Artificial intelligence: Getting computers to pretend to converse is an "extremely hard computational problem"
A computer engineer of some importance has written to say that he thinks me a bit off the mark here, where I deny that computers actually think. He writes,
modern computers are often programmed to be adaptive, in that rules are given for learning (e.g, generalizing or updating "beliefs") based on experience. So in fact computers can be (and are) programmed to "learn" things that their programmers don't know.He argues that behaviorally, this is thinking, but that it does not include consciousness (which means, I suppose, that he disagrees with Kevin Warwick).
He does, however, say,
And the important part that you and I apparently agree on is that there is no compelling reason to believe that a computer program is doing something altogether like what a human being does.Now, there, he must certainly be right. In the original post, I had discussed a problem I was having with a computer-based book order system that did not allow me to buy ten copies of a book (because no one had thought to program in the possibility of multiple orders).
The stupidest human clerk would have understood immediately.
He is really annoyed with my saying that
Most people will believe that the computer is human if it just sounds wittier or sexier than they do. In fact, the only reason this isn't yesterday's news is that so many computer nerds are inarticulate, and wouldn't have any idea what to program the computer to say.
He calls that "gratuitous" and "wrong!".
Hey, I only said that to see who I would get a rise out of. Turned out to be him, imagine!
Anyway, he advises me that making computers respond convincingly to unscripted dialogue (natural language processing) is "an extremely hard computational problem."
But I am hardly surprised. Dialogue is fiendishly difficult to write well, as my novelist friends aver. It is one reason why I went into non-fiction rather than fiction.
Labels: artificial intelligence