Craig Silverstein, technology director at Google, was recently interviewed by Stephanie Olsen at CNET. The discussion got around to Artificial Intelligence, or at least Artificial Reference Librarians:
You have portrayed the ideal search engine as one resembling the intelligence of the Starship Enterprise or a world populated with intelligent search pets. Can you talk a little bit about those ideas?
Well, the third idea is having the computer be as smart as a reference librarian. That's interesting, because reference librarians, of course, use computers, use Google to help them search, but they put some element of intelligence into it that the computer cannot do by itself.
So, part of the goal is to make computers smart enough so that when you interact with them, they can do something with that information to help you actually get better results. That is certainly something Google thinks about to improve quality.
When do you think that kind of artificially intelligent search will happen?
I think that understanding language is kind of the last frontier in artificial intelligence, and then talking to a computer will be just like talking to a reference librarian, because they will both be equally knowledgeable about the world and about you.
The big difference, and this is where the search pets come in, is that the reference librarian will understand emotions and other nonfactual information that even a fully intelligent computer may have trouble with.
In terms of timing, I typically say about 200 to 300 years. I think it is probably closer to the 300th year end of it. But if it ends up being closer to the 200th year, I would not be around in any case, and I will not be able to have anyone gainsay me.
Going back further, even 30 years, the people who were working on artificial intelligence in the '60s thought all these problems would be solved by today--and we are basically not very much closer in terms of those overall high AI goals of understanding language.
So basically, Silverstein agrees with Marvin Minsky -- and just about everybody else -- that AI is brain dead. But further, he's making a quantitative estimate that it will take between 200 and 300 years to create an artificial reference librarian.
SIlverstein is only 31, and he's not a scientist, so we can't apply Arthur C. Clarke's dictum that "When a distinguished but elderly scientist states that something is possible, he is almost certainly right. When he states that something is impossible, he is very probably wrong."
So we need another proverb, along the lines of "When a bright young technology hotshot predicts that something will be developed next year, (s)he's almost certainly ____. When (s)he predicts that something will not be developed for two or three centuries, (s)he's very probably ____." Fill in the blanks for yourself. There's more than one right answer, I'm sure.
In the absence of a compelling "young technologist" substitute for the "elderly scientist" role in Clarke's Law, we might instead turn to Kernighan's Law (at least I think it was Brian Kernighan from whom I first heard this): "When a programmer says that a piece of code will take X amount of time, double the estimate and raise it to the next higher unit." In other words, "half an hour" means "one day"; "one day" means "two weeks"; "two weeks" means "four months"; and so on.
According to Kernighan's Law, 200-300 years translates to 400-600 decades, i.e. four to six millennia. On that view, the time period between AI and the development of the computer would be about the same as the time period between the development of the computer and the invention of writing. That would have a nice symmetry, but I doubt that it's true, if only because cultural prediction on a time scale of millennia (or even centuries, at this point) seems absurd.
[Silverstein interview via Blogos.]Posted by Mark Liberman at May 27, 2004 08:03 PM