Last month I was happily playing with a little electronic game of 20 questions that a friend of mine had. Basically you think up an object and after 20 questions with yes, no or maybe answers it will make a guess at the object you're thinking of. Most of the time it was right and the times that it wasn't quite there, it was pretty close. Whilst at the time it was a bit of fun, I didn't really appreciate the amount of time and effort had been put into teaching this machine what it knows. Apparently it started learning back in 1988 by adding to it's neural network memory. Since then it has had the most popular objects and inferences burned into a chip for use in this cheap toy. The toy itself can't learn but the game you can play on the 20q website is still teaching itself for every game that is played. I just tried it out and it guessed first time.