I worked with AI last semester and am doing more this semester (though at the moment it is mainly related to image processing).
What is it exactly you're curious about? AI is still pretty rudementry, it's not nearly as "Magical" as you'd think.
AI generaly boils down to one of 3 catagories of aproaches (or a combination):
1) Searches, a lot of AI is just tricky search techniques, this kind of AI is what you use when you create game playing computer AI, or things of that sort. The real trick is getting good results without your searches taking forever. There are a few techniques that are quite popular.
2) Y function Aproximators: These are your "learning" algorithms. They work by taking up a bunch of learning inputs and their corresponding output. Given that data they can then attempt to classify new sets of data. A lot of useful prediction algorithms use this approach. It is also quite useful for things like computer vision and computer driven vehicles.
3) Statistical Analysis, this is quite often incorporated with either (or both) of the approaches above. Basically this leverages things like Bayesian assumptions to predict the most likely correct output. This is very often used as part of the Y function approximators in part 2. It is a complex version of this kind of algorithm that powers Watson.
Anyway, yea.. that is pretty much a super rough outline of AI. There are 3 main issues with AI at the moment:
1) Getting algorithms to run acceptably fast is hard, this is most often an issue when using search based techniques.
2) Over-fitting data is a huge problem, overly complex classifications schemes can be counterproductive for classifying new data.
3) It's very hard to get an algorithm that determines what "values" of input to look at. By that I mean, most algorithms depend on the programmer defining the attributes to be used in classification. For example if you wanted to write a program to predict how healthy a person was, you'd have to decide what values about that person to input. Reasonable examples would be things like age, weight, history of heart disease, and so on. It would be much more useful if we could just sort of set our algorithms use on all the knowledge we had about a person, and it would decide what was useful. This is especially key in fields such as computer vision.
Anyway.... I hope this stupidly long post helped. Feel free to ask me more specific questions, though I am by no means an expert in the field.