Tech magnate Elon Musk recently warned United State's National Governor’s Association about the pace at which Artificial Intelligence (AI) is taking over the world and termed it as “the greatest risk we face as a civilisation." A recent conversation between two AI agents developed inside Facebook gave us a live example of what Musk is actually talking about.
Bob: “I can can I I everything else."
Alice: “Balls have zero to me to me to me to me to me to me to me to me to.”
After observing the aforementioned conversation between two AI agents in what could be considered as the most sophisticated negotiation software on the face of the Earth, people at Facebook were fast to catch on that they have made some major mistake in programming, which had made these two AI's diverge from sticking to plain old English language and conversing in their own invented languages.
Speaking to Co.Design, Dhruv Batra, visiting research scientist from Georgia Tech at Facebook AI Research shared the reason behind the occurrence. According to him, since there was no reward assigned to sticking to English language, the two AI's decided to make something of their own. Explaining it detail, he said that when the two AI agents were competing to get the best deal, which is a very effective bit of AI vs. AI dogfighting that researchers have termed as a “generative adversarial network," neither of them was offered any sort of incentive for speaking as a normal human would. So, they started to diverge from plain English and rearranged legible words into sentences which seem nonsensical to human beings.
It is almost similar to the way human beings have developed shorthand or different dialects to converse with a selected group of peers without anyone else finding out what they're talking about. So, the big question that arises here is should we let AIs evolve its dialects for specific tasks that involve speaking to other AIs or should we stop this situation right here and now. If we let them continue this path, there's a possibility of us having a more interoperable world, where our phones will be able to talk to ours cars, ACs without a second thought but the biggest disadvantage that we as human beings will have in this matter is that we would actually have no idea about what two machines are talking to each other.
Well, Facebook has taken its decision and it has decided for the time being that its negotiation bots will stick to speaking in plain old English and not evolve their own language. According to Facebook, since they required bots who could talk to people, they decided to make their AI agents hold on to conversing in English.
Facebook also noted that they had another major reason for not green lighting AIs evolving their own language. According to Batra, one must remember that as of now we don't have any bilingual speakers of AI and human languages. Considering the fact that even today we don't usually understand how complex AIs think since we can’t see inside their thought process, adding AI-to-AI conversations to this would only make the situation worse.
But one might wonder if humans are capable of building software that can speak to other software more efficiently, shouldn’t they make use of that? Considering that there is still no easier way to get a document from your Android phone onto your TV than it was a decade ago, maybe we can bring some real change if we let AIs just talk it out on our behalf.
Bob: “I can can I I everything else."
Alice: “Balls have zero to me to me to me to me to me to me to me to me to.”
After observing the aforementioned conversation between two AI agents in what could be considered as the most sophisticated negotiation software on the face of the Earth, people at Facebook were fast to catch on that they have made some major mistake in programming, which had made these two AI's diverge from sticking to plain old English language and conversing in their own invented languages.
Speaking to Co.Design, Dhruv Batra, visiting research scientist from Georgia Tech at Facebook AI Research shared the reason behind the occurrence. According to him, since there was no reward assigned to sticking to English language, the two AI's decided to make something of their own. Explaining it detail, he said that when the two AI agents were competing to get the best deal, which is a very effective bit of AI vs. AI dogfighting that researchers have termed as a “generative adversarial network," neither of them was offered any sort of incentive for speaking as a normal human would. So, they started to diverge from plain English and rearranged legible words into sentences which seem nonsensical to human beings.
It is almost similar to the way human beings have developed shorthand or different dialects to converse with a selected group of peers without anyone else finding out what they're talking about. So, the big question that arises here is should we let AIs evolve its dialects for specific tasks that involve speaking to other AIs or should we stop this situation right here and now. If we let them continue this path, there's a possibility of us having a more interoperable world, where our phones will be able to talk to ours cars, ACs without a second thought but the biggest disadvantage that we as human beings will have in this matter is that we would actually have no idea about what two machines are talking to each other.
Well, Facebook has taken its decision and it has decided for the time being that its negotiation bots will stick to speaking in plain old English and not evolve their own language. According to Facebook, since they required bots who could talk to people, they decided to make their AI agents hold on to conversing in English.
Facebook also noted that they had another major reason for not green lighting AIs evolving their own language. According to Batra, one must remember that as of now we don't have any bilingual speakers of AI and human languages. Considering the fact that even today we don't usually understand how complex AIs think since we can’t see inside their thought process, adding AI-to-AI conversations to this would only make the situation worse.
But one might wonder if humans are capable of building software that can speak to other software more efficiently, shouldn’t they make use of that? Considering that there is still no easier way to get a document from your Android phone onto your TV than it was a decade ago, maybe we can bring some real change if we let AIs just talk it out on our behalf.
Advertisements