Too cheap to meter
In
the news a few days ago I was happy to see that Ontario has decided to
build new nuclear reactors to solve the province's energy crunch. As
expected, all kinds of "concerned activists" (and we know what that is
code for) are already out in force, swearing to fight this decision.
Well, at least as long as the coal and petroleum industries will keep
bankrolling their efforts. But at least there is some hope for the
future of this great province and its people.
Speaking of the future, as I mentioned earlier, I recently read Ray Kurzweil's new tome "Singularity Is Near". The basic content of this book is the prediction that massive improvements in information and physical technology will take place within the next few decades, and that these improvements will change the very human life in ways that are virtually unimaginable today. As technology progresses in an ever-increasing pace as better technology can always be used to create the next generation of even better technology, we will eventually reach the technological singularity, at which point all bets are off. Kurzweil is also well-known for his passion for life extension, so this book goes to great lengths to discuss how genetic engineering and nanotechnology could be used to rebuild and improve human body so that we would no longer be slaves of our ancestral DNA that evolved for a totally different primitive world and wasn't "designed" by the blind evolutionary forces to be optimal for our current needs, life expentancy and pleasure. (Once again, the artificial option will be vastly superior to natural, when technology reaches the level needed to offer it.)
The book spends quite a few pages discussing nanotechnology. It also basically takes for granted the eventual appearance of human-level artificial intelligence within a few decades, which will then soon be followed by a superhuman intelligence, and eventually a planet-level intelligence that is as incomprehensibly smart next to us as we are next to ants. Myself, I am much more skeptical about the possibilities and future of artificial intelligence, in the sense of human-intelligent machines that this word is typically used. If we define the term "artificial intelligence" so that it means algorithms for representing knowledge and choosing optimal actions, especially in the level of complexity and difficulty that is just above what we currently know how to do, so that advances in the field gradually become part of algorithmics (for example, few people would say today that the minimax algorithm for playing chess or other board games is "artificial intelligence"), then artificial intelligence can be a meaningful field of research and it already has useful applications in business and games.
Since we don't really any kind of useful theory of computational intelligence of how knowledge might be represented and operated on intelligently with computational means, it wouldn't really help even if computers got a thousand times faster, the way Moore's law and Kurzweil predict will happen in about 15 years. We will then just be going nowhere thousand times as quickly. Of course, this objection mostly rules out the "good old-fashioned AI" that is based on formal operations with symbolic logic, but not the connectionist, emergent or other similarly hand-waving automagical forms of AI. On the other hand, it is hard to argue that if a future supercomputer simulated the behaviour of every neuron of the human brain (as is the goal of Blue Brain project), the result wouldn't be "intelligent" or "conscious" the way we usually define these terms. It probably wouldn't be very useful (after all, we already know how to make more than enough human brains), but at least it would be interesting. Even if this mind has no physical mouth, can it still scream?
Kurzweil's book starts with a general discussion of exponential growth, and presents a whole bunch of historical curves about how certain technological advances such as computing speed have improved over time. Drawn in a logarithmic scale, these curves are almost too linear, showing that the underlying progress is steadily exponential. Since unlimited exponential growth is impossible in limited space, the growth eventually turns into a S-curve, but Kurzweil claims that these S-curves will follow each other so that each S is tighter and rises higher than the previous one, together forming an exponential growth curve that will lead to singularity.
One thing that was immediately evident in these curves was the dizzying thought that during the vast majority of human existence, everything in life has been pretty horrible for even the richest kings. Everybody just toiled and moiled away while no visible progress whatsoever happened or could even be imagined in their rigid society where everything just remained as bad it always was. Among the billions of people who have ever lived (and the pre-humans who came before them, all the way back to the countless single-cell organisms that we evolved from), all doomed to their unimaginably wretched misery, I somehow had the tremendous luck to be born me in this time at the End of History at the visible knee of the exponential curve, where everything just keeps getting better. I mean, Jesus, what are the odds? Vanishingly small, certainly.
Of course, the pessimist in me will immediately predict that I will probably be so unlucky as that I will die a day before the indefinite life extension technology becomes widely available. So I won't get to be one of the lucky ones who get to enjoy the World of Tomorrow, a land of infinite abundance where the life of the average person would be as incomprehensible for me as my life would be to people hundred years ago.
Speaking of which, I have never really been able to understand this notion of consciousness uploading, so perhaps somebody who knows better could clarify it to me. If the essential structure of my brain is copied into a computer simulation in which Simulated Ilkka lives in some kind of Matrix, how exactly does that make my life better, or allow me to live forever? Once I climb out of the machine that did this scanning and the monitor says "uploading complete" (and perhaps shows me images of how Simulated Ilkka is happily running around the Big Rock Candy Mountain or sipping a margarita at the Lipstick Lesbian Pool), I am still the same physical me as I was before the uploading, fated to the same destiny of gradually withering away.
The only way I could see consciousness uploading doing anything for me is if nanotech is used to gradually replace my brain cells with artificial cells that are functionally equivalent but will live forever. This would actually be neat, and since I am a functionalist instead of a vitalist, I wouldn't be the least afraid of losing my conscious mind at the point where enough biological brain cells have been replaced with their artificial functional equivalents. I remember when I once went to the dentist, and the dentist drilled into a cavity and replaced the weak organic tooth matter with hard ceramic substance that has none of the weaknesses of the frail biological matter. I sure wish there was some way to repeat this for every other part of my body. Dammit, get on to work with this project already, transhumanists!
In its concluding chapters, Kurzweil's book goes on to discuss and dismiss the famous risks that are supposedly present in genetic engineering, nanotechnology and robotics, although he predicts that nanotechnology and protections against it will be hot political issues as soon as in the twentytwenties. The last chapter then tackles various objections to the predictions that he is making. Of these objections, I would actually consider "The Criticism from Software" to be the most serious one: in essence, software engineering will be the bottleneck and not keep up with the complexity that these advances would require, and we will never be able to produce the threshold "seed AI" that would then go on to produce better AI systems indefinitely until the singularity. Second interesting piece of criticism, "The Criticism from Ontology", discusses John Searle's famous "Chinese Room" thought experiment and the ensuing argument that an uploaded/augmented/improved mind would not be conscious. Most of the other criticisms such as the totally irrelevant "The Criticism from the Church-Turing Thesis" or the patently ridiculous "The Criticism from the Rich-Poor Divide" were probably featured in this book mostly for their humour value.
Speaking of the future, as I mentioned earlier, I recently read Ray Kurzweil's new tome "Singularity Is Near". The basic content of this book is the prediction that massive improvements in information and physical technology will take place within the next few decades, and that these improvements will change the very human life in ways that are virtually unimaginable today. As technology progresses in an ever-increasing pace as better technology can always be used to create the next generation of even better technology, we will eventually reach the technological singularity, at which point all bets are off. Kurzweil is also well-known for his passion for life extension, so this book goes to great lengths to discuss how genetic engineering and nanotechnology could be used to rebuild and improve human body so that we would no longer be slaves of our ancestral DNA that evolved for a totally different primitive world and wasn't "designed" by the blind evolutionary forces to be optimal for our current needs, life expentancy and pleasure. (Once again, the artificial option will be vastly superior to natural, when technology reaches the level needed to offer it.)
The book spends quite a few pages discussing nanotechnology. It also basically takes for granted the eventual appearance of human-level artificial intelligence within a few decades, which will then soon be followed by a superhuman intelligence, and eventually a planet-level intelligence that is as incomprehensibly smart next to us as we are next to ants. Myself, I am much more skeptical about the possibilities and future of artificial intelligence, in the sense of human-intelligent machines that this word is typically used. If we define the term "artificial intelligence" so that it means algorithms for representing knowledge and choosing optimal actions, especially in the level of complexity and difficulty that is just above what we currently know how to do, so that advances in the field gradually become part of algorithmics (for example, few people would say today that the minimax algorithm for playing chess or other board games is "artificial intelligence"), then artificial intelligence can be a meaningful field of research and it already has useful applications in business and games.
Since we don't really any kind of useful theory of computational intelligence of how knowledge might be represented and operated on intelligently with computational means, it wouldn't really help even if computers got a thousand times faster, the way Moore's law and Kurzweil predict will happen in about 15 years. We will then just be going nowhere thousand times as quickly. Of course, this objection mostly rules out the "good old-fashioned AI" that is based on formal operations with symbolic logic, but not the connectionist, emergent or other similarly hand-waving automagical forms of AI. On the other hand, it is hard to argue that if a future supercomputer simulated the behaviour of every neuron of the human brain (as is the goal of Blue Brain project), the result wouldn't be "intelligent" or "conscious" the way we usually define these terms. It probably wouldn't be very useful (after all, we already know how to make more than enough human brains), but at least it would be interesting. Even if this mind has no physical mouth, can it still scream?
Kurzweil's book starts with a general discussion of exponential growth, and presents a whole bunch of historical curves about how certain technological advances such as computing speed have improved over time. Drawn in a logarithmic scale, these curves are almost too linear, showing that the underlying progress is steadily exponential. Since unlimited exponential growth is impossible in limited space, the growth eventually turns into a S-curve, but Kurzweil claims that these S-curves will follow each other so that each S is tighter and rises higher than the previous one, together forming an exponential growth curve that will lead to singularity.
One thing that was immediately evident in these curves was the dizzying thought that during the vast majority of human existence, everything in life has been pretty horrible for even the richest kings. Everybody just toiled and moiled away while no visible progress whatsoever happened or could even be imagined in their rigid society where everything just remained as bad it always was. Among the billions of people who have ever lived (and the pre-humans who came before them, all the way back to the countless single-cell organisms that we evolved from), all doomed to their unimaginably wretched misery, I somehow had the tremendous luck to be born me in this time at the End of History at the visible knee of the exponential curve, where everything just keeps getting better. I mean, Jesus, what are the odds? Vanishingly small, certainly.
Of course, the pessimist in me will immediately predict that I will probably be so unlucky as that I will die a day before the indefinite life extension technology becomes widely available. So I won't get to be one of the lucky ones who get to enjoy the World of Tomorrow, a land of infinite abundance where the life of the average person would be as incomprehensible for me as my life would be to people hundred years ago.
Speaking of which, I have never really been able to understand this notion of consciousness uploading, so perhaps somebody who knows better could clarify it to me. If the essential structure of my brain is copied into a computer simulation in which Simulated Ilkka lives in some kind of Matrix, how exactly does that make my life better, or allow me to live forever? Once I climb out of the machine that did this scanning and the monitor says "uploading complete" (and perhaps shows me images of how Simulated Ilkka is happily running around the Big Rock Candy Mountain or sipping a margarita at the Lipstick Lesbian Pool), I am still the same physical me as I was before the uploading, fated to the same destiny of gradually withering away.
The only way I could see consciousness uploading doing anything for me is if nanotech is used to gradually replace my brain cells with artificial cells that are functionally equivalent but will live forever. This would actually be neat, and since I am a functionalist instead of a vitalist, I wouldn't be the least afraid of losing my conscious mind at the point where enough biological brain cells have been replaced with their artificial functional equivalents. I remember when I once went to the dentist, and the dentist drilled into a cavity and replaced the weak organic tooth matter with hard ceramic substance that has none of the weaknesses of the frail biological matter. I sure wish there was some way to repeat this for every other part of my body. Dammit, get on to work with this project already, transhumanists!
In its concluding chapters, Kurzweil's book goes on to discuss and dismiss the famous risks that are supposedly present in genetic engineering, nanotechnology and robotics, although he predicts that nanotechnology and protections against it will be hot political issues as soon as in the twentytwenties. The last chapter then tackles various objections to the predictions that he is making. Of these objections, I would actually consider "The Criticism from Software" to be the most serious one: in essence, software engineering will be the bottleneck and not keep up with the complexity that these advances would require, and we will never be able to produce the threshold "seed AI" that would then go on to produce better AI systems indefinitely until the singularity. Second interesting piece of criticism, "The Criticism from Ontology", discusses John Searle's famous "Chinese Room" thought experiment and the ensuing argument that an uploaded/augmented/improved mind would not be conscious. Most of the other criticisms such as the totally irrelevant "The Criticism from the Church-Turing Thesis" or the patently ridiculous "The Criticism from the Rich-Poor Divide" were probably featured in this book mostly for their humour value.
I guess the problem is the same as with the duplicate Voyager on the eponymous TV serial. One of the Voyagers survives, one is destroyed when Janeway activates the self-destruction sequence. Which one is the real Voyager? What about a dying Trill who has been disconnected from the symbiont? Where does the consciousness actually reside?
The only thing that keeps you connected to your former self is your memory. What will happen when the virtual Ilkka in the Matrix and the real Ilkka share the same memories up to the point of the upload, but part company after that? How will it feel?
Posted by Panu | 12:26 PM
I think an Immortal would for some considerable time be destroyable by extreme blowing apart violence. Would he be known to be immortal, he would at once be torn to pieces by the mob, so he must keep his nature secret. Some say this has already happened.
Posted by Catilina | 2:15 PM
Re: 'the Chinese room argument':
Who wouldn't want a computer with John Searle sitting inside it? Unfortunately, such a design has no practicality, as there is only one John Searle.
Instead of Searle, we should get Chinese people to sit in the computers: they're plentiful and mostly small, which would result in the computers being more compact.
Posted by C. Van Carter | 2:16 PM
Ilkka, did you at all read the fairly large portion of the book where Kurzweil discusses reverse-engineering the brain as one of the most feasible strategies to develop strong artificial intelligence? He discusses at length the currently exponentially improving resolution, speed, and cost-effectiveness of brain imaging.
Posted by Markku | 3:46 PM
Its just a case of managing careful transition The AI version of you is activated after you are sedated. When you 'wake up' you are the AI, no reason for you to even know as long as continuity is maintained. The old you is put to sleep, never waking again.
Posted by Anonymous | 4:08 PM
Reverse engineering the brain to a reasonable level would require being able to somehow map all the ways in which the 100 billion neurons, 900 billion glial cells and loads of other cells interact with another in a soup of hormones, nerotransmitters and whatnot. This includes biochemical, electromagnetic and mechanic interactions. And every single one of the cells is complex chemical factory with individual features. I'm not holding my breath anytime soon.
Posted by Anonymous | 4:15 PM
(I have been blabbing about this subject for a while now, so forgive me for repeating myself, but..)
One seemingly promising approach to building intelligence is described in a book called On Intelligence written by Jeff Hawkins. Hawkins argues that the neo-cortex of the human brain has a common algorithm which is embodied in a hierarchical temporal memory.
Hawkins claims that the human intelligence is based on a hierarchy of "nodes" that take input from the senses or lower level nodes (which are effectively the same thing). Nodes store common spatial patterns and temporal patterns and propagate their predictions up the hierarchy. No single node knows that it's looking at, say, an apple. These bits and pieces of information propagate up the hierarchy and the knowledge (ie. a prediction) emerges at the top level.
Now, the beauty of this theory is that Hawkins can map it to the current knowledge of the anatomy of the neo-cortex. He used the knowledge of the anatomy as a restriction on his model, ie. it had to be something that the brain really could do. And indeed it does. The six layers of visual cortex, for example, can store these spatial and temporal patterns and they propagate their prediction to other regions (higher levels).
The book, On Intelligence, is still accurate, but Hawkins et al. have been updating the theory and have managed to make a mathematical formalism for the model. (Which, not surprisingly, involve a variation of Bayesian Networks.)
Hawkins's theory is based on a lot of brain research, so it's not in that sense a ground-breaking idea. Hawkins just ties these strings into one coherent and beautiful model. But that's not all. We can actually build intelligent machines based on this model.
Along with Dileep George, Hawkins has produced an impressing demonstration of the theory in the field of "computer vision." It's a simplish example, sure, but it's concentrated to solve the invariance problem: how does the human brain resolve indefinitely varied visual inputs into distinct objects in mind.
Now, I don't know a lot about the state of the art of computer vision, but their demo seems pretty impressive. (Look at the table of correctly identified pictures in page 6.) It's just simple 32x32 black-and-white pictures that they use, but the robustness of the system is incredible. You can add noise, resize, rotate, etc, the picture to nearly unidentifiable and the system produces the correct result.
Not only that, but Hawkins is actually busy building a business based on his theory. So it's not just empty talk. They should have concrete deliverables in less than a year. (If the vision invariance demo doesn't impress you.)
Posted by Jarno | 4:41 PM
Markku: Ilkka, did you at all read the fairly large portion of the book where Kurzweil discusses reverse-engineering the brain as one of the most feasible strategies to develop strong artificial intelligence?
I did, and believe it to be the best technique for creating AI. It would also be useless technique, since such an AI could not do anything that a human brain couldn't do.
Posted by Ilkka Kokkarinen | 4:43 PM
Markku: "... Kurzweil discusses reverse-engineering the brain as one of the most feasible strategies to develop strong artificial intelligence?"
Ilkka: "I did, and believe it to be the best technique for creating AI. It would also be useless technique, since such an AI could not do anything that a human brain couldn't do."
If Jeff Hawkins is right (see my previous post), then the end-result of reverse-engineering the brain is not the brain, but an algorithm to produce human-like intelligence. The algorithm (or the "machine" based on the algorithm) basically just finds persistent causes in the world through sensory data.
Now, if we attach embodiment of this algorithm (which Hawkins calls an HTM, hierarchical temporal memory) to some other sensory data, then the HTM can "see" causes in the world which are not visible to us. One example Hawkins uses is the weather system. We can think of the weather measurement stations as a giant retina of the ongoing weather. If the HTM looks at this kind of sensory data and finds the persistent causes in it, that would produce something quite unreachable for our brains.
Other possible "visions" could be based on 4 or more (spatial) dimensional sensory data; effectively producing 4D vision. (Or n-dimensional vision.)
Posted by Jarno | 4:56 PM
Reverse engineering the brain to a reasonable level would require being able to somehow map all the ways in which the 100 billion neurons, 900 billion glial cells and loads of other cells interact with another in a soup of hormones, nerotransmitters and whatnot. This includes biochemical, electromagnetic and mechanic interactions. And every single one of the cells is complex chemical factory with individual features. I'm not holding my breath anytime soon.
Don't understimate the power of exponential growth.
Posted by Markku | 5:27 PM
I did, and believe it to be the best technique for creating AI. It would also be useless technique, since such an AI could not do anything that a human brain couldn't do.
A useless technique? Your argument is wrong in so many ways. First of all, reverse-engineering the brain does not mean duplicating every feature of the human brain. Reverse-engineering involves taking something apart and understanding how it works by analysing its parts and their interactions. That's precisely where vast computational resources come in extremely handy: they allow performing rapid simulations and tests.
Secondly, even if reverse-engineering the brain meant only duplicating all its features, you could make as many copies of the AIs and run them as fast as your (exponentially expanding) hardware resources permit. Imagine billions upon billions of equivalents of scientists doing nothing but trying to build better AIs, all working millions of times faster than their (unenhanced) human counterparts.
Posted by Markku | 5:36 PM
Ilkka,
You mentioned the coal and petroleum industries underwriting their (the activists) efforts. Man, Ive been screaming this to people I know for YEARS. France gets 78% of her juice from Nuclear reactors and liberals just worship France. Keep spreading the word about this. Sen. John Warner has attemtped to kill windpower more or less in the states contending that they might threaten military radar in HEARTLAND America. Makes me sick. Its the best racket in the world, energy.
Posted by Anonymous | 7:23 PM
As anon said, don't focus on the meatspace "you" that lays down in a scanner and gets up, and later dies. There's a second "you", the machine version; that version is the one which lives forever and which is eternally grateful it decided to lay down and get itself immortalized.
Both of you feel they are you. You, the meatspace version, probably are a bit snotty about claiming you're the "real" you. But after 50 years, that argument is moot.
Posted by Leonard | 9:39 PM