Training the AI through his dreams only worked as long as the dreams provided him some kind of positive feedback. If those dreams failed to live up to reality…
As a more than a mimicry of human life, the AI responded to the same kinds of reinforcement that human beings did. His programmers would have to do better than to make the AI's positive reinforcement a higher score or cheese at the end of a maze.
I hope that my department impresses me more next semester. Although my AI class was interesting (I learned how to train my Pavlovian Pacman player, how to write recursive Prolog, and how to store n-grams in graph form), the other course only taught me how I should conduct myself more professionally.
I want to be a professional who gives my colleagues more than a day's notice; who keeps track of seemingly "missing" colleagues; and who gives criticism using emotionally even, easy-to-understand language. As a writer, I find my last goal particularly interesting to think about. Would I feel comfortable with being a poetic grader? How would words like "vomit" affect the students whom I was grading?
Here are, perhaps, questions that merit research: Does a style manual exists for coders? Are method names automatically nouns although we (or just I?) think of them as code components that perform specific actions? (For example, would you write a Java method that shuts something down as "shutDown()" or "shutdown()"?) What other ways exist to store and search n-grams besides parse trees? Just what is the best way to train a robotic arm to flip pancakes?