Previously, you said that a lot of work in Artificial Intelligence is “5% intelligence and 95% rigged demo”. What would you consider an example of something that has a higher “intelligence ratio”, if there is one, and what efforts do you consider most likely to increase this ratio?
Good point. I remember that in that context, Eliezer Yudkowsky had spoken highly of Sebastian Thrun’s CES (C++ for embedded systems). I started reading Thrun’s exposition of CES but never finished it.
Still, I’d like to hear Eliezer’s answer to my question, in case there’s more he can say.
Eliezer called the example in the link 95% rigged by virtue of how much the problem was constrained before a program attacked it. Chess is likewise the very definition of a constrained problem.
Certainly, the Kasparov match wasn’t “rigged” (other than being able to review Kasparov’s previous games while Kasparov couldn’t do the same for Deep Blue), but when the search space is so constrained, and tree pruning methods and computer speed only get faster, it’s bound to surpass humans eventually. There was no crucial AI insight that had to be overcome to beat Kasparov; if they had failed to notice some good tree-pruning heuristics, it would have just delayed the victory by a few years as computers got faster.
Previously, you said that a lot of work in Artificial Intelligence is “5% intelligence and 95% rigged demo”. What would you consider an example of something that has a higher “intelligence ratio”, if there is one, and what efforts do you consider most likely to increase this ratio?
DARPA’s Grand Challenge produced several intelligent cars and was definitely not a rigged demo.
Juergen had some things to say about that actually.
Something to say about “intelligent cars” or “not a rigged demo”?
Good point. I remember that in that context, Eliezer Yudkowsky had spoken highly of Sebastian Thrun’s CES (C++ for embedded systems). I started reading Thrun’s exposition of CES but never finished it.
Still, I’d like to hear Eliezer’s answer to my question, in case there’s more he can say.
Chess-playing AI has had a lot of decent-quality work done on it, and Deep Blue beating Kasparov was definitely not a rigged demo.
Eliezer called the example in the link 95% rigged by virtue of how much the problem was constrained before a program attacked it. Chess is likewise the very definition of a constrained problem.
Certainly, the Kasparov match wasn’t “rigged” (other than being able to review Kasparov’s previous games while Kasparov couldn’t do the same for Deep Blue), but when the search space is so constrained, and tree pruning methods and computer speed only get faster, it’s bound to surpass humans eventually. There was no crucial AI insight that had to be overcome to beat Kasparov; if they had failed to notice some good tree-pruning heuristics, it would have just delayed the victory by a few years as computers got faster.