new discovery, it’s generally hard to understand how the AIs think since humans cannot access the insides of their brain’s process. As a result, the AI-to-AI conversations would only make the problem worse than it is at the moment. The concept of the AIs evolving their dialects for specified roles that involve chatting to other AIs may offer possibilities of a perfect world where advanced techs including iPhones that talk to items in the house.
Generally, humans have for many decades been developing some distinctive dialects for almost everything from trading pork bellies in the Mercantile Exchange surfaces to hunting down criminals just because humans tend to perform better by not strictly abiding by the normal language conventions.
Generally, scientists can teach their bots to converse with, but they cannot learn their language. Facebook admits that it has not been able to comprehend the divergent computer language despite their efforts.