Slashdot Mirror
Comparing Clarke/Kubrick's 2001 To Now
angkor wrote us about a recent Economist article
that explores and compares the differences between Clarke/Kubrick's vision of 2001, and what we've got. Of course, I'd point out that the literary one wasn't meant to be a literal 2001; but this an interesting comparasion nonetheless.
regardless of what we didn't have achived, look at what we have.
Process of the univers birthed an intelligence: evolution
Evolution birthed a greater intelligence: Us
We birthed(or are birthing) a greater intelligence than us: technology (ai)
What will technology birth?
The universe is doing nothing less than attempting to become aware of itself... piece by piece.
--
What is the sound of this sentence?
Leveraged buy-outs, insider trading, junk bonds, corporate mergers, golden parachutes - all this has destroyed what was once the paradigm for how to do things right. When 2001 was made, a 10 or 20 year corporate game plan was not unusual. Now you'd be luck to find any corporate plans looking ahead more than 10 or 20 months. Oh, and need I mention the "dot-com" crash as a perfect example of what this new culture breeds?
Our problem is that we think like the humans we are.. That includes a pretty large amount of overestimation of our own abilities. The human kind of intelligence is probably *not* the only one that can exist. Trying to copy the human brain (neural networks etc.) is not only hardly possible, it wouldn't be what we want. The human brain does not provide the best kind of intelligence for analyzing stock data, creating optimized electrical circuits or whatever. It is optimized on remembering pictures, sounds, faces and communicating with other humans. An intelligent machine would require different abilities. Let's not be too arrogant and conclude that because our first attempts of creating intelligence failed we'll never achieve it.. Maybe just rethink what intelligence actually is.
Artificial Intelligence has arrived right on time in 2001 as predicted by Stanley Kubrick, but not as the Heuristically programmed ALgorithmic (HAL) computer that tried to get Dave to open the pod bay door. Instead, the A.I. is a primitive, low-intelligence virtual entity striving to establish itself in such forms as Visual Basic Mind.VB and Java-based Mind.JAVA -- earthbound AI Minds incapable of space flight.
When the film 2001: A Space Odyssey came out in 1968, we had not yet even heard of the now onrushing Technological Singularity beyond which no science fiction writer can even imagine what things will be like. because it's a Singularity .
It's a cool book to read if you're interested in AI (but not an expert, then it could be all old news I guess), but it is a bit expensive (at least here in Europe)..
'HAL's Legacy', edited by David G. Stork, MITpress, ISBN 0-262-19378-7. Oh, I just found an online version at MIT, check it out: http://mitpress.mit.edu/e-books/Hal/
NachtVorst
Considering the AI 'boom' of the 80s failed to produce anything concrete on almost every level, there's still a deep seated resentment against AI and AI researchers in some circles.
Considering Huxley wrote that novel in 1932 (the structure of DNA wasn't even found until the 1950s!), its rather amazing how accurate both the technology (in general, not the details, since when he was writing it a lot of this was far off fantasy) and the social aspects of it are compared to the current day.
Simple amazing...
1984 was a completely symbolic date. The book was written in 1948 as a critique of the british society of the day by reversing the digits of the time Orwell cast a dystopian future metaphor for the subject of his ire.
"The first thing to do when you find yourself in a hole is stop digging."
Humanity is interested in space exploration, it's just that the people in charge can not find the profitability in space exploration.
In the beginning space exploration was about showing off how powerful one's defense industry could be to the point that America proved it could put a man on the moon and therefore also establish a lunar base from which to lob missles at the former USSR.
The science of the lunar missions and the subsequent Mars missions were simply funded by the excess money generated by the defense industry to make space exploration seem legitimate in the first place with the veil of scientific inquiry.
Back in the good old days of space exploration (late fifties to mid seventies) there was profit in space exploration. Sadly today NASA works on a shoestring (for space exploration) budget making things which could realize the dreams of mankind just dreams.
Our problem is that we think like the humans we are.. That includes a pretty large amount of overestimation of our own abilities. The human kind of intelligence is probably *not* the only one that can exist....Let's not be too arrogant and conclude that because our first attempts of creating intelligence failed we'll never achieve it.. Maybe just rethink what intelligence actually is.
But that's precisely the problem with trying to "achieve AI"--defining what the hell "intelligence" is. For better or worse, people have traditionally defined "intelligence" roughly as "the things people can do but animals can't," or, "the things people can do but it makes our noggins hurt after a while." When put this way, the deficiencies in this definition become pretty apparent, but no one has come up with an obviously better version. Instead we usually approach the question of whether a thing is "intelligent" using the standards of the old Supreme Court decision defining obscenity--we think we know it when we see it.
Or more often, we think we know what it isn't when we see that. The history of "the quest for AI" (I put that in quotes very advisedly) is full of problems that, if solved, would surely be proof of AI...until they are solved, in which case it's still a dumb computer. Computers are now the world champions or competitive with world champions in chess, checkers, backgammon, othello, poker, bridge, and almost any game of mental skill with the significant exception of go. Computers have both proven several important and previously unproved mathematical theorems (e.g. the 4 color map coloring conjecture) and have come up with elegant and/or novel proofs for existing theorems (e.g. a computer proof of Godel's Incompleteness Theorem which "invented" Cantor's diagonalization technique on its own).
On the other hand, we have yet to make a computer which can navigate and react to its environment as well as, say, a pet dog can (sorry AIBO), nor one which can understand human language in any but the most limited domains. (Of course "understand" is a similarly difficult to define term. As an example of what I mean, look at CYC, a company which gets its name from its initial mission when it was founded IIRC back in 1984--to program a computer which understood enough concepts to understand language well enough that it could read an enCYClopedia (or any other descriptions in natural language) and learn what it didn't already know. While CYC has developed a useful system, it's still a ways from passing the encyclopedia test.)
Even though we're used to thinking of playing championship-level chess or doing advanced mathematics as hallmarks of particularly intelligent humans, while navigating an environment or understanding language is something that even the dumbest people can do, we find that computers are good at different things. (Or rather, we know how to program computers to be good at some things but not other things.)
The "problem" has been that in the early days of computers and on into the "golden age" of AI, we didn't know squat about how the human brain worked, nor even about what sorts of steps were needed in order to e.g. understand natural language. Back then, most AI researchers--brilliant people, mind you--figured all that would be necessary for a computer to understand language would be a link to a dictionary and maybe some rudimentary ability to parse grammar. Indeed, in many ways the field of linguistics arose as a result of the attempts and failures of computer scientists to get computers to understand language. Similarly, the successes and failures of AI have been instrumental in guiding or even creating the field of computational neuroscience.
What we are coming to understand is that the things that only "more intelligent" people can do are not really the hallmarks of "intelligence" but rather are examples of people fitting their brains to tasks they were not really designed for. For AI to truly "be achieved", we will have to get much better at making computers succeed at the tasks which a monkey can do just as well as a human, rather than those which humans can do but monkeys can't.
Also, we're learning that our instinctive idea of "intelligence" demands that techniques be general rather than specific. In other words, we don't consider exhaustive depth-limited minimax search with static evaluation to be a truly intelligent game playing technique--even though it can allow a computer to become the world chess champion--because it really sucks at go. The fact that go has a branching factor (i.e. avg. # of legal moves) of over 300 while chess has one of around 30 doesn't mean that similar thinking techniques (so far as we can tell) can't be used for a human to play both, but it does mean that exhaustive search is a feasible technique for a chess-playing computer but not a go-playing computer; we tend to interpret this (rightly or wrongly) as saying that exhaustive search is not an "intelligent" technique.
Next, it's time to stop tossing around that crap about how computers are so much faster or more powerful than human brains. That's complete hogwash. A modern CPU has roughly 10^6 gates, compared to ~10^11 neurons in a human brain. A computer might have 10^9 bits of memory (or even 10^10 if we go really high-end), and 10^11 bits of storage space, but a human brain has ~10^14 synapses, which can be viewed as encoding part of what the brain knows. A human brain has a remarkable 10^14 bits/sec of data bandwidth, compared to ~10^10 bps for a PC and 10^11 bps for e.g. the upcoming Alpha EV7. The only category computers lead in is cycle time, roughly 10^-9 for computers compared to ~10^-3 for the human brain. The upshot of all this is that, when it comes to computers programed as neural networks, a computer can only perform about 10^6 neuron updates/sec compared with 10^14 for a human brain, and the largest computer networks (limited by feasibility not by space) are maybe 10^5 neurons compared to 10^11 in the brain. So, roughly 100,000,000 times slower and 1,000,000,000 times smaller than a brain. (Figures based upon those in _Artificial Intelligence: A Modern Approach_, updated for the 7 years since the book was published.) No wonder computers aren't as intelligent as a human brain! And yet despite the huge disadvantage, neural nets are still the best technique for many AI problems, especially if we are worried about coming up with a technique which seems to be generally intelligent.
And finally, while it's interesting to talk about why we haven't created HAL yet, it's important not to confuse this with the idea that "the field of AI is a failure". AI is *not* a failure. While some problems have proven much harder than we initially expected, this is almost entirely because our initial expectations were completely ignorant, rather than because progress has not been made. Most importantly, we need to realize that people who are working in the field of AI are not sitting there day after day trying to create Lt. Commander Data or pass the Turing Test. Rather they're working on solutions to limited domain problems where computers can augment or replace the efforts of humans--and they're succeeding in many, many instances. The only real "problem" with the field of AI is defining what exactly it is.