Normally I would find this an absolutely paranoid remark, but according to ZDNet Microsoft paid about $40 mln for bungie, which is (in my opinion) a bargain for a developper that
A) Produced several known titles already
B) has a large enough infrastructure to be producing two 3d shooters of which one is promissing to be the hit of the year.
If I was selling out to MS I'd at least make it worthwhile for my soon to be assimilated employees.:P
Anyone with a little more financial perspective care to disagree ?
Penrose's argument using Goedel's theorem may be consistent in the context where all machine's can be described by turing machine. However there is one important feature that human brains have and some machines may emulate, but is not part of turing's machine theory: true parallelism. This abbillety allows the thinking entity to reflect on itself and it's own thinking process, thus avoiding (not solving) the halting problem. Furthermore Penrose's alternative desciption of computation (using quantum machanics) is absurd to say the least.
>I wonder if anyone's simulated this type of >"behaviour" on a (digital) computer? Imagine if >such a program, responding to external stimulii >could reprogram itself in response and "grow >more circuits" ie learn. Actually the only thing kinda like revolutionary to this is that they did it without simulating it in a digital computer. Systems that emulate neural behaviours are being studied for several decades now at AI labs and universities. One special point of interest has been the growing of neural networks that you mentioned, but the results are not very encouraging until now. Just the last couple of years the growing of neural networks using genetic alghoritms has produced some results that actually solved some problems, but simulating evolution may take some time... (Like 5 billion years or so) >Perhaps we have the first tentative steps to >passing the Turing test? Not likely that this will come from this side of AI research, from a bug to a human is a long way. Some Results in this direction are soon to be expected from the language side of AI. (I hope) >Time for humans to sit back and let the machines >do all the work. >Well, maybe not. You were probably sitting back behind your computer when you wrote this:)
Although many researchers are tackling this problem (neural nets, HMMs), any real learning is still far away. Currently, computers are told by us humans what is right and what is wrong
This is not entirely true especially the part about NN's. They can be trained in two ways: 1 using supervised learning 2 using unsupervised learning(duh). These two methods form the major part of learning in all mammals, altough the later seems to be the more intelligent(=human) one. It's basically learning from your mistakes. (Hey if I run head first into a wall it hurts) and generalising it (maybe i shouldn't run into oak trees either). The unsupervised learning method is computationally very intensive especially for processor based computers, since they have to emulate the mechanisms used by parallel computers (like that thing in your head, you know, the one that hurts when you bang it against that wall:) ) Which brings me to the subject of the original post, speed is not the only factor, well yes it is when you need to emulate things. The only way to recreate a brain is to grow it out of dna, which is ok, but ethical questions arise when you stuff a piece of brain in your proxy server and tell it to remember all the pages you went to.
Slashdot Mirror
Normally I would find this an absolutely paranoid remark, but according to
:P
ZDNet Microsoft paid about $40 mln for bungie, which is (in my opinion) a bargain for a developper that
A) Produced several known titles already
B) has a large enough infrastructure to be producing two 3d shooters of which one is promissing to be the hit of the year.
If I was selling out to MS I'd at least make it worthwhile for my soon to be assimilated employees.
Anyone with a little more financial perspective care to disagree ?
Penrose's argument using Goedel's theorem may be consistent in the context where all machine's can be described by turing machine. However there is one important feature that human brains have and some machines may emulate, but is not part of turing's machine theory: true parallelism. This abbillety allows the thinking entity to reflect on itself and it's own thinking process, thus avoiding (not solving) the halting problem. Furthermore Penrose's alternative desciption of computation (using quantum machanics) is absurd to say the least.
>I wonder if anyone's simulated this type of >"behaviour" on a (digital) computer? Imagine if >such a program, responding to external stimulii >could reprogram itself in response and "grow >more circuits" ie learn. :)
Actually the only thing kinda like revolutionary to this is that they did it without simulating it in a digital computer. Systems that emulate neural behaviours are being studied for several decades now at AI labs and universities. One special point of interest has been the growing of neural networks that you mentioned, but the results are not very encouraging until now. Just the last couple of years the growing of neural networks using genetic alghoritms has produced some results that actually solved some problems, but simulating evolution may take some time... (Like 5 billion years or so)
>Perhaps we have the first tentative steps to >passing the Turing test?
Not likely that this will come from this side of AI research, from a bug to a human is a long way. Some Results in this direction are soon to be expected from the language side of AI. (I hope)
>Time for humans to sit back and let the machines >do all the work.
>Well, maybe not.
You were probably sitting back behind your computer when you wrote this
Although many researchers are tackling this problem (neural nets, HMMs), any real learning is still far away. Currently,
:) )
computers are told by us humans what is right and what is wrong
This is not entirely true especially the part about NN's. They can be trained in two ways:
1 using supervised learning
2 using unsupervised learning(duh).
These two methods form the major part of learning in all mammals, altough the later seems to be the more intelligent(=human) one. It's basically learning from your mistakes. (Hey if I run head first into a wall it hurts) and generalising it (maybe i shouldn't run into oak trees either).
The unsupervised learning method is computationally very intensive especially for processor based computers, since they have to emulate the mechanisms used by parallel computers (like that thing in your head, you know, the one that hurts when you bang it against that wall
Which brings me to the subject of the original post, speed is not the only factor, well yes it is when you need to emulate things. The only way to recreate a brain is to grow it out of dna, which is ok, but ethical questions arise when you stuff a piece of brain in your proxy server and tell it to remember all the pages you went to.