When Jason Yosinski grad “cobbled together” a classroom demonstration of two laptops talking to each other side by side, he never expected it to become the talk of YouTube.
Built with Igor Labutov grad, the video of the talking laptops, entitled “AI vs. AI. Two chatbots talking to each other,” was originally intended as a demonstration for a class on artificial intelligence. Yet as of Sept. 12, the video has gone viral, compiling more than 2 million views on YouTube and a 99-percent “thumbs-up” rating.
The moment of revelation, Labutov describes, “It was when we joined 2 ELIZA chatbots (primitive chatbots) and saw a very silly conversation.” It was just an infinite loop of random questioning. To make it more interesting, they hooked up two cleverbots and recorded the conversation. After some weeks, when Prof. Hod Lipson, engineering, wanted to show it in his first Artificial Intelligence class, they refined the video. The code would automatically convert text to audio and vice-versa. Also, a delay of 20- 30s in the response of each cleverbot was removed.
The video shows two chatbots, a male and female figure powered by the computer program Cleverbot, conversing about seemingly random topics. Although Labutov and Yosinski did not create Cleverbot, they used the code, virtual avatars and a text-to-speech synthesiser to make the video.
The conversation begins with the female figure greeting the male with a “Hello there,” to which the male figure responds, “Hiay!”
The robots soon move on to heavier topics. After accusations that the other is a robot and an existential discussion about the existence of God, the robots sign off. The video ends showing the logo for Cornell’s Creative Machines lab.
While acknowledging that the particulars of Cleverbot’s algorithm remain ambiguous, Yosinski said the bots seem to recall what previous humans have typed into their programming and regurgitate whatever phrases seem most appropriate.
Yosinski compared how Cleverbot “learns” to the way that human babies learn language.
“First they’re saying ‘mama’ or ‘dada’ and then they’re learning actual words and soon enough they’re putting sentences together,” he said.
Yosinski and Labutov said that they decided Cleverbot would be perfect for their demonstration because of its advanced algorithm.
Lubotov said that, for this project, they had initially used a program called ELIZA, a “therapist-like” chatbot published in 1966 by Joseph Weizenbaum.
However, Lubotov said ELIZA was not sophisticated enough to simulate human conversation.
“As you can imagine when you put two together it goes into this endless loop with ‘how do you feel,’” he said. “We wanted a chat that was more intelligent and Cleverbot is definitely one of the best.”
Cleverbot also shows much about humans’ preconceived notions of artificial intelligence, Yosinski said, citing interactions between people and the bots that were later reflected in the video.
“People have been trying random phrases trying to trick this robot … at first they try things like 2+2=4 but then they realize that robots are good at math so they move to existential questions and questions about feelings,” Yosinski said, noting that the regurgitation of these more complex questions were what made the video appealing. “The cool thing about that is the robot learns about that stuff and then it’s able to talk about that which makes it seem more human.”