6 ms·

Watson is what I think is called ANI (Artificial Narrow Intelligence). There is a lot of things we need to figure out to move from a narrow intelligence to a ge

by watson 11y ago

Watson is what I think is called ANI (Artificial Narrow Intelligence). There is a lot of things we need to figure out to move from a narrow intelligence to a general intelligence (AGI) and then (quickly) to a super intelligence (ASI). The big question of cause is how far we are from AGI - i.e. an intelligence on par with a human. No one knows of cause but a lot of smart people say there is a good chance it's happening in around 30 years. It's a development I for one will be following closely - with both fear and excitement ;)

jacquesm 11y ago
> No one knows of cause but a lot of smart people say there is a good chance it's happening in around 30 years. I've been hearing that since I started using computers, 36 years ago.
- ExpiredLink 11y ago
  Old wine in old bottles. Why the hype again?
- jsolson 11y ago
  To be fair, today we have what we believe to be (much) more accurate models of how much computation a human mind is capable of and how long it will take to build computing machines operating at that scale. 36 years ago the argument that AGI was coming soon could be made in tandem with the argument that we'd make some fundamental advance that allowed computers to express intelligence with less computational capacity than humans (by orders of magnitude). Today we can make an argument that we'll achieve it (at least initially) by leveraging computational capacity on par with or orders of magnitude greater than a human mind.
  - jacquesm 11y ago
    > To be fair, today we have what we believe to be (much) more accurate models of how much computation a human mind is capable of How much, yes, how, not so much and definitely not at the powerbudget the brain has. > and how long it will take to build computing machines operating at that scale. We don't actually know that. There have been some WAGs but so far those appeared to be totally off based on the developments since. > 36 years ago the argument that AGI was coming soon could be made in tandem with the argument that we'd make some fundamental advance that allowed computers to express intelligence with less computational capacity than humans (by orders of magnitude). Yes, that was a crucial mistake and it led directly to the AI winter. > Today we can make an argument that we'll achieve it (at least initially) by leveraging computational capacity on par with or orders of magnitude greater than a human mind. Chances are that we're missing a very important piece of the puzzle for which there is no known solution even in theory. The problem is that there are many candidates for that important piece none of which have currently proposed workable solutions no matter what the computational budget or the accepted slowdown (they are equivalent). So I think some caution when throwing around projections numbering 'just a couple of years' is warranted, after all it's 'merely a matter of programming' but in this case we don't have a working model that we understand.
  - zachrose 11y ago
    Maybe somebody who knows about AI can answer this for me: what do we expect of an AGI machine to be able to do when it "turns on?" How will you know it works? To compare to a human, you might say that it can cry, suck milk, and learn to walk and talk over the span of a few years, and then slowly, with a lot of trial and error, learn to do more.
- david927 11y ago
  Sure, because "30 years" is code for, "far enough in the future to obviate that nothing we see today would predicate such a jump -- but still, it will happen." The timelines are silly; the prediction is not.
  - jacquesm 11y ago
    It's the timeline I have a problem with. Any timeline with just two digits in it for AGI would do well to address the steps required to divide those years up into discrete steps of what bits and pieces will be accomplished by what (say) 5 year interval. Saving all the hard bits for the last 5 years would be considered cheating. For now - as far as I can see - we are no closer to the goal than where we were 35 years ago but we know better how much we are still missing and all the hard parts are still in front of us.
    - Houshalter 11y ago
      Predicting AI progress isn't like predicting other things were you can lay out all the steps that need to be done and make a plan. It's more like solving a series of hard technical problems. No one knows what the solutions will be or how long it will take someone to invent them. There were perfectly educated people that said there would never be flying machines, 2 years before the Wright brothers succeeded.
      - jacquesm 11y ago
        That's exactly it. This is like saying there will be flying machines in 30 years after seeing a bird for the first time without even realizing that you need to know about aerodynamics and that you need a technology based society to support the development of engines as the prime mover before you can have powered flight. It's not about saying it can't be done it is all about saying it can be done within a specific time-frame. Key inventions don't happen 'on command' unless the only thing required is a brute force search for something that is already possible in principle (say electric light).
api 11y ago
You are assuming that ANI is anything like AGI.
- watson 11y ago
  No I know it's not even close. I might be foolish and I might mix up some numbers (here it's closer to 50 than 30 years) - but when reading papers like this I can't help feeling slightly optimistic: http://www.nickbostrom.com/papers/survey.pdf http://www.nickbostrom.com/papers/survey.pdf
  - api 11y ago
    Estimate surveys have a poor record. Technologies tend to be perpetually 10-30 years away until they are 0 years away, or never happen at all. ANI and AGI are qualitatively different. Progress in making cars faster will not lead to teleportation or warp drive. It's not even applicable. There are many unknown unknowns in AGI. I could not even pretend to give you an estimate.
anonymfus 11y ago
I think that there are more dimensions than such classification implies. "General" naturally means that it can be applied to any intellectual task (dimension of generality), but "super" means that it can solve them faster/better than humans (dimension(s) of speed and quality). Looks like slow general and fast narrow AIs already exist, so realistically the thing people call AGI (AI at least as general and as fast as human mind at anything) will also be ASI (faster/better than human mind at something). Russian classification traditionally also have distinction between terms "artificial intelligence" and "artificial mind" ("искусственный разум") to keep the problems of autonomy/agency/consciousness out of field of AI.

by watson 11y ago

jacquesm 11y ago

> No one knows of cause but a lot of smart people say there is a good chance it's happening in around 30 years. I've been hearing that since I started using computers, 36 years ago.

ExpiredLink 11y ago
Old wine in old bottles. Why the hype again?
jsolson 11y ago
To be fair, today we have what we believe to be (much) more accurate models of how much computation a human mind is capable of and how long it will take to build computing machines operating at that scale. 36 years ago the argument that AGI was coming soon could be made in tandem with the argument that we'd make some fundamental advance that allowed computers to express intelligence with less computational capacity than humans (by orders of magnitude). Today we can make an argument that we'll achieve it (at least initially) by leveraging computational capacity on par with or orders of magnitude greater than a human mind.
- jacquesm 11y ago
  > To be fair, today we have what we believe to be (much) more accurate models of how much computation a human mind is capable of How much, yes, how, not so much and definitely not at the powerbudget the brain has. > and how long it will take to build computing machines operating at that scale. We don't actually know that. There have been some WAGs but so far those appeared to be totally off based on the developments since. > 36 years ago the argument that AGI was coming soon could be made in tandem with the argument that we'd make some fundamental advance that allowed computers to express intelligence with less computational capacity than humans (by orders of magnitude). Yes, that was a crucial mistake and it led directly to the AI winter. > Today we can make an argument that we'll achieve it (at least initially) by leveraging computational capacity on par with or orders of magnitude greater than a human mind. Chances are that we're missing a very important piece of the puzzle for which there is no known solution even in theory. The problem is that there are many candidates for that important piece none of which have currently proposed workable solutions no matter what the computational budget or the accepted slowdown (they are equivalent). So I think some caution when throwing around projections numbering 'just a couple of years' is warranted, after all it's 'merely a matter of programming' but in this case we don't have a working model that we understand.
- zachrose 11y ago
  Maybe somebody who knows about AI can answer this for me: what do we expect of an AGI machine to be able to do when it "turns on?" How will you know it works? To compare to a human, you might say that it can cry, suck milk, and learn to walk and talk over the span of a few years, and then slowly, with a lot of trial and error, learn to do more.
david927 11y ago
Sure, because "30 years" is code for, "far enough in the future to obviate that nothing we see today would predicate such a jump -- but still, it will happen." The timelines are silly; the prediction is not.
- jacquesm 11y ago
  It's the timeline I have a problem with. Any timeline with just two digits in it for AGI would do well to address the steps required to divide those years up into discrete steps of what bits and pieces will be accomplished by what (say) 5 year interval. Saving all the hard bits for the last 5 years would be considered cheating. For now - as far as I can see - we are no closer to the goal than where we were 35 years ago but we know better how much we are still missing and all the hard parts are still in front of us.
  - Houshalter 11y ago
    Predicting AI progress isn't like predicting other things were you can lay out all the steps that need to be done and make a plan. It's more like solving a series of hard technical problems. No one knows what the solutions will be or how long it will take someone to invent them. There were perfectly educated people that said there would never be flying machines, 2 years before the Wright brothers succeeded.
    - jacquesm 11y ago
      That's exactly it. This is like saying there will be flying machines in 30 years after seeing a bird for the first time without even realizing that you need to know about aerodynamics and that you need a technology based society to support the development of engines as the prime mover before you can have powered flight. It's not about saying it can't be done it is all about saying it can be done within a specific time-frame. Key inventions don't happen 'on command' unless the only thing required is a brute force search for something that is already possible in principle (say electric light).

api 11y ago

You are assuming that ANI is anything like AGI.

watson 11y ago
No I know it's not even close. I might be foolish and I might mix up some numbers (here it's closer to 50 than 30 years) - but when reading papers like this I can't help feeling slightly optimistic: http://www.nickbostrom.com/papers/survey.pdf http://www.nickbostrom.com/papers/survey.pdf
- api 11y ago
  Estimate surveys have a poor record. Technologies tend to be perpetually 10-30 years away until they are 0 years away, or never happen at all. ANI and AGI are qualitatively different. Progress in making cars faster will not lead to teleportation or warp drive. It's not even applicable. There are many unknown unknowns in AGI. I could not even pretend to give you an estimate.

anonymfus 11y ago

I think that there are more dimensions than such classification implies. "General" naturally means that it can be applied to any intellectual task (dimension of generality), but "super" means that it can solve them faster/better than humans (dimension(s) of speed and quality). Looks like slow general and fast narrow AIs already exist, so realistically the thing people call AGI (AI at least as general and as fast as human mind at anything) will also be ASI (faster/better than human mind at something). Russian classification traditionally also have distinction between terms "artificial intelligence" and "artificial mind" ("искусственный разум") to keep the problems of autonomy/agency/consciousness out of field of AI.