February 9, 1994
The thing I like about neural networks and genetic algorithms is that you don’t have to tell the computer how to solve the problem, you just tell it what your goals are (in different ways). I’ve always felt that the current approach to AI is wrong because we’re specifying how to solve the problem, and therefore the program will never do any more. The program is limited by what we tell it; I don’t want to limit the program that way. Last summer I worked for Microsoft, and I remember at the Bill Gates party he was going off about how neural networks are the wrong approach to doing things (presumably AI) because the programmer doesn’t understand the algorithm. It seems to me that if the programmer understands the algorithm, then the algorithm is limited by the programmer! I didn’t tell him that at the time, probably because he was surrounded by 15 interns who immediately would agree with anything he said, but I still feel that neural networks and genetic algorithms are far underused methods of solving problems. (Especially the latter, which don’t require training.)