Slog News & Arts

Line Out

Music & Nightlife

« Will the Last California Roll ... | Worst Post Ever About Pit Bull... »

Friday, March 14, 2008

I’m Afraid We Can’t Do That

posted by on March 14 at 15:29 PM

Researchers at Rensselaer Polytechnic Institute are convinced they can, with the help of IBM’s new supercomputer, build true artificial intelligence.

Human common sense reasoning is extremely hard to model. Consider how much unconscious reasoning goes into your morning routine: hearing the radio, figuring out your alarm just woke you up, turning it off, getting up and walking to the kitchen (avoiding anything on the floor in the process), deciding you need a cup of coffee … you get the idea. In the 1980s, a researcher named Doug Lenat had an idea: Just get enough people filling in enough facts into a huuuuuuge database, apply some simple straightforward algorithms to them, and - poof! - you’ll have the software apparatus to build robots who can run around figuring out how not to burn your toast. Thus the CyC project was born. They went off and spent long hours compiling collections of simple information about everyday life, like “Toasters make toast”. And entered some more. And more. And more. It’s now eighteen years later, and CyC has done some cool things, but they’re still off filling gaps in the data, and we still don’t have toast-making robots.

Since then other researchers have come up with similar models for AI - trying to decompose intelligence into multiple simple layers which can be built upon, trying to automatically learn that same common sense information about the world from libraries of text, or from the World Wide Web, and so on. Projects based on these models have solved very specific problems well and brought us wonderful technology - the Roomba that cleans my floor, for example - but the task of building a robot that can interact with and think about the world as humans do remains largely unsolved. We still don’t really understand what intelligence is - how observation and memory are connected and built upon - let alone how to model it.

These researchers have gone about the task of modeling intelligence via a Second Life character that shows the cognitive skills of a child. From this, they expect to build a machine that will eventually solve the famed Turing test. History suggests they will instead solve a couple of tiny problems very well - say, building convincing personalities for video game characters - and then move on. This AI problem has turned out to be a lot harder than we thought forty or so years ago, when Marvin Minsky declared that within a generation the “AI problem” would be solved. But the most efficient process found for creating intelligent beings so far has taken about 3.7 billion years, so I guess we shouldn’t feel so bad.

hal.jpg

RSS icon Comments

1

This is news?? Zzzzzzzz.

Posted by christopher Frizzelle | March 14, 2008 3:34 PM
2

Yeah, but can they make a computer that gets drunk and pisses away the rent money at the racetrack?

Posted by Fifty-Two-Eighty | March 14, 2008 3:35 PM
3

Most people suffer from artificial intelligence. I'd be happy if someone would strive to discover the real thing.

Posted by Sargon Bighorn | March 14, 2008 3:40 PM
4

So how long before the Internet becomes self-aware and starts eating us?

Posted by kebabs | March 14, 2008 3:49 PM
5
Posted by Peter F | March 14, 2008 3:49 PM
6

@2 - they already did that - his name is King George.

Posted by Will in Seattle | March 14, 2008 3:53 PM

Comments Closed

In order to combat spam, we are no longer accepting comments on this post (or any post more than 14 days old).