The AI Singularity
Since the idea of computers and robots was conceived by some geek in a cave people have been talking about the day that computers would be smarter than us. Some have written books about the utopia that will exist when the robots meet our every desire without us ever having to work again. Others talk of the day when the computers fight back and try to destroy the inefficient humans. But in all cases the critical moment comes when computers start thinking for themselves.
But what does it mean for a computer to "think"? All existing and future computers will always just run a set of instructions through a glorified calculator. You can add some randomness to make it seem less formulaic, but there's still a program underneath it somewhere.
In the early days of computing there were hopes that some really smart guy would write the AI program that made the computer "think". As time passed many nerds got many PhDs describing why it was so hard, but no one made anything that even slightly resembled thinking. Even today there are lots of engineers trying to make "neural networks" and "adaptive algorithms" that they hope will be the AI program that makes computers think.
So what will we do when we get this amazing program. Well, anything we currently spend brainpower trying to do, the harder the better. Interestingly, one of the hardest problems we look at today is making a better AI. What if a program could make a better version of itself? What would we do with that better version? We'd use it to make an even better version, of course.
This is where it gets scary... Today, our advances in AI are limited by the rate at which humans can improve their algorithms. Our tools (i.e. computers) get better all the time, and so does our state of research. But it still takes months or years to achieve notable improvements. What happens when a program can compile a better version of itself in a few days or a few hours?
Researchers call it The Singularity; the day our rate of research in AI improves by a few orders of magnitude. Suddenly our computers are getting dramatically better every day. A human goes to school for years to learn the basics of knowledge, even longer to become an expert. A computer program could replicate itself thousands of times in a day, and each would have the full expertise of the original.
Even tasks that are better suited for human thought than machine algorithms would quickly be switched to hardware when thousands of AI experts can work on a problem 24 hours a day for the cost of a computer and some electricity. Humans will never work again (or we'll be enslaved by the robots depending on your optimism). Our rate of scientific discovery will grow by orders of magnitude overnight... The world will never be the same...
So when will this first true AI get written? When is the singularity? Is this all hogwash? Well, it's certainly possible that making a real AI is harder than any of us can imagine, or it might even get proven impossible by some genius of the millennium. Then again, the program might already have been written, and is just in the process of getting itself into a good first working version.
Since the first computers people have been predicting AI is just around the corner, but I predict that the day is going to come soon. After all it's not that the computer needs to become a perfect thinker, it just has to get better at designing AI than us. We've had over 60 years with computers and we haven't achieved a half decent AI yet. That isn't such a high bar for the AI to beat...
But what does it mean for a computer to "think"? All existing and future computers will always just run a set of instructions through a glorified calculator. You can add some randomness to make it seem less formulaic, but there's still a program underneath it somewhere.
In the early days of computing there were hopes that some really smart guy would write the AI program that made the computer "think". As time passed many nerds got many PhDs describing why it was so hard, but no one made anything that even slightly resembled thinking. Even today there are lots of engineers trying to make "neural networks" and "adaptive algorithms" that they hope will be the AI program that makes computers think.
So what will we do when we get this amazing program. Well, anything we currently spend brainpower trying to do, the harder the better. Interestingly, one of the hardest problems we look at today is making a better AI. What if a program could make a better version of itself? What would we do with that better version? We'd use it to make an even better version, of course.
This is where it gets scary... Today, our advances in AI are limited by the rate at which humans can improve their algorithms. Our tools (i.e. computers) get better all the time, and so does our state of research. But it still takes months or years to achieve notable improvements. What happens when a program can compile a better version of itself in a few days or a few hours?
Researchers call it The Singularity; the day our rate of research in AI improves by a few orders of magnitude. Suddenly our computers are getting dramatically better every day. A human goes to school for years to learn the basics of knowledge, even longer to become an expert. A computer program could replicate itself thousands of times in a day, and each would have the full expertise of the original.
Even tasks that are better suited for human thought than machine algorithms would quickly be switched to hardware when thousands of AI experts can work on a problem 24 hours a day for the cost of a computer and some electricity. Humans will never work again (or we'll be enslaved by the robots depending on your optimism). Our rate of scientific discovery will grow by orders of magnitude overnight... The world will never be the same...
So when will this first true AI get written? When is the singularity? Is this all hogwash? Well, it's certainly possible that making a real AI is harder than any of us can imagine, or it might even get proven impossible by some genius of the millennium. Then again, the program might already have been written, and is just in the process of getting itself into a good first working version.
Since the first computers people have been predicting AI is just around the corner, but I predict that the day is going to come soon. After all it's not that the computer needs to become a perfect thinker, it just has to get better at designing AI than us. We've had over 60 years with computers and we haven't achieved a half decent AI yet. That isn't such a high bar for the AI to beat...
posted by Ze Ace at
3:38 PM
1 Comments:
It cant be too long - moore's law will end in another 10 years approx. At this time we will have 64core CPU's running at around 10Ghz. I am writing Image processing routines on a 2ghz single core and getting rudimentary shape tracking & recognition for usefull feedback for autonomous robot guidance in real time etc and my code is not even properly optimised or well written (other much better software is out ther ) . With a 100 fold increase in CPU power i could implement a vision processing pipeline equal in complexity to the human visual system, just a matter of more methods and features being added the right way. Since my code took me 4 months part-time and just me working on it , You can be sure that a core team of say 10 people fulltime over 10 years would easily have a visual system that equals the human visual capability we take for granted. Add teams for other things like movement , sound , survival and replication strategy. And around 10 years from now one can expect machine species fully capable of self sufficiency just like any other species. So we can all look forward to saying hello to AI pretty soon.
By
Anonymous, at 10/14/2007 4:39 AM
Post a Comment
<< Home