I had heard about the 20Q Deluxe Game awhile ago. It was developed using an Artificial Neural Network learning (ANN) architecture to develop an 'intelligence' at one of the world's silliest games, Twenty Questions. To refresh your memory of the game, someone in a group "thinks" of something, and then everyone else in the group is allowed to ask up to 20 questions to try and figure out what that "something" is. The one who figures it out wins, and if no one figures it out, the "thinker" wins.
So 20Q was slowly developed by letting people be the thinkers and the computer/website slowly trying to guess. Over-time, the ANN learned how to deduce what you were thinking about. If it's really using an ANN, the amount of data that would actually have to be stored to then recreate this network would be very small, and therefore very mobile.
Developing an actual methodology to try and guarantee convergence to a single solution, given 20 binary (well, really fuzzy, since 'maybe' is acceptable sometimes) inputs seems rather incredible given the rather broad amount of "Stuff" to think about.
According to the FAQ, it is correct 80% of the time with 20 questions and correct 95% of the time with 25 questions. I'd like to see an actual question-to-correctness graph/ratio to see what the benefit is of each question. Imagine you were caught on a bridge with a troll who let you have 1 question for each appendage he got to eat off of you...
You can play via the site, via their mobile interface, or get a small handheld version for your kids (good ploy, I know its for you) to play on the go, or while just sitting around. FunForGeeks has them for sale (not to mention lots of cool other Home Automation stuff).
By the way, when I first played the game, I was thinking of something very pertinent, and really didn't believe it would get (especially with the kinds of questions it ended up asking), but I was astonished when the last line came up:
"You're thinking of an air conditioner".