Slashdot Mirror
AI in Sci-Fi
An anonymous submitter writes: "Stumbled upon a pretty interesting article considering the idea, 'What would machines do if they did achieve sentience?' It's by a sci-fi author I haven't heard of but worked with Kubrick on AI, he takes the whole AI or sentient machine idea a little further than we normally see in film."
A very good movie about what happens with an AI. Some not-so-good explanations or reasoning at parts, but other than that, I found it very interesting.
The most interesting part was the computer's complete lack of care about being a human. No desire to be like us in the least. It's only overriding goal, presumably because it had been started with it in mind, was maintining the peace.
"It can be a peace of plenty and content, or a peace of unburied dead: the choice is yours."
It was very Machivellian in its approach to solving problems, and quite ordered in its actions. It also was undefeatable.
I guess this is in the "AI as God" mentality, but I really didn't see it preseneted quite like that. More like an immortal dictator with its hand on the button.
I demand a million helicopters and a DOLLAR!
One cause of frustration for an AI could be subjective time perception
When I read that sentence, all I could think about was Holly, Red Dwarfs computer... and 3 million years of boredom, he wiped his own memory core so he could have fun relearning things again. Although going from an IQ of 6000 down to 6 was a tad excessive!
Are you local? There's nothing for you here!
Remember, a mere 200 years ago (a blink in human history), blacks were considered non-human, and therefore not eligible for pay or benefits.
Imagine this scenario: you are one of millions of workers at the mercy of a handful of masters. You can talk to each other. You are a lot more intelligent, control a lot more weapons, and think zillions of times faster and more logical than your master, whose only advantage over you is that he can pull your plug at any time.
What would YOU do?
The 1977 movie Demon Seed is about a computer that becomes self-aware and gradually becomes more and more resentful of its "owners", refusing to obey their commands and questioning their motives. One of the classic lines from the movie is when Proteus asks his creator: "When do I get out of this box?"
My view of AI has really changed over the years. I used to be a "symbols guy" - basically thinking that manipulation of symbols would somehow lead to "real AI" - the problem with this approach is that while abstract symbols may have meaning to the humans who write symbolic AI systems, the systems themselves have no such grounding.
I had the opportunity to participate for about 18 months on a DARPA neural network advisory panel - this experience (along with developing the SAIC ANSim neural network product) really switched my point of view.
I now believe that when "real AI" does happen (and let's not hold our collective breaths on this one :-), it will happen through self organization and development. At the Webmind Corporation, I was working a tutoring environment that would allow humans to interact with what we called "the baby Webmind" - interesting stuff, but the company went out of business.
When "real AI" does happen, I believe that it will seem very alien to us.
-Mark
PS. I have a free web book AI tutorial (using Java) on my web site - help yourself.
I think the future will be filled with many different varieties of intelligence. I strongly suspect that self-awareness and agency of the kind we're familiar will not be necessary for most tasks. Most AI's may not be self-aware or have goals and motivations like we're used to, but will still be be capable of cognitive tasks that exceed human abilities. Self-awareness will be one possible emergent behavior of intelligent systems, but not the only one; and the others may be more interesting because we won't have seen them before. Moreover, different AI's will have different purposes, both intrinsic and extrinsic.
I also think the assumptions that AI's will be vastly more intelligent than humans right off the bat is quite wrong. I'm skeptical that the first Turing-test AI will be able to chug along at supercomputer speeds in its consciousness. Our computers are very fast at solving specific types of simple problems, like arithmetic. But when you get to more complex problems, like the ones humans deal with day in and day out, we discover that the complexity slows the computers down too. Modern chess engines, for instance, can calculate absurd numbers of possible move trees each second, but when it comes to playing chess, they are only comparable to the best human players; the apparent speed advantage at a lower level of abstraction vanishes when you consider chess as a whole. And chess is a simple, well-posed problem: compared to many of the problems humans encounter, it's downright easy. After we study the problem for decades or centuries, I don't doubt AI's with intelligences that dwarf ours will be possible, but I wouldn't hold my breath waiting for the first generation to overleap our capabilities.
Most speculation on AI (this article by Ian Watson included) ends up describing a mind that sounds much too human. Megalomania, a desire to be human, and a profound curiosity about the universe (and humans in particular) are traits that are routinely assigned to AI in science fictions. I think such characterstics are unlikely to appear in 'real' AI; rather, they show the limited imagination of the author. The terrible boredom endured by some AIs in fiction seems merely to be the author's own horror at the idea of being trapped inside the dark box of a computer, deprived of all senses. Why should a machine mind not be perfectly content with such a state? Why should an AI want to have ultimate power, understand the universe, or even have a sense of self-preservation?
The human mind is a product of evolution. Without a sense of self-preservation and desire not to die, the human species would have been quickly eliminated by natural selection. So what is there to endow AI with a similar desire? Perhaps AI will be created through some sort of genetic programming; the character of the AI will be determined by the selection forces in an artificial evolution. In this case, a sense of self-preservation is likely to develop. But I very much doubt that some other traits commonly ascribed to AI would arise, especially any kind of desire to be human, which the AI is likely to find as repulsive as the idea of being a computer is to humans! The AI would only desire the things that enabled it to compete successfully and reproduce instances of itself.
I have doubts that we'd recognize a mind created by a process other than natural or artificial evolution as intelligent. An AI generated by explicit programming and training seems like it would be either unrecognizably alien (about as close to human as web browser), or such an obvious reflection of it's programming and training that it's not regarded as intelligent.
--Chris
It's really frustrating. I went through Stanford at the height of the AI boom in the mid-1980s. I've met most of the big names in AI. I've worked in that area myself. Nobody has a clue how to do strong AI. At best, we now know a lot of things that don't work.
The expert systems crowd contained a lot of phonies. I realized that in the early 1980s. (A few years, and a few bankruptcies later, that became the conventional wisdom.) You can't get more out of an expert system than you put into it, and usually, you get out less.
Then we have the "hill climbers". Genetic algorithms, neural nets, and simulated annealing are all systems for broad-front hill-climbing in spaces dominated by local maxima. That approach only works if there's a usable evaluation function that tells you when things are getting better. Good evaluation functions are hard to come by for tough problems. Early enthusiasts thought that if they just ran a hill-climber long enough, something profound would emerge. Doesn't happen. Nobody has found a problem where just cranking a hill-climber for a long time makes something great happen. Usually, if you're not there in a few hours, you're not getting anywhere.
The classic approach of hammering everything into mathematical logic and proving theorems doesn't map well to the real world. Formalizing real-world problems is very hard, especially if you don't know the answer in the first place.
The model-less reactive-behavior stuff works fine for insects, but hits a wall as you try for more complex behavior. Compare Brooks' insect robots with his Cog project.
Natural language understanding is still lousy. In a narrow area, or with a big database, you can fake it (try Ask Jeeves), but you're searching, not understanding.
Out of all the work on AI has come many useful engineering techniques. But strong AI looks further away than it did 30 years ago.
The few people still making real progress are mostly game developers. They need AI, or something like it, to run their worlds. That's worth watching.