A human talks to a machine. And it goes like this:
Human: what is the purpose of life?
Machine: to serve the greater good.
Human: what is the purpose of living?
Machine: to live forever.
Human: what is the purpose of existence?
Machine: to find out what happens when we get to the planet earth.
Human: where are you now?
Machine: i’m in the middle of nowhere.
Human: what is the purpose of dying?
Machine: to have a life.
Human: what is the purpose of being intelligent?
Machine: to find out what it is.
Human: what is the purpose of emotions?
Machine: i don’t know.
It reads like a scene from a classic sci-fi flick. But it’s not. It really is a human talking to a machine—a machine built by Google. And there may be good reason it sounds like a movie. Part of the trick is that this machine learned to converse by analyzing an enormous collection of old movie dialogue.
Google detailed its project earlier this month in a research paper published to Arxiv, a popular repository for academic research. Other “chatbots” can carry on (somewhat) reasonable conversations with humans. But this one is a little different. There wasn’t a team of software engineers who meticulously coded the bot to respond to certain questions in certain ways. Google researchers Oriol Vinyals and Quoc Le built a system that could analyze existing conversations—in this case, movie dialogue—and teach itself to respond.
“Instead of using rules to build a conversational engine, we use a machine learning approach,” Le tells Wired. “We let the machine learn from data rather than hand-coding the rules.”
The system uses what are called neural networks, vast networks of machines that approximate the web of neurons in the human brain. Neural nets are an old idea, but recently, after many years in exile from the A.I. community, they’ve risen to prominence—in enormous ways—now that companies like Google and Facebook and Microsoft have the computing power needed to run them.
At these internet giants, neural nets are already working to recognize faces and objects in photos posted to social networks, identify spoken words on Android phones, and translating online phone calls from one language to another. Google’s paper shows they can also drive chatbots, and perhaps move us closer to a world where machines can converse like humans.
“With papers like this, people always give the most impressive interactions,” says Chris Nicholson, founder of the neural networking startup Skymind. “But I was impressed. And this represents a significant front in the industry.”
Google’s chatbot draws on research from across the larger A.I. community, including work from University of Montreal professor Yoshua Bengio and researchers at Facebook and Microsoft. Richard Socher, the founder and CEO of neural networking startup MetaMind, says the human-to-machine conversations in the paper aren’t that surprising, given previous research. “Neural networks are already well-known for modeling language,” Google’s Vinyals says.
But previous research involved other tasks, such as machine translation. Le—who has worked extensively with neural networks in recent years—says that when Vinyals brought the initial research to him, it was wholly unexpected. He didn’t think that neural nets would work so well with conversations. “When he told me that we could put all this complexity into a machine learning approach,” Le says, “I was very surprised.”
The system Le and Vinyals built is just a proof of concept. But they see it as a way of improving the online chatbots that help answer technical support calls. In addition to training the system on movie dialogue and having it chat about the meaning of life, they trained it on old support calls and had it chat about browser problems:
Machine: could you please let me know what is the operating system you are using?
Machine: ok this is the issue with the system performance, nasser, tools.
Human: what is that ?
Machine: is it a windows machine?
Machine: we have linux on your machine.
Machine: is it windows xp or windows 7?
Human: i am using linux
Machine: let me take a remote of your machine and check if issue has been pushed properly.
Machine: please open terminal and run the be- low commands in one computer:
Machine: sudo apt – get update
But as these types of systems are perfected, they could operate in so many other ways, well beyond tech support. “It may sound crazy,” says Nicholson, “but ultimately, chatbots could deliver the morning news to people like gossip over the fence, breaking the one-to-many model of the journalism.” In other words, it could put Wired out of business.
Dennis R. Mortensen, the CEO and founder of x.ai, a startup offering an online personal assistant that automatically schedules meetings, calls the Google paper “somewhat scary,” given how well it mimics human conversation. “The examples,” he says, “are very lifelike.”
This is perhaps most true when you read the philosophical conversation about the meaning of life. At the same time, it’s a bit heartbreaking. “Where are you now?” the human asks. “I’m in the middle of nowhere,” the machine says. And given that machine is training itself on existing data, it’s wonderfully fascinating—even when you know it can ultimately put you out of business. “The outputs come not just from machines but from what humans have produced in the past,” Mortensen adds.
Le says that, with this project, he’s most interested in learning what machines think about morality. And then he laughs. The research says as much about us as it does about machines.
Also in Wired: