There is a stick figure on the screen with a wedge-shaped head. The position is one in which it moves with its knee dragging on the ground, taking up a crouched position. That's just walking! No, that's about it.
 |
| Image Credits: TechnologyReview |
Wang Rui, on the other hand, is ecstatic. He says, “Every day, when I stroll into my office, I turn on my computer and wonder what I will find.”
Uber AI researcher Wang likes to leave his laptop running with Paired Open-Ended Trailblazer, an application he helped develop, turned on overnight. If you're interested in learning about virtual bots, a training dojo for POET might be a good choice. To date, they've learned little to nothing. They're trying to make their way through a somewhat simplistic topography of railings and ravines while avoiding falling down.
But what's really interesting is not what the machines are learning, but how they're learning. Without human involvement, POET will design, build, and assign the next challenge for the robots, all of which are things that only computers can do. The bots inch their way up step by faltering step. Like a kung fu master, it may suddenly jump over a cliff, says Wang.
Wang is one of a small group of researchers investigating a new approach to creating super intelligent machines: getting AI to "automate" itself.
Jeff Clune, who was previously Wang's colleague, is one of the biggest supporters of this concept. At the School of Wyoming, Clune was part of a team of researchers that included Anthony, who was also working at Uber AI Labs at the time. At the moment, she divides her time between both the University of British Columbia and OpenAI, who have backed her.