5 ms·

> not even to maintain or develop the AI's, since they can, by definition, do that themselves, with infinite parallelizeability to boot. I don't think it works

by Ygg2 13y ago

> not even to maintain or develop the AI's, since they can, by definition, do that themselves, with infinite parallelizeability to boot. I don't think it works like that. I think AI and intelligence in general is more akin to optimization of code where there are few big wins and constant diminishing returns. I think the basic premise of Kurzweil is wrong. I think brain works like a simulation, where you get more finer picture, the more accurate the data and the more accurate the data the more of data you need to process.

Houshalter 13y ago
The thing is, even if that's the case, each improvement still decrease the time it would take to get to the next improvement. Also computer hardware can be improved as well. Imagine Moore's law where instead of improving every 18 months, the scientists designing the computers are also getting twice as fast. Now imagine they are also getting smarter from improvements made by code optimizations. And then economic growth as well (more and more computers running the AIs or bigger factories to make them.) Even if progress does plateau, it doesn't have to happen anywhere near human scales. For example computers already have millions of times more serial processing ability than human brains. Any optimization that can take advantage of that, even if for only some problems and domains, would automatically give them a massive advantage over us.
- Ygg2 13y ago
  If scientist were getting twice as fast, I doubt it would make a difference. As I said, ok scientist are twice as fast, but can they think of a novel solution twice as fast? What is the last great physics discovery that was by a single (wo)man? Why are there no medals in sciences now won mostly by teams? Again I don't think it won't reach super human levels. I just don't think the Singularity is within next thousand years (I'd say several thousands but I don't want to underestimate humans). Progress in AI (or other sciences), to me seems more like an Evolutionary Boom, followed by long periods of nothing. Basically lots of information about how 'X' works piles up on a floor until someone or something picks it up and sorts it into a neat pile that explains lots of stuff, then more and more stuff piles up and now we need even smarter person(s) to sort that problem. I suspect that is because the pile of data to sort out is getting bigger and bigger and bigger, so that no single human can filter it any more. You now need teams and computers and God knows what else to make a breakthrough. If for breakthrough of same magnitude you need more smart people than you did before, means that problems we're dealing now got harder. Another thing that strikes me is that perhaps the thing that contributes to discovery of how Intelligence works isn't the same thing as intelligence. I.e. we're making our computers faster but not smarter. We assume that speed of thought is all, but what if there is a choke-point, like entropy? Chaos and randomness are essential to with creativity and discovering novel solutions. So basically we make our computers 20 times more intelligent, but they aren't chaotic enough to find what fixes them.
  - Houshalter 13y ago
    >If scientist were getting twice as fast, I doubt it would make a difference. As I said, ok scientist are twice as fast, but can they think of a novel solution twice as fast? What is the last great physics discovery that was by a single (wo)man? Why are there no medals in sciences now won mostly by teams? I think you are misunderstanding what "fast" means. Their brains would be literally running twice as fast, they would be having twice as many experiences and thoughts in the same amount of real time. They really would think of novel solutions twice as fast as they otherwise would. This isn't even taking into account increased parallelization, like adding more neurons to your brain or adding more people to work on the problem. >Again I don't think it won't reach super human levels. I just don't think the Singularity is within next thousand years (I'd say several thousands but I don't want to underestimate humans). I've heard pessimistic people estimate it at 100 years, and optimistic people estimate it within 20 years, but your prediction is way out there. The worst case is we have to examine the human brain and build a model of it in the computer (not easy but certainly within the realm of possibility.) Likely we'll figure out how to build intelligence long before then. >Chaos and randomness are essential to with creativity and discovering novel solutions. So basically we make our computers 20 times more intelligent, but they aren't chaotic enough to find what fixes them. Random number generators are simple enough, though randomness itself is not a good thing and does not help intelligence: http://lesswrong.com/lw/vp/worse_than_random/ http://lesswrong.com/lw/vp/worse_than_random/
    - Ygg2 13y ago
      I don't think IQ is same as speed of thought. I think IQ is defined ability to adapting to various environmental challenges. Again I don't think that having twice as many experiences is going to help. Sure you have twice as many experiences and twice as many thoughts but you still require time to sort those experiences into meaningful information. Basically if your computer was twice as fast at absorbing information, it means instead of X information it absorbs 2X information. Then when it process it it process 2X information and 2Y speed of processing information, you'd get same performance overall. Parallelization again yields only good results on problems that are parallelizable. Each node in a parallel network yields diminishing returns (i.e. The Mythical Man Month) because - surprise now there is more communication than processing. I mean this: http://en.wikipedia.org/wiki/Computational_creativity#Stephen_L._Thaler.27s_work_on_a_unified_model_of_creativity http://en.wikipedia.org/wiki/Computational_creativity#Stephe... Thaler's work on the subject indicate that creativity might be actually a byproduct of damage/noise done to neurons, i.e. imperfections and noise. So we make a perfect super fast computer with average intelligence that can't think of a solution to fix their predicament. I think we are 100 years from modelling a proper brain simulation with all forty ion canals and various other subsystems. Super human AI is probably farther than that. We right now (presumably) know all about the routes of the brain. But that is like saying that because you've been on all highways in Japan you have visited the entirety of Japan's mainland.

by Ygg2 13y ago

Houshalter 13y ago

The thing is, even if that's the case, each improvement still decrease the time it would take to get to the next improvement. Also computer hardware can be improved as well. Imagine Moore's law where instead of improving every 18 months, the scientists designing the computers are also getting twice as fast. Now imagine they are also getting smarter from improvements made by code optimizations. And then economic growth as well (more and more computers running the AIs or bigger factories to make them.) Even if progress does plateau, it doesn't have to happen anywhere near human scales. For example computers already have millions of times more serial processing ability than human brains. Any optimization that can take advantage of that, even if for only some problems and domains, would automatically give them a massive advantage over us.

Ygg2 13y ago
If scientist were getting twice as fast, I doubt it would make a difference. As I said, ok scientist are twice as fast, but can they think of a novel solution twice as fast? What is the last great physics discovery that was by a single (wo)man? Why are there no medals in sciences now won mostly by teams? Again I don't think it won't reach super human levels. I just don't think the Singularity is within next thousand years (I'd say several thousands but I don't want to underestimate humans). Progress in AI (or other sciences), to me seems more like an Evolutionary Boom, followed by long periods of nothing. Basically lots of information about how 'X' works piles up on a floor until someone or something picks it up and sorts it into a neat pile that explains lots of stuff, then more and more stuff piles up and now we need even smarter person(s) to sort that problem. I suspect that is because the pile of data to sort out is getting bigger and bigger and bigger, so that no single human can filter it any more. You now need teams and computers and God knows what else to make a breakthrough. If for breakthrough of same magnitude you need more smart people than you did before, means that problems we're dealing now got harder. Another thing that strikes me is that perhaps the thing that contributes to discovery of how Intelligence works isn't the same thing as intelligence. I.e. we're making our computers faster but not smarter. We assume that speed of thought is all, but what if there is a choke-point, like entropy? Chaos and randomness are essential to with creativity and discovering novel solutions. So basically we make our computers 20 times more intelligent, but they aren't chaotic enough to find what fixes them.
- Houshalter 13y ago
  >If scientist were getting twice as fast, I doubt it would make a difference. As I said, ok scientist are twice as fast, but can they think of a novel solution twice as fast? What is the last great physics discovery that was by a single (wo)man? Why are there no medals in sciences now won mostly by teams? I think you are misunderstanding what "fast" means. Their brains would be literally running twice as fast, they would be having twice as many experiences and thoughts in the same amount of real time. They really would think of novel solutions twice as fast as they otherwise would. This isn't even taking into account increased parallelization, like adding more neurons to your brain or adding more people to work on the problem. >Again I don't think it won't reach super human levels. I just don't think the Singularity is within next thousand years (I'd say several thousands but I don't want to underestimate humans). I've heard pessimistic people estimate it at 100 years, and optimistic people estimate it within 20 years, but your prediction is way out there. The worst case is we have to examine the human brain and build a model of it in the computer (not easy but certainly within the realm of possibility.) Likely we'll figure out how to build intelligence long before then. >Chaos and randomness are essential to with creativity and discovering novel solutions. So basically we make our computers 20 times more intelligent, but they aren't chaotic enough to find what fixes them. Random number generators are simple enough, though randomness itself is not a good thing and does not help intelligence: http://lesswrong.com/lw/vp/worse_than_random/ http://lesswrong.com/lw/vp/worse_than_random/
  - Ygg2 13y ago
    I don't think IQ is same as speed of thought. I think IQ is defined ability to adapting to various environmental challenges. Again I don't think that having twice as many experiences is going to help. Sure you have twice as many experiences and twice as many thoughts but you still require time to sort those experiences into meaningful information. Basically if your computer was twice as fast at absorbing information, it means instead of X information it absorbs 2X information. Then when it process it it process 2X information and 2Y speed of processing information, you'd get same performance overall. Parallelization again yields only good results on problems that are parallelizable. Each node in a parallel network yields diminishing returns (i.e. The Mythical Man Month) because - surprise now there is more communication than processing. I mean this: http://en.wikipedia.org/wiki/Computational_creativity#Stephen_L._Thaler.27s_work_on_a_unified_model_of_creativity http://en.wikipedia.org/wiki/Computational_creativity#Stephe... Thaler's work on the subject indicate that creativity might be actually a byproduct of damage/noise done to neurons, i.e. imperfections and noise. So we make a perfect super fast computer with average intelligence that can't think of a solution to fix their predicament. I think we are 100 years from modelling a proper brain simulation with all forty ion canals and various other subsystems. Super human AI is probably farther than that. We right now (presumably) know all about the routes of the brain. But that is like saying that because you've been on all highways in Japan you have visited the entirety of Japan's mainland.