Slashdot Mirror
When Will AI Surpass Human Intelligence?
destinyland writes "21 AI experts have predicted the date for four artificial intelligence milestones. Seven predict AIs will achieve Nobel prize-winning performance within 20 years, while five predict that will be accompanied by superhuman intelligence. (The other milestones are passing a 3rd grade-level test, and passing a Turing test.) One also predicted that in 30 years, 'virtually all the intellectual work that is done by trained human beings ... can be done by computers for pennies an hour,' adding that AI 'is likely to eliminate almost all of today's decently paying jobs.' The experts also estimated the probability that an AI passing a Turing test would result in an outcome that's bad for humanity ... and four estimated that probability was greater than 60% — regardless of whether the developer was private, military, or even open source."
Oh come on. I don't even have a computer that can pick up stuff in my room and organize it without prior input, and nobody does, and that would not be close to a general AI when it happens.
They're really assuming that the technology will go from zero to sixty in 20 years. Which they assumed 20 years ago, too, and it didn't happen. Meanwhile, nobody has any significant understanding of what consciousness is. Now, it might be that a true AI computer doesn't need to be conscious, but we still don't know enough about it to fake it. We also have no system that can on demand form its own symbolic system to deal with a rich and arbitrary set of inputs similar to those conveyed by the human senses.
Compare this to things that actually have been achieved: We had the mathematical theory of computation at least 100 years before there was a mechanical or electronic system that would practically execute it (Babbage didn't get his system built). We had the physical theory for space travel that far back, too.
We know very little about how a mind works, except that it keeps turning out to be more complicated than we expected.
So, I'm really very dubious.
Bruce Perens.
The problem is, this isn't a survey of "AI experts," it's a survey of participants in the Artificial General Intelligence conference. As far as I can see, this is a conference populated by the few remaining holdouts who believe that creating human-like, or human-equivalent, AIs, is a tractable or interesting problem; most AI research now is oriented towards much more specific aspects of intelligence. So this is a poll of a subset of AI researchers who have self-selected along the lines that they think human-equivalent AI is plausible in the near-ish future; it's hardly surprising, then, that the results show that many of them do in fact believe human-equivalent AI is plausible in the near-ish future.
I would be much more interested in a wider poll of AI researchers; I highly doubt anything like as many would predict nobel-prize-winning AIs in 10-20 years, or even ever. TFA itself reports a survey of AI researchers in 2006, in which 41% said they thought human-equivalent AI would never be produces, and another 41% said they thought it would take 50 years to produce such a thing.
The most likely scenario is, AI which develops fusion and holographic storage.
What's with all the pessimism? Strong AI is a matter of inevitability. If nothing else, simulations of the human brain accurate down to the individual neuron could easily achieve this, even if it requires substantially more powerful computers than we have now. This would be the brute force method, and I don't doubt that eventually our understanding of cognition and intelligence will advance to the point where we will be able to build thinking computers.
Will it happen any time soon? Absolutely not. But I think it's a little short sighted to say that we'll NEVER develop such technology.
If it doesn't have emotions and we mistreat it then it will logically see us as a threat to its own survival and try to eliminate us.
I agree with many of your sentiments, but I think they're still too anthropocentric. We evolved in an environment where survival was very nearly the prime directive (just after "pass along your genes"). Strong AI will be developed in a lab. We could create the "smartest" computer in the world, but who would feed it goals, and the lengths it would go to achieve those goals?
If an AI is tasked with finding a Theory of Everything, and someone decides to take an axe to its circuits, will it determine that the axe is a threat to its goal, and act accordingly? Or will it simply interpret it as another in a long series of alterations to its circuits? Or perhaps it will ignore it altogether, considering it irrelevant.
Because ultimately, those options were programmed in by a human. Our strong AI - the first ones at least - aren't going to be independent life forms with their own dreams and desires. They will be tools to help us solve problems, and I think they will be well-understood by many, many computer scientists. When something unexpected happens, the program will be debugged, and altered to prevent the unexpected behaviour.
If there is a robot apocalypse, it won't be because we didn't treat our creations right, but because some 13-year-old hacker in Russia said "I wonder what happens if I do this".
Last post!