I'm really intrigued by the limits of artificial intelligence, especially in a technology (text-based chat) that is supremely widespread and accessible for programmers.
Cleverbot, a chat bot launched in 2009, bases its replies on a growing database of over 20 million actual online conversations. This means that, at times, it can give surprisingly human-sounding replies. But, at other times it completely misses the point, because it's only picking from fragments of other conversations.
The best AI I've seen is called SILVIA, from 2007, which seems able to store functional and relational vocabulary in memory to construct very meaningful sentences from scratch. A rather lengthy video is posted above, but you can get a great example of SILVIA's ability just from random sections. At various parts of the video, when asked identical questions from earlier, SILVIA responds with the same answers but in very different words. It's quite impressive.
Give Cleverbot a try and let me know what you think. Or if you know of any more impressive ones.