I’ve recently revisited the state of AI (as available to the common man, at least — who knows what’s going on in government research), and frankly the news isn’t too good. Some friends from Slashnet have been mucking about with AI bots to run over IRC, and I am seriously underwhelmed.
MegaHAL doesn’t do a bad job of learning new phrases from conversations he observes, but has no real ability to parse them out into parts of speech. He cuts and pastes from his store of knowledge with some facility, but it’s usually nonsense, and if he does
manage to make sense, it’s almost always rather surreal. A.L.I.C.E. is a much more realistic conversationalist, albeit in an ELIZA-like manner, and can learn things, but all of her data has to be marked up in AIML — she’s unable to learn things on the fly from conversations — which means that she’s a fairly static creature. You can spend lots of effort coding knowledge into her, but then she’s done.
Most promising of all seems to be Hal2003, an honest-to-goodness neural net, which begins life as the veritable mewling infant and is instructed through interaction. The user tells Hal2003 when he has made an incorrect statement and provides an example of a correct response. And so he learns. I look forward to seeing what comes of Hal2003. Unfortunately for those of us looking for an AI to run right now, Hal2003’s source has not been released to the public, and he is only accessable through an interface on his creators’ website.