Gigaom AI Minute – January 25

Transcript

There's a difference between narrow AI, what we have now, that is an artificial intelligence that can do a task, a defined task. And a general AI, an AI like you see in the movies, it can do things it hasn't been trained to do. And the question that I would like to address today is whether the techniques that we use to build a narrow AI, are those eventually going to get us to a general intelligence?

Opinion is mixed on this, but I personally think that it is not. But the way we think about artificial intelligence right now is we take a bunch of data, and we say that it's a representation of the universe in this one domain space. And then we get the computer to study it and figure out an optimal solution, as it were, to that data, whatever the goal is. It then applies that learning to things going forward. There's a difference between narrow AI, what we have now, that is an artificial intelligence that can do a defined task, and a general AI, an AI like you see in the movies, it can do things it hasn't been trained to do. And the question that I would like to address today is whether the techniques that we use to build a narrow AI, are those eventually going to get us to a general intelligence? Opinion is mixed on this, but I personally think that it is not. But, the way we think about artificial intelligence right now is we take a bunch of data, and we say that it's a representation of the universe in this one domain space. And then we get the computer to study it and figure out an optimal solution, as it were, to that data, whatever the goal is. It then applies that learning to things going forward.

It's not at all evident to me that this is anything like how general intelligence works. We don't do one narrow thing and learn that, and then perfect, and then can't do any slightly different variant of that. The way humans do knowledge is completely different than how a computer does. The way a brain orders information is very different from how a computer does. And so, if we're ever going to build a general intelligence, I believe it's going to require an entirely different way to think about computers and their intelligence. It's not at all evident to me that this is anything like how general intelligence works. We don't do one narrow thing and learn that and then perfect and then can't do any slightly different variant of that. The way humans do knowledge is completely different than how a computer does. The way a brain orders information is very different from how a computer does. And so, if we're ever to build a general intelligence, I believe it's going to require an entirely different way to think about computers and their intelligence.