November 21, 2006
writing
This past week I’ve picked up reading Douglas Hofstadter’s Metamagical Themas again. Whenever I read Hofstadter I’m filled with an overwhelming compulsion to quit everything I’m doing, drop out of school and (practical considerations aside) go join the AI lab at the University of Indiana. There are some great articles in MT, “Who Shoves Who around Inside the Careenium?” and “Waking Up from the Boolean Dream”, that describe consciousness as an emergent phenomenon of a “dumb” substrate; “Boolean Dream” goes on to talk about ways of mechanizing that activity for a computer. Hofstader writes:
Although I am not pressing for a neurophysiological approach to AI, I am unlike many AI people in that I believe that any AI model eventually has to converge to brainlike hardware, or at least to an architecture that at some level of abstraction is “isomorphic” to brain architecture (also at some level of abstraction).
Coincidentally, this month’s WIRED had an interview with AI pioneer Marvin Minsky and philosopher-cum-cognitive-scientist (or is it the other way around?) Daniel Dennett that addressed similar topics. They take a slightly more hardware-oriented approach, “semi-autonomous [...] independently evolved agencies” to Hofstader’s symbols, but ultimately come to the same conclusion; namely, when we develop an artificial intelligence it will be one that mirrors the functioning of the human brain on a lower level than currently expected. Asked “What would a machine that worked this way look like?”, Dennet replies: “Like us”.
I used to dream of growing up to be The Man who Created AI. I’ve pretty much given up on that now – it may be that growing up in this age I’ve lost all sense of perspective, but I won’t be at all surprised if someone has built a thinking machine as described by Hofstadter et al. by the time I graduate from University. Minsky himself says, “If I could afford to get three or four first-rate systems programmers, we could do it.” I’m starting to think, though, that a human-style AI is just the entry point into something much broader.
Creating a piece of machinery that can accurately model a human mind on any level will be a revolutionary achievement, but it will only confirm what we already suspect about how the human mind functions. We will have learned a great deal, both about computers and about our own minds, but in the end we’re left with the same kind of computer currently manufactured with some old-fashioned lovemaking and a good grade school education.
Where the fun really begins, in my opinion, is when we start creating non-human intelligences. Programs that navigate conceptual spaces entirely alien to us. Minds incapable of making mistakes, or forming half-complete thoughts. Minds that think purely in numbers. Even (dare I say it?) the top-down AI that seems to be the butt end of so much criticism now.
This is hinted at in the postscript to “Boolean Dream”:
What bothers me is a kind of “hardware chauvinism” that we humans evince. This chauvinism says “Real Things live in three dimensions; they are made of atoms. Photons bounce off of Real Things [and so on...] The idea that being able to maneuver about in a “space” or “universe” of pure abstractions might entitle a robot to be called “sentient” would be ridiculed to the skies, no matter if the maneuvering in that abstruse high-dimensional space were as supple and graceful as that of the most skilled Olympic ice-skating champion or the greatest jazz pianist.
But Hofstadter is himself guilty of a kind of “software chauvinism” by assuming that even his robot will have to be built from a bottom-up, symbol-based intelligence. He gives excellent reasons why a human-type AI would need to be constructed this way, but doesn’t linger on the incredibly compelling idea of building intelligence that perceives the world in a way completely different from the way we experience ours.
If any one of these “alternative intelligences” were left on the doorstep of an AI lab at MIT tomorrow, I would be surprised if most people would immediately leap to call it intelligent. After all, it’s unlikely that all or even most of the AIs described above could pass the Turing test. But it’s important to remember that, in spite of all objections to the contrary, the Turing test is, by construction, anthropocentric – at the end of the day, it’s a measure of humanness rather than intelligence. I suspect that once we’ve developed the AI that everyone’s waiting for (and that will be a significant achievement) we’re going to discover just how much more territory there is to explore.
Blogged with Flock
