Google's director of research talks AI

SAN FRANCISCO--Google's influence on the Web can be likened to game theory, said Google Director of Research Peter Norvig.

Very simply defined, game theory applied to economics is a study of the strategic interactions among players and their influences on each other.

"We (once) thought of ourselves as observers of the Web. We made a copy of it, and we thought it was just a reflection of the Web," Norvig said Sunday while speaking here at the Singularity Summit, a two-day conference on artificial intelligence.

"Now we understand that we're co-evolving. When we make a change, it changes. Search engine optimizers watch us, and when we make a move, then they make a move. The Web moves in different directions because of the interaction between us," he said.

Norvig, who wrote a widely read book on artificial intelligence, gave the opening keynote address at the second day of the Singularity conference, which drew about 800 attendees. The so-called singularity refers to the point at which advances in artificial intelligence will bring about self-improving machines that are smarter than humans. Some technologists believe that the current rapid advancement in computer hardware and software are leading us to this point.

But Norvig pooh-poohed that idea, saying that current data doesn't necessarily show that we're at a time of accelerating change brought on by technology. For example, he referred to a chart of U.S. GDP growth over the last 100 or so years that showed constant progress, without great spikes because of space flight or the introduction of the personal computer. Similarly, a chart of the annual growth in the world's GDP shows a 1 percent to 2 percent fluctuation in growth from 1970, with no solid trend up or down, he said.

"From any point in time, it can look like things are happening more rapidly," said Norvig, who's also the former head of the NASA Ames computational group.

Similarly, when it comes to artificial general intelligence, Norvig said that he hasn't seen much indication that we're at an increased point of breakthrough versus 20 or 30 years ago. "We need more data and more models of what that data does," he said.

However, Norvig said he was encouraged by recent research in AI, such as work on "probabilistic first-order logic" at Stanford University and UC Berkeley. He said that probabilistic first order logic, or programming software that can quantify data over multiple states, will be key to developing artificial general intelligence. Other prerequisites include software that can draw hierarchical representations, such as vision systems that can make leaps from pixels to faces to people in a crowd, he said. Machines will also be able to learn from lots of data, online and in an efficient way, he said. "Once we have those components then we have artificial general intelligence."

Earlier in the talk, he compared a child to a baby chimpanzee, saying that there's more similarity than difference. But when comparing the cultures of humans and chimpanzees, that's where the real differences are visible. "That's the promise of what humans are. It's not the individual intelligence but this collective intelligent culture," he said.