QUESTIONS
I was sent a link to www.20q.com via the RHIZOME mailing list. It came with this paragraph of interpretation:
Most artists interested in the relations between media and consciousness are familiar with the ‘Turing test,’ a procedure which forces us to confront how an artificial intelligence might be able to pose as human. The ‘20Q test’ may soon become just as important for media artists. Invented by Ottawa-based developer Robin Burgener, 20Q is a version of the traditional ‘twenty questions’ game in which an online intelligence reads a human player’s thoughts with startling accuracy. Working with some 10,000,000 synaptic connections, the website is even able to account for false steps in players’ reasoning. The more people visit 20Q, the better it gets at guessing, making startling connections based on a logic that transcends any one individual’s ideas (the site’s handlers even claim that it ’seems to be developing a warped sense of humor’ all on its own.) The ‘20Q test’ turns the ‘Turing test’ around. In the latter, an artificial intelligence can pose as authentic. In 2OQ, one experiences how ‘authentic,’ personal thoughts can be reduced to chains of connections that seem completely artificial
I guess that’s pretty cool. I have a sort of love/hate relationship with anything AI-ish these days, just because so much of it seems like a complete joke, and for a long while I was had. (FOOL ME ONCE CAN’T GET FOOLED AGAIN.) Well, maybe not a joke, there’s plenty of great AI out there behind the scenes, just nothing like the mechanical colleagues or slaves we all thought we’d be fraternizing with by now.
Fifty-five years ago, Turing thought that in fifty years we’d have machines that would be capable of enjoying the taste of blueberries and cream. (When you reflect on the fact that he was saying these things amidst the huge, hot mazes of vacuum-tube and wire that were the computers of 1950, his optimism is even more astounding.) Instead, we got folks writing quirky little knowledge bases like
BEER -> IS_WET;
BEER -> IS_DELICIOUS;
and then hooking it up to a query system and simple grammar parser to create dialogs like
Is beer wet?
> yes
Is beer like water?
> yes
Is beer delicious?
> yes
Is beer made with hops?
> PARSE ERROR 2838659!!!! ‘UNKNOWN OBJECT OR PROPERTY’
And then the authors says something like “Hey, it at least it knows that beer is wet and delicious! Now all I need to do is apply for grants to pay research assistants to type billions of objects and properties and relations into the database. Then we can ask it anything and, (as long as it can be answered through valid inference on stable, uncontroversial objects), it will tell us!” As a result you get CYC, a worthless bloated monstrosity to which graduate students are ritually sacrificed to feed it’s terrible hunger. Or so I’ve heard.
Anyways, 20Q reminds me of this research programme, only without the pretension to creating something useful as a result. It solves AI’s “tell me something I don’t know” problem by making the whole point to tell you something you do know. Very clever. It’s a neat little toy you can play with once and say “haha wow I WAS thinking of a banana!” and then forget about forever. This is the MIT Media Lab approach to AI, which has long given up on representationalist, knowledge-base AI, and instead devotes itself almost exclusively to creating art and novelties. A big improvement, for sure.
However, contrary to the author of the quoted passage above, I did not feel in the slightest as if my authentic, personal thoughts about a banana were being reduced to a chain of “artificial” connections. Truth be told, I wasn’t even thinking of a banana, I just answered the questions as if I were thinking of one. I was instead relying on the mysterious ability of sentences to be true of some things and not of others. Turns out, it doesn’t really make a difference.
I’ll be really impressed when someone comes up with a system which plays a reasonable game of “Questions,” from Rosencrantz and Guildenstern Are Dead. That would be great.