You're right -- that was a bit of a flippant remark on my part; sorry.

I guess I should confess a certain zeal for the minimalist approach of Alicebot, which is all about dumping conventional NLP and AI techniques out the window, and going whole-hog instead for a really hardcore stimulus-response model. The typical reaction people have to this approach is to say "Isn't that just like Eliza?", to which the answer is "Yes, but several orders of magnitude smarter"...followed by a long explanation (there's a lot on the site).

I didn't invent this, Richard Wallace did. But I have to admit a certain enthusiasm for the approach, which sometimes leads me to be a little over-energetic in making claims about it.

I guess for me the point about "really works" is framed by the Turing Test -- and that is certainly a presumption that ought to be stated up front. Alice won last year's Loebner Prize, which although not an exact implementation of Turing's original experiment is the closest thing the world has right now. In that contest, the main goal is to imitate a human (in Turing's proposal, the goal was a little more constrained and had a scientific control attached). Anything that does reasonably well on that score can be said to "work", from one point of view: a more empirical one. Of course, from other points of view (and there sure are many!), anything that doesn't incorporate structures/concepts/rules that are generally conceived of as being more closely analogous to brain functionality (neural networks for instance) is off the mark.

So Alicebot has absolutely no neural network, no notion of a lexicon, etc. AIML is about as simple as you can get. This confuses people sometimes -- as though Wallace & others working on this somehow "overlooked" the need for a "learning" mechanism or a syntactic parser or a link with an "ontology of common sense" or what-have-you. In point of fact, Alicebot is just actively rejecting most other approaches, based not only on the disappointing results that they've generally produced, but also just on the premise that it might be (and has proved to be) interesting to pursue this particular, minimal, pattern-matching-based approach. Clearly no approach has "won" hands-down, but Alicebot has certainly proved a lot with a lot less financial backing than some projects (like Cyc). Time will tell...and framing the question better will help (which is what I failed to do in my response -- sorry, one more time ;-) ).

Noel