File:Blaise Agüera y Arcas.jpg

The next big frontier is the mind and brain

"When we were kids, we felt like the space age was imminent," says Google machine learning expert Blaise Aguera y Arcas. "But in a funny way, the big frontier for our generation is the mind, the brain -- these inward spaces."

The engineer, who was the architect of Bing Maps, was joined on stage at WIRED2014 by DeepMind Technologies founder Demis Hassabis and Ben Medlock, CTO of Swiftkey. The trio have a great deal in common: a love of chess, a love of space invaders and a drive to create the next artificial intelligence revolution. For Medlock, it's about predicting natural language to make typing easier and machines preemptive and smarter, while Hassabis -- whose London-based company was acquired by Google this year -- is working to one day integrate machine learning in everything, to tackle big world problems like climate change and the flagging healthcare system. For now, his technology is being trialled and tested on Space Invaders. And it's pretty darn good.

"The computer loses three lives almost immediately," says Hassabis, indicating to a video of his machine playing Space Invaders for the first time ever. "Now, this is it after overnight training on our servers -- it's now better at playing the game than any human. It has perfectly modelled this complex stream."

The system even taught the DeepMind team a thing or two after training on Pong for 240 minutes -- it developed a perfect system of engineering a tunnel to beat the game every time.

DeepMind's applications are in ecommerce and games for now, but as with Swiftkey, the horizon is looking broad. Both Medlock and Hassabis were compelled by the idea of the intelligent machine as children and "that sense of wanting to explore the unknown," as Medlock puts it. They are engineers, and by nature problem solvers -- creating a truly intelligent machine is the ultimate problem left wide open to tackle.

The solution, as all three see it, is to "integrate all of these sources of [big data] into one unified system" as Medlock explains.

We're still a long way off from basing machine intelligence on human reality. Hassabis points out that Deep Blue, the IBM computer that can beat the world's best human chess players, and was at one point seen as the pinnacle of machine intelligence, would be stumped if it had to play a game of noughts and crosses.

"You could trivially teach that to a person in a few minutes -- but Deep Blue, you'd have to totally reprogram it. All that chess knowledge would be useless."

"We're interested in the self-learning type of AI that can learn how to master tasks directly from experiences."

www.wired.co.uk/news/archive/2014-10/16/google-zeitgeist