|
|
|
Interlude -- two stories about intelligenceThere are two distinct ways in which intelligence can be increased. Here are two stories to illustrate the difference. The first story is from Blood in the Streets, by Davidson and Rees-Mogg, chapter one:
This was like a battle between grownups and children. The Chinese and the foreign devils were both human, but they were clearly on different levels. But the British and other "westerners" had a long way to go. Soon after this battle, they raised themselves to a higher level. In Men, Machines, and Modern Times, Elting Morison relates the fascinating story of how the British and American navies adopted the practice of aiming guns before firing them. This was about a hundred years ago, about the same time as the battle described above. Before that, they could fire all afternoon at a ship and not hit it. That's not quite as egregious as the Chinese firing at an embassy for two months and not hitting it, but it's bad enough. The ship's cannon were fired at a fixed angle. A sailor could adjust the angle once, to take into account the distance of the target; but when the angle was set, it was set. The ship, of course, was moving up and down in the water, but the sailors had no way to compensate for this motion. At least no easy way. But one sailor (whose name was not recorded) tried to compensate for the motion of the ship. Instead of locking his gun in one position, he tried to adjust the angle in real time. The gears weren't set up in such a way as to allow this, but he tried to do it anyway. His officer noticed that he was more accurate than the other sailors, and watched what he was doing. The officer thought about how the guns could be redesigned so the sailors could aim them continuously. When his design was adopted, the accuracy of the British Navy increased by 3,000 percent in a few years. What does 3,000 percent mean, specifically? Morison explains:
This is impressive, and it certainly makes a decisive difference in a battle, but it's not the same kind of difference as the previous example. The difference between the Chinese and the foreign devils was fundamental: the Chinese had an animist model of reality, but the "westerners" had a scientific model. This is a qualitative difference, not just a matter of degree. Now, of course, the Chinese fight battles the same way we do, because they think like we do. Mao Tse-tung introduced Marxism into China. Marxism is a combination of Jewish philosophy and European philosophy. Mao taught the Chinese to think like Europeans. They are no longer animists. They fire material bullets at material targets. Their third-world mentality has been replaced by a first-world mentality. Whether this change is complete, and whether it's permanent (in China or anywhere else), are separate issues that I won't go into here. The point I want to make is that the change from animist to scientific was a discrete change from one mental state to another, not a continuous change. It was not a movement along an exponential curve, it was a step function. When the British Navy adopted a new way of aiming guns, this was an incremental improvement within the scientific model. An improvement of 3,000 percent is a large increment, but it required no fundamental change in the British model of reality. It was a refinement of what they were already doing. They were already aiming their guns at the target. They just figured out how to do it a lot better. When the British started aiming their guns in real time, that was part of an exponential curve, but strictly speaking they didn't increase their intelligence, they increased their precision. This is just one more example of the continually increasing precision in the mechanical arts that began in the middle ages. The curve was almost flat for several hundred years. It didn't turn up perceptibly until the French started making muskets with interchangeable parts about 1790. The steepness of the curve increased steadily in the 19th and 20th centuries. In the 1960's the curve appeared to be exponential. Gordon Moore noticed that the number of transistors on a chip was doubling every year or two, and he called it Moore's Law. This is a special case of a phenomenon with a long history and an uncertain future. It is our ability to make more and more fine-grained machines that is increasing exponentially, not our intelligence per se. Is there another discrete level above the scientific model of reality? When the Chinese fired material bullets at the supposed spirits above their opponents' heads, that was a category mistake. It's an equally fundamental mistake to think that battles are only fought on a material level. It is possible to fire spiritual weapons at spiritual targets. When you hold a three year old on your knee and tell him stories, he lives in the world you create for him. Your mind contains his mind. That's the relation between the storytellers in Hollywood and the rest of us: we are children living in their world. They are winning battles in a war we are only dimly aware of. This too is a war between grownups and children, and we are the children. Lenin said the surest way to destroy a nation is to debase its currency. An even more fundamental way to destroy a nation (or a mind) is to debase its language. The destruction of the Tower of Babel isn't just a mythical event in the ancient past, it's an ongoing process. The destruction of language is the most insidious weapon there is, and those who use it are indeed on a level above the rest of us. They are as far above us as the British were above the Chinese a century ago. Superhumanity -- ?Back at the beginning of his essay, Dr Vinge does not merely assume that AI systems reach our level and "wake up." He makes a stronger assumption:
Superhuman computers are supposed to follow immediately from human-level computers:
Excuse me? Little doubt?! This is the exact point that requires an argument. If we can create human equivalence in a machine, then it would certainly be possible to construct "beings" that are more intelligent, but how much more? Of course incremental improvement is always possible, but more intelligent is not the same as superhumanly intelligent. This latter concept needs to be examined. Professor Vinge makes a distinction between "weak superhumanity" and "strong superhumanity." He says,
It certainly is hard, but that's what needs to be done. He continues,
No, it wouldn't. But it's not obvious how this applies to humans. Speed may or may not make a difference. Speed is not sufficient for intelligence: if you start with a fool and speed up his brain, he is still a fool. He will just miss the point a hundred times faster. Nor is speed necessary for intelligence: there is an old proverb, "Still water runs deep." Albert Einstein was a perfect example. He wasn't a fast student, but he saw as deeply into physics as any of his faster colleagues. On the other hand, it would be rash to assert that speed is irrelevant or useless. If you take somebody who is already bright, and speed up his brain, this should be a benefit. John von Neumann could read math and physics journals at full speed, the way ordinary people read a newspaper. If I could do that, I would accomplish a lot more than what I'm doing now. I would be "weakly superhuman" compared with my present level. What if you could take von Neumann himself and speed him up? Would he be a genius on an even higher level? Not necessarily. Sometimes my mind races. Thoughts fly by so fast that I can't keep track of them or make sense of them. I have not found this state to be productive (fun, maybe, but not productive). Some of my best ideas have come to me when I have taken my favorite elixir, which slows me down to just the right pace. Recently I have been trying to learn the trick of adjusting the pace myself. Creative thought isn't the same as data processing. To say that thinking faster is always better is as absurd as saying that "Ave Maria" is more beautiful when you play it back at ten times the normal speed. There is an optimal tempo for music, and for creative thought. Nevertheless data processing is an important part of what we do, and when we learn to do it faster, the difference may be dramatic. When we use computers to process data, it's like the British learning to aim their guns in real time. It was only an incremental improvement in their capability, not a fundamental improvement, but it gave them an unanswerable advantage in battle (until their opponents adopted the same technology). To sum up this section: I agree with Dr. Vinge's point about speed. Sometimes it helps, and sometimes it doesn't, but in any case it only takes us so far. Going from fast to faster is not like going from an animist model of reality to a scientific model. If we are looking for the essence of superhuman intelligence, speed is not the place to look. Dr. Vinge continues:
Actually Dr. Vinge doesn't say much about strong superhumanity later in the paper. He just takes it for granted that there is such a thing, and that we already know what it is. He is more concerned with trying to decide how we should feel about it than with defining it or establishing its existence. But he does say a couple of things about it. Two paragraphs later he says this:
If one of the components is already strongly superhuman in its own right, this comment gets us nowhere. If not, he needs to demonstrate that a combination of human-level intelligences (or weakly superhuman intelligences) can be strongly superhuman. In any case, this comment is no help at all in arriving at a definition of "strongly superhuman." Toward the end of the paper he says
A high bandwidth connection doesn't necessarily accomplish anything. It depends on who is being connected to whom. If the entities are not already strongly superhuman, a high bandwidth connection isn't going to take them to that level. This is like the point discussed above, where you speed up a fool's brain and still have a fool. If you wire fools together with a high bandwidth connection -- even a telepathic connection! -- you just have a bigger fool. We already have millions of people sending messages back and forth via AOL instant messenger, and many of them have high bandwidth connections, but that doesn't create a superhuman entity. We also have people who are not fools sending messages to each other on the internet, and this does multiply their intelligence, but they are still basically on the same level. The scientific community is at most a "weakly superhuman" entity. Bandwidth by itself doesn't take us from human-level intelligence to a higher level, any more than speed does. If strongly superhuman intelligence is going to be fundamentally different from human intelligence, then it will require a fundamentally new language: a new syntax, a new concept of "entity," and a new concept of "concept" (among other things): a new model of the world. The mind has a model of itself, and going from one level of intelligence to a higher level would involve creating a new model, which would be expressed in a new language. But is there a model of the mind, or of the universe, that is qualitatively better than the one we have? Is there a model as far above the scientific model as the scientific model is above the animist model? I think there must be, for three reasons:
If there is a model of reality beyond our present scientific model, there is no way to know a priori how long it will take us to discover it. We don't have to wait for AI systems to catch up with us. Our reaching the next level of intelligence doesn't depend on computers reaching our level. Those two things have nothing to do with each other. If AI systems do catch up with us, that doesn't imply that we (or they) can take the next step. We may discover a new model of reality before AI systems catch up with us, or a long time after. If they never catch up, that doesn't prevent us from discovering the new model. On the other hand, there is no guarantee that we can do it at all. If there is a model of reality beyond our present scientific model, and hence a level of intelligence far enough above us to qualify as "strongly superhuman," this does not imply that we could bootstrap ourselves to the strongly superhuman level. Nor does it imply that a human-level AI could bootstrap itself to the strongly superhuman level. A human-equivalent AI system would face exactly the same problem we face when we try to invent or discover the new model. Nor does the existence of such a model imply that there is another model beyond that, and another one beyond that, ad infinitum. The existence of a strongly superhuman intelligence does not imply that such an entity could design an even higher intelligence, and so forth. We are constantly improving our model of the world. Whenever you learn a new concept such as "epigenetic" or "eigenvector," you have extended your language and increased your intelligence incrementally. When you learn basic concepts such as "definition," "theorem," and "experiment," you have increased your intelligence in a more fundamental way. When you learn a whole new subject such as differential equations or economics, you acquire a very powerful set of conceptual tools that didn't come with your original human toolkit, and you become "superhuman" compared with your previous state. People do this all the time. It's an ongoing process, not an intelligence explosion. When robots reach our level (where "our level" refers to scientists, not morons or animists), they will have to extend their model of the world in the same laborious way, subject by subject, concept by concept. It's not going to be any easier for them than it is for us. This applies in particular to improving their model of their own minds. When robots reach our level, they won't be born knowing how their minds work. They will be the product of decades of evolution. Their brains will be a black box to them just as much as our brains are a black box to us. They will have to construct their own neuroscience. They will have to figure out how brains work (both their brains and ours), and it won't be any easier for them than it is for us. If and when they reach our level, it will be exactly as hard for them as it is for us. The great changes in our future don't depend on AI systems reaching our present level. Nothing depends on that. That whole issue is a red herring. The possibility of a next step -- redesigning our hardwareBack at the beginning of the paper, Professor Vinge lists four ways in which superhuman intelligence might emerge. The fourth is
Of the four, this is the one that he doesn't pursue. At this point I'm going to leave Vernor Vinge behind. From here on I'm just going to discuss my own ideas. Suppose we start with a human being. We incorporate nanoscale devices into his or her cells, including brain cells. A nanodevice is like a new organelle, which monitors the activity of the cell and tweaks it or assists it as necessary. The artificial organelles are designed to mesh with the cellular machinery that is already there. The next step is to suppose that some of the natural organelles in the cell are replaced by artificial organelles. In other words, at this stage the new organelles don't just act as adjuncts to the natural organelles, they take over certain tasks, so the original organelles are no longer necessary. For example, we could design an artificial lysosome that would clean up the cell and remove waste products better than natural lysosomes. Suppose this happens in all cells, throughout the body, including the brain. Suppose that, over a period of time, more and more functions of the cell are taken over by artificial organelles, and finally the entire cell consists of artificial lysosomes, artificial mitochondria, artificial everything. The new organelles are not foreign objects inserted from the outside. There are artificial ribosomes which make the new organelles right there on site, in the cell. The chromosomes have been modified so that when the cell divides, the new cell has artificial organelles. The DNA code for the natural organelles has been removed. It isn't necessary to carry this scenario all the way to its conclusion. Even at the earliest stages of this process, the humans who morph themselves into transhumans will be more intelligent than we are. To the extent that imperfections in our brains are holding us back, they will be superior. They will never be fogged out. They will have impeccable memories. Their attention spans will be longer than ours. Their brain cells will be able to secrete caffeine, THC, or any other molecule in the exact amount needed at any given time for maximum alertness and creativity (unless the government prevents this -- a caveat which applies to the whole scenario). They will be like we are at our best moments, only far beyond that, and their lucidity won't last just a moment, it will go on and on. This is transhuman consciousness, or hyperlucidity. On page 15, "There is no Moore's Law for software," I wrote:
Well, of course there is. Our own brains run "software" that embodies an abstract model of a mind, which is (more or less) correct on the most general level. That's not to say there is no room for improvement, but our model of ourselves is correct enough so that we are conscious. When we have better brains, we will be able to design still better brains. When we start redesigning our brain cells, Moore's Law will apply to cells. They will improve exponentially. There is no obvious limit to how far this can go. Maybe it won't go all the way to The Singularity. I don't really care. We can definitely get as far as the Transhuman Metamorphosis. That's the important thing. What happens after that remains to be seen. We will find out when we get there. This line of thought is continued on the cellular transformation page. The possibility of a next step -- redesigning our language
the replacement of Roman numerals with Arabic numerals -- graphs (Nicole Oresme, 14th century) -- the development of mathematical notation in the 16th, 17th, and 18th centuries, which made modern science possible -- this notation is still used and has become fossilized -- it's holding us back -- the possibility of making a fresh start and inventing a radically new notation for mathematics the possibility of non-cryptic programming languages -- the application of user interface design to programming languages -- it's time to do the same thing in software that Doug Engelbart did in hardware: change the very idea of what a program looks like and how we interact with it human language as a model of the world -- the appearance of grammar in ancient Greece and India (the first step toward an explicit model of language itself) -- Plato's Republic, the first explicit model of the human mind -- the appearance of the first dictionaries in the 17th century -- Kant, Bolzano, Frege, Russell, Principia Mathematica, Carnap, the idea of a metalanguage -- simulating the evolution of language -- the origin of language according to Calvin and Bickerton in Lingua ex Machina -- root words in Hebrew, Sanskrit, and German -- generating other concepts from root concepts -- the possibility of starting over with a new set of root concepts -- the application of genetic programming to language design -- the possibility of entirely new languages -- the convergence of computer languages and human languages the quasi-three-dimensional nature of ordinary human consciousness (i.e. we think in three dimensions, but not very well) -- geometry in ancient Greece -- the development of explicit three dimensional perspective in the 15th century -- the discovery of non-Euclidean geometry in the 19th century -- the discovery of the quantum domain in the 20th century -- the possibility of projecting our consciousness into the quantum domain and seeing the world from the point of view of an electron (or a photon, or any subatomic wave/particle) -- the possibility of incorporating abstract mathematical space (non-Euclidean, higher dimensional, complex, topological, etc) into our consciousness and basing our language on it On page 170 of Lingua ex Machina, William Calvin says
Human language is like that. It's an accretion of kluges. There is nothing stopping us from starting over with a clean sheet of paper and designing a language that accurately reflects the structure of reality. This is one of the main tasks for philosophy in the 21st century. As far as I'm concerned, this an ongoing activity. My thoughts are constantly in motion. My language continually transforms itself. Concepts are brought into consciousness, re-examined in the light of everything I have learned, and sent back to my memory with new connections. When I think about mathematics, I try to "look through" the notation to the underlying mathematical reality, and I am trying to conceive some kind of notation (using computer graphics) that will display the math in a more perspicuous way. Likewise for all the other things mentioned above. This is something I do, not something an AI will do for me at some magical moment in the future. It's not something an AI could do for me. If I'm going to do it at all, I have do it myself. This is Biocentric Transhumanism: it's not so much an ism as an activity. "In him was life"In discussions of AI and the Singularity, it is always assumed that it's up to us to create a superhuman entity. The idea that a superhuman mind already exists, and it creates us, is outside the bounds of the discussion. But I'm not sure why I have to stay within those bounds. In the 915 Manifesto, I wrote:
That's still what I'm trying to do. At this point in my life, pot no longer has that effect, and I stopped smoking it about three years ago. But that doesn't alter the fact that our minds are part of a larger mind. What marijuana showed me in the first few months after the breakthrough is still valid. Going from our present level to a higher level requires both invention and discovery. It's partly a matter of inventing a new model of reality, as described above, but it also involves tuning in to a mind that is already there. From pre-Cambrian to post-CambrianVernor Vinge quotes Stan Ulam paraphrasing a conversation with John von Neumann: we appear to be "approaching some essential singularity in the history of the race beyond which human affairs, as we know them, could not continue." Here the word "singularity" is used not in the technical sense we find in math or physics, but in a more general sense: an event so extraordinary that our usual categories no longer apply. As I said above, I think this is a misnomer, but let that pass. There isn't going to be an "intelligence explosion," as conceived by Vernor Vinge, but there could still be an event beyond which human affairs as we know them could not continue. Such an event could come from the advent of artificial life. In his book In the Blink of an Eye, Andrew Parker asserts that the nature of the Cambrian explosion is generally misunderstood. According to him, the internal body plans of all animal phyla were already there, and had been there for tens of millions of years. They evolved during the "Precambrian surge in evolution." Precambrian animals were soft wormlike creatures. "The Cambrian explosion is all about external body parts only," he says. The thesis of his book is that "Between 544 and 543 million years ago a revolution took place. During this one million year period, vision was born." Vision made predator/prey relationships possible, and that was what led to the development of hard body parts. His argument is plausible, but I don't know enough about the subject to decide whether it's true. In any case, I think the Cambrian explosion is the best analogy for the transformation we are about to go through. As Moore's Law continues on its inexorable path toward the nanoscale, eventually nanodevices will mesh with cells, and the result will be a revolution comparable to the Precambrian surge and the Cambrian explosion, combined. The availability of cheap intelligence will have an effect similar to the development of eyes. It will revolutionize predator/prey relationships. The availability of new materials (carbon nanotubes, diamond...) will lead to hard body parts of a new order. Cheap energy (i.e. a new energy generating organelle to replace mitochondria) will cause as profound a change as the transition from anaerobic to aerobic life. Not only that, we have to consider the possibility of entirely new body plans, and even new substrates for life. There is no reason why we couldn't have organisms made of plasma, and there is no reason why life couldn't exist at the subatomic level. And when trillions of such organisms are wired together... To repeat what I said above, combination organisms will soon include smart dust and grid computers which will use genetic algorithms to redesign themselves with less and less human supervision. If the capability of such combination organisms continues to grow exponentially, the result will indeed be a profound change. As Andrew Parker says, 544 million years ago "Life was about to be stirred up."
External links: Vinge's own comments on his article - The Spring 2003 issue of Whole Earth magazine contains an edited version of Vinge's 1993 article, with his own comments added a decade later. Singularity Watch -- John Smart's site The Singularity Institute |