Mark Zuckerberg may believe Elon Musk’s warnings about the dangers of artificial intelligence are “irresponsible,” but a recent incident at the social network appears to suggest the Tesla boss could have a point: Facebook researchers decided to shut down an AI they invented after it started speaking its own made up language.
In a conversation between two bots called Bob and Alice, Bob states: “I can i i everything else.” Alice responds with the equally bizarre: “balls have zero to me to me to me…” The rest of the exchange consisted of variations of these sentences.
Rather than this being the first step toward the system concluding humans are a virus that should be wiped out, it’s merely the way in which the intelligence operates. The bots are negotiating an exchange of balls, with repeated use of words like “i” and “me” representing the number of items.
Modern technology uses a “reward” system where it expects actions to have a “benefit.”
“There was no reward to sticking to English language,” Dhruv Batra, a research scientist from Georgia Tech who was at Facebook AI Research (FAIR), told Fast Co. Design . “Agents will drift off from understandable language and invent code-words for themselves. Like if I say ‘the’ five times, you interpret that to mean I want five copies of this item. This isn’t so different from the way communities of humans create shorthands.”
“It’s definitely possible; it’s possible that [language] can be compressed, not just to save characters, but compressed to a form that it could express a sophisticated thought,” Batra added.
This isn’t the first instance of AIs abandoning English in favor of “shorthand.” Google’s AI, for example, can efficiently translate between language pairs it hasn’t explicitly been taught by using its own made up ‘universal language’ as a buffer.
As for Facebook ’s AIs, the researchers decided to shut down the system as they wanted the bots to communicate with people.