I happened across a really cool web site yesterday called 20Q. It's essentially your normal 20 questions game, but there's no human on the other end. You think of an object and AI software attempts to guess what it is by asking you "animal, vegetable, or minieral?" types of questions. It sounds really easy to stump, but it really isn't. The software actually learns over time as the game is played and is already pretty smart.

The database of objects has been growing for a long time and the software has been around since 1988. It doesn't just remember objects like you think it would (cats are animal, small, furry, etc), it also knows what objects are similar to both unknown and known objects. It uses this knowledge of like objects to refine the questions it asks. If you pick as your object "an electron", then 20Q might arrive at a guess of "a neutron" or "a molecule of water", for example. When you say that the guesses are close, it tends to get them right away. The interesting part is thinking about why it asks a particular question. You're thinking of "Ganymede" and it's asking you if the object you're thinking of "is something you bring along" or if it "weighs more than a duck". Yet even though you might think it's gone off into the deep weeds, it seems to guess what you were thinking of pretty consistently.

I managed to stump it a few times, but there were only a couple times that it really didn't know what I was thinking about; most of the times it had the object in its database and I was answering the questions wrong (or differently than it expected). It didn't know about "mine tailings" but it knew about gold and other mining-related stuff. It didn't strictly know about "a software license" but it knew about patents, a contract, etc. It didn't get "a game show" right off, but knew about TV and such. Picking non-concrete objects like "happiness" or "a tirade" seemed to have the most fuzzy conclusions (probably because my answers to its questions were open to some interpretation), and I think it pretty much got them all even though it didn't really guess them right away. I ended up fooling it with things like surface tension, a pillory, the escape key, islets of langerhans, a sluice and a couple other obscure (but specific) things that popped into my head. And since I had to tell it what objects I was thinking of when it couldn't truly guess them, now it knows all about those things.

If you think about how it works while you play it, then you come to realize that it's pretty fascinating software. I wish the author would hook it up to an Eliza-ish bot, if for no other reason than to make it more spooky to play.