4 ms·

It's always fun to watch non-programmers underestimate the difficulty of human-equivalent AI.

joe_the_user 17y ago
Actually, I think Ray Kurweil has a good point in saying that every technical specialist has trouble extrapolating for the increase in the power of the tools that the specialty will soon be working with. That applies to programmers as much as other engineers. When programmers start working with a machine having the level of complexity of the human brain itself, our present ideas about the difficulty of human-equivalent AI may turn out to entirely erroneous. After all, we certainly don't have such a machine to practice on today.
- yters 17y ago
  Nope, human intelligence is inherently different than anything machines can do: http://www.box.net/shared/u13u3agxqg http://www.box.net/shared/u13u3agxqg
  - joe_the_user 17y ago
    Love a quote from this 'study': "While the results have not been quantified with a statistical test, it is CLEAR ON OBSERVATION that there is a significant performance difference between the human and non human agents." Human-level AI is an extremely complex problem and I wouldn't claim know everything that might be involved in creating it. But by token, refuting any possibility of AI is an equally complex problem. Neither "Clear on observation" conclusions nor the other haphazard observations of this paper cut it for such an analysis.
    - yters 17y ago
      Eh, well I've carried out a goodness of fit test since then for stats geeks (p < 0.05), but it doesn't really add anything substantial. What is haphazard about the rest of it? If you can point out specifics it'd help me strengthen the work.
ramchip 17y ago
I don't really know what you mean by "human-equivalent", but I'll assume you mean an AGI. I don't think an actual programmer is the most qualified to build it. It could equally be a neuroscientist or a mathematician (who does some Maple or Python on the side for their job), and more realistically a multidisciplinary team. I've seen plenty of programmers underestimating the problem of AGI as well!
- ewjordan 17y ago
  I've seen plenty of programmers underestimating the problem of AGI as well! ...though for every one that underestimates it, there are plenty that overestimate it. Solving this problem doesn't necessarily mean that we do anything at all clever; it really may just boil down to having decent brain scanning technology and lightning fast processors to run the simulations. We'll definitely have the computers, in time, so the only question is whether brain scanning improves. If that's what it comes down to, programmers and computer scientists are pretty much irrelevant, and the real heroes will be the guys doing medical tool development. Granted, it would be deeply unsatisfying to "solve" the problem by blindly copying nature. But there's a very real possibility that it's going to turn out to be our first viable solution to the problem, and let's be honest: it's not going to be very difficult to do if the enabling technologies are already in place.
  - robotrout 17y ago
    > deeply unsatisfying to "solve" the problem by blindly copying nature The image that jumped to my mind, was the wing flapping airplane prototypes that, quite literally, never got off the ground, and that we all laugh at today. Perhaps, in the end, our solution won't match natures. That also leads me to another thought, which is the definition of the problem. Going back to the airplane. If the problem had been defined as "to build a machine with birdlike agility and control", than we still wouldn't be there. Instead of patting ourselves on the backs, we would still be miserably failing, over 100 years later. But we didn't really need birdlike agility, we just wanted to fly, which, it turns out, is much easier. The problem statement for a lot of AI research seems similarly ambitious. Do we really want to create some sentient being, trapped in a box? Or is what we're really after, a little more practical and obtainable? I think folks like the military, get this. They have a definite objective that they are trying to accomplish. Autonomous desert navigation, autonomous urban navigation, autonomous enemy identification. These are things you can define, measure, and make progress on.

Actually, I think Ray Kurweil has a good point in saying that every technical specialist has trouble extrapolating for the increase in the power of the tools that the specialty will soon be working with. That applies to programmers as much as other engineers. When programmers start working with a machine having the level of complexity of the human brain itself, our present ideas about the difficulty of human-equivalent AI may turn out to entirely erroneous. After all, we certainly don't have such a machine to practice on today.

yters 17y ago
Nope, human intelligence is inherently different than anything machines can do: http://www.box.net/shared/u13u3agxqg http://www.box.net/shared/u13u3agxqg
- joe_the_user 17y ago
  Love a quote from this 'study': "While the results have not been quantified with a statistical test, it is CLEAR ON OBSERVATION that there is a significant performance difference between the human and non human agents." Human-level AI is an extremely complex problem and I wouldn't claim know everything that might be involved in creating it. But by token, refuting any possibility of AI is an equally complex problem. Neither "Clear on observation" conclusions nor the other haphazard observations of this paper cut it for such an analysis.
  - yters 17y ago
    Eh, well I've carried out a goodness of fit test since then for stats geeks (p < 0.05), but it doesn't really add anything substantial. What is haphazard about the rest of it? If you can point out specifics it'd help me strengthen the work.

ramchip 17y ago

I don't really know what you mean by "human-equivalent", but I'll assume you mean an AGI. I don't think an actual programmer is the most qualified to build it. It could equally be a neuroscientist or a mathematician (who does some Maple or Python on the side for their job), and more realistically a multidisciplinary team. I've seen plenty of programmers underestimating the problem of AGI as well!

ewjordan 17y ago
I've seen plenty of programmers underestimating the problem of AGI as well! ...though for every one that underestimates it, there are plenty that overestimate it. Solving this problem doesn't necessarily mean that we do anything at all clever; it really may just boil down to having decent brain scanning technology and lightning fast processors to run the simulations. We'll definitely have the computers, in time, so the only question is whether brain scanning improves. If that's what it comes down to, programmers and computer scientists are pretty much irrelevant, and the real heroes will be the guys doing medical tool development. Granted, it would be deeply unsatisfying to "solve" the problem by blindly copying nature. But there's a very real possibility that it's going to turn out to be our first viable solution to the problem, and let's be honest: it's not going to be very difficult to do if the enabling technologies are already in place.
- robotrout 17y ago
  > deeply unsatisfying to "solve" the problem by blindly copying nature The image that jumped to my mind, was the wing flapping airplane prototypes that, quite literally, never got off the ground, and that we all laugh at today. Perhaps, in the end, our solution won't match natures. That also leads me to another thought, which is the definition of the problem. Going back to the airplane. If the problem had been defined as "to build a machine with birdlike agility and control", than we still wouldn't be there. Instead of patting ourselves on the backs, we would still be miserably failing, over 100 years later. But we didn't really need birdlike agility, we just wanted to fly, which, it turns out, is much easier. The problem statement for a lot of AI research seems similarly ambitious. Do we really want to create some sentient being, trapped in a box? Or is what we're really after, a little more practical and obtainable? I think folks like the military, get this. They have a definite objective that they are trying to accomplish. Autonomous desert navigation, autonomous urban navigation, autonomous enemy identification. These are things you can define, measure, and make progress on.