You guys should check out this lecture.
http://msl.stream.yorku.ca/mediasite/Viewer/?peid=a19b20bb1af6433b85778661e2ea2564
It's the most incredible AI I've ever seen. Basically this guy came up with a better algorithm to recognize shapes. It does so by representing generic data in a neural-net type of model. At first he had it analyze handwriting, with some incredible results. I love the part where he has it "dream" of the number 2 continuously, and you see it coming up with all sorts of variations on the number. But because it represents data in a generic way, they started plugging all sorts of inputs to it. Audio, text, pictures, etc. And this thing manages pretty well.
I do wonder its applicability to games. At the end he shows how the AI handles motion-capture data. It can essentially blend between them on a per-bone basis (it seems) and construct new animations from parameters. I can see great potential for this in games and films for crowds and secondary characters.
Oh, and Google is now using it. Inc Skynet in 3...2...1...