3 ms·

I sympathize with your position. Kurzweil seems to regularly makes strong remarks that shock me in their assertive naivety. However, our species is incredibly

I sympathize with your position. Kurzweil seems to regularly makes strong remarks that shock me in their assertive naivety. However, our species is incredibly poor at intuitively understanding exponential trends. Kurzweil's essential point is that our gut-feeling predictions of the future will be off substantially. Specifically, our intuitive sense of future progress significantly underestimates actual progress and the gap systematically widens as time progresses. Neither you or I could have intuitively predicted twenty years ago the level of today's technological advancement. And we'd be even more wrong in the predicting the next twenty. It's far too easy to dismiss Kurzweil's work as that of a superficial nutjob who wants to "sell more books and speaking engagements." Kuzweil's crunched the numbers. While he might not be right on everything, his predictions are based on data, but yours are based on human intuition that has historically been wrong. I would be very careful of the knee-jerk reaction you display to Kurzweil's work. Keep in mind we are cognitively disabled at intuitively predicting exponential progress.

Retric 15y ago
The problem is not all tasks are of linear difficulty. If building a 1% more intelligent AI takes 1% more computational power then he might have a point. However if building a 1% more intelligent AI takes 2x the computational power then you don't end up with a singularity because you don't get a strong feedback loop. Consider starting at the low end of the scale booting windows 7 on a 2KHz CPU is about twice as fast as a 1KHz CPU and all things being equal faster processors let you get more things done but having a 1GHz vs 2GHz CPU has little effect on boot times. We might be able to build a 10x human level intelligence but if that can only build a 11x human intelligence and that can only build a 11.1x human intelligence you quickly hit a wall. Or for a more natural looking progression (10x, 15x, 18.75x, 21.09x, 22.4x). x sub 1 = 10; x(sub n) = x(n-1) * (2^n+1)/ (2^n)
- MatthewPhillips 15y ago
  The point of TS is that once you get to 1.0001x smarter than human you get to wall really quickly. If the wall is 2x or 11x, self-improving AI will reach it in no time. And we'll still have tremendous benefits from it, no matter how close the wall is to human intelligence. AI only need be a tiny bit smarter than humans to achieve things such as interstellar travel, eternal life, etc.
  - Retric 15y ago
    Slightly more intelligent than human is not the same as magic. Self improving AI is a nice theory but it assumes you can get better at everything at the same time without tradeoffs and that takes better hardware not just software. Science fiction is full of magic masquerading as technology, but it ignores the possibility that reality may not support that capability. Predicting the future requires more than simply extrapolating trends it also requires understanding physical limitations, early engines saw rapid increases in both HP and efficiency but that trend slowed down for a reason. EX: Literal eternal life requires a universe that is capable of lasting that long. If the fundamental laws of physics limit either the age of the universe or the amount of useful work that can be done then you at some point everything dies. So no we are down from eternal to just really long life. Also, 1.0001x a smart as a human is not really significant. 1,000 AI's 1.0001x smarter than a human is not going to be able to say beat 1 grand master at go without training or hundreds of years of preparation. I have no problem saying AI get's to quickly build better software, but building hardware at scale takes time.
Dn_Ab 15y ago
Data can be manipulated to confirm biases - hence the flaw of averages and jokes about 3 types of lies. And human intuition has also led to discoveries. Consider that Poincare felt logic was inferior to intuition. Point being just because someone uses data and another uses intuition doesn't inherently make the intuitive wrong. The intuition could be built from a large experiential base with a model whose generative process is much more accurate than a naive application of data. Also see http://en.wikipedia.org/wiki/Technological_singularity#Criticism_of_the_accelerating_returns_argument http://en.wikipedia.org/wiki/Technological_singularity#Criti...
arethuza 15y ago
"Kuzweil's crunched the numbers." From what I have seen only in the most superficial and optimistic sense - long term technological prediction is hardly a recognized scientific field.
- DavidChouinard 15y ago
  I was skeptic at first too. Read the first few dozen pages of The Singularity Is Near and you'll be impressed the quality of the data. By almost all accounts, technological progress datapoints match very closely an exponential curve. Here's a good mashup from a number of third party sources: http://en.wikipedia.org/wiki/File:ParadigmShiftsFrr15Events.svg http://en.wikipedia.org/wiki/File:ParadigmShiftsFrr15Events.... This is hardly a wishy washy thesis by a history major. The data are pretty robust.

Retric 15y ago

The problem is not all tasks are of linear difficulty. If building a 1% more intelligent AI takes 1% more computational power then he might have a point. However if building a 1% more intelligent AI takes 2x the computational power then you don't end up with a singularity because you don't get a strong feedback loop. Consider starting at the low end of the scale booting windows 7 on a 2KHz CPU is about twice as fast as a 1KHz CPU and all things being equal faster processors let you get more things done but having a 1GHz vs 2GHz CPU has little effect on boot times. We might be able to build a 10x human level intelligence but if that can only build a 11x human intelligence and that can only build a 11.1x human intelligence you quickly hit a wall. Or for a more natural looking progression (10x, 15x, 18.75x, 21.09x, 22.4x). x sub 1 = 10; x(sub n) = x(n-1) * (2^n+1)/ (2^n)

MatthewPhillips 15y ago
The point of TS is that once you get to 1.0001x smarter than human you get to wall really quickly. If the wall is 2x or 11x, self-improving AI will reach it in no time. And we'll still have tremendous benefits from it, no matter how close the wall is to human intelligence. AI only need be a tiny bit smarter than humans to achieve things such as interstellar travel, eternal life, etc.
- Retric 15y ago
  Slightly more intelligent than human is not the same as magic. Self improving AI is a nice theory but it assumes you can get better at everything at the same time without tradeoffs and that takes better hardware not just software. Science fiction is full of magic masquerading as technology, but it ignores the possibility that reality may not support that capability. Predicting the future requires more than simply extrapolating trends it also requires understanding physical limitations, early engines saw rapid increases in both HP and efficiency but that trend slowed down for a reason. EX: Literal eternal life requires a universe that is capable of lasting that long. If the fundamental laws of physics limit either the age of the universe or the amount of useful work that can be done then you at some point everything dies. So no we are down from eternal to just really long life. Also, 1.0001x a smart as a human is not really significant. 1,000 AI's 1.0001x smarter than a human is not going to be able to say beat 1 grand master at go without training or hundreds of years of preparation. I have no problem saying AI get's to quickly build better software, but building hardware at scale takes time.

Dn_Ab 15y ago

Data can be manipulated to confirm biases - hence the flaw of averages and jokes about 3 types of lies. And human intuition has also led to discoveries. Consider that Poincare felt logic was inferior to intuition. Point being just because someone uses data and another uses intuition doesn't inherently make the intuitive wrong. The intuition could be built from a large experiential base with a model whose generative process is much more accurate than a naive application of data. Also see http://en.wikipedia.org/wiki/Technological_singularity#Criticism_of_the_accelerating_returns_argument http://en.wikipedia.org/wiki/Technological_singularity#Criti...

arethuza 15y ago

"Kuzweil's crunched the numbers." From what I have seen only in the most superficial and optimistic sense - long term technological prediction is hardly a recognized scientific field.

DavidChouinard 15y ago
I was skeptic at first too. Read the first few dozen pages of The Singularity Is Near and you'll be impressed the quality of the data. By almost all accounts, technological progress datapoints match very closely an exponential curve. Here's a good mashup from a number of third party sources: http://en.wikipedia.org/wiki/File:ParadigmShiftsFrr15Events.svg http://en.wikipedia.org/wiki/File:ParadigmShiftsFrr15Events.... This is hardly a wishy washy thesis by a history major. The data are pretty robust.