I see that you are interested in exploring the 'should we' question, but I want to insert here Lee Smolin's second response to Lanier's
Edge presentation which responds to the 'can we' question concerning both 1) the ThX project of constructing a human-like artificial intelligence in a machine substrate with which to replace humans, and 2) the ThX project of re-engineering our species.
Lee Smolin
Date: September 27, 2000
Jaron is raising some very important points that deserve closer examination and discussion. Among them is his challenge to the idea that the optimization of present day computers could produce anything with the capabilities of living, intelligent animals, cats let alone people. I think Jaron is right to point out that the arguments for this thesis rest on incorrect assumptions. I believe that Jaron's argument can be strengthened and I would like to explain how. The following is just a sketch, but I hope it suffices to stimulate the debate.
The problems to be addressed are 1) what kinds of problems can computers solve and whether they differ in kind from the kinds of problems humans solve. 2) What kind of problem is it to design a computer and whether it differs in kind from the problem of designing a human, or a creature with equal capabilities.
To approach these questions it helps to begin with the idea that some design problems involve searching a space of possible design parameters. We know that in these cases there are simple optimization algorithms that will find the local extrema in whatever basin of attraction one happens to be in. However, optimization is a small part of design because it can be used reliably to solve only a small subset of possible design problems. To talk about this we may distinguish five classes of design problems.
CLASS 1: Local optimization problems problems which can be solved with standard hill-climbing techniques.
CLASS 2: Locate a pretty good, but not necessarily global extremum in a configuration space with many local extrema and many basins of attraction.
CLASS 3: Locate the global extremum in a configuration space with many local extrema and many basins of attraction.
CLASS 4: Find local extrema in a landscape which changes unpredictably on the same time scale it takes to find local optima.
CLASS 5: find local extrema in cases in which the computation time required to construct the configuration space and/or calculate the fitness function is either infinite or much longer than the time available. These are the class of problems which have to be invented or discovered before they can be solved, as there is no algorithm that can lead to their formulation or complete specification.
Let us first discuss the first question. At least so far, computers are very good at solving CLASS 1 problems, and there are decent algorithms for simple CLASS 2 problems. But we do not have good methods for finding global extrema and hence solving CLASS 3 problems. To my knowledge computers can do decently at some simple CLASS 4 problems, but can easily fail when they become more complex. By definition computers have problems solving CLASS 5 problems, as the computation time to set up the extremization problem is prohibitive. However humans can often solve CLASS 3 problems and are also quite good at CLASS 4 problems. This should be no surprise, this is part of our biological specialization. This is what is required to flourish in a new environment, domesticate a new species, become farmers, populate almost all the ecological zones on the planet and so forth.
But humans can do even better than that, we can both invent and solve CLASS 5 problems. This is what poetry, art, music and science, are about. We invent the forms and traditions and then we master them. We can thrive in a domain in which we create optimal versions of things that did not even exist a short time before. We are not extremizing in a landscape, we are building the landscape on the same time scale that we master it.
One correspondent suggested that anyone who thinks people are different from machines are naive romantics. This is not true, we are different because we have vastly different capabilities. It is irrelevant to talk of the universality of Turing machines, for Turing machines are entities that run programs that must be written by an external entity. So far at least the only entities we know of who can function as those external programmers are humans.
Humans are intelligent creatures that do not need to be programmed by any external agency. Turing machines are designed, we are the result of natural selection. We need then to examine the second question, whether designing or programming a computer is in the same CLASS of problems as the problems natural selection solved in the course of evolution.
Of course inventing the idea of a digital computer was a CLASS 5 problem. But once we had the idea, the optimization of digital computers is mainly a CLASS 1 problems. This is what Moore's law is about, it tells us how quickly local optimization can work when ample resources are available. One of the points Jaron is making is that the design of software required to do justice to the exponentially increasing capabilities of our machines are not CLASS 1 problems. Moore's law tells us that the fitness landscape for software is changing on a time scale comparable to the time required to write and debug software. Thus writing software involves problems of at least CLASS 4. This is of course just a different way of making one of Jaron's arguments.
For there to be a danger of robots taking over, or even being able to do a decent job entertaining us, replacing songwriters and singers, artists, scientists and comedians, one of two things have to happen. Either we will be able to design a machine that could replace us, which means a machine that can solve problems of CLASS 5, or we will be able to design a machine that could in turn design a machine that could solve CLASS 5 problems.
But
while we can solve problems up to CLASS 5, so far we have only been able to design machines that can solve CLASS 2 problems reliably. And so far machines are not able to design other machines to solve even CLASS 1 problems. When one puts it this way it is clear that it is not just a matter of Moore's law, designing one of us is a very different kind of problem then optimizing a programmable digital computer.
What kind of problem is it to design an entity that can solve CLASS 5 problems? We know we were created by natural selection, acting on not only us but the whole collection of living species. This is at least a CLASS 4 problem, but it is very likely at least a CLASS 5 problem. The interactions among many species as they evolve under the rules of natural selection is a CLASS 4 problem, as is shown by models of Bak and Sneppen, Kauffman, Sola and others. But there are good arguments, summarized in Stuart Kauffman's forthcoming book, that natural selection and cultural evolution are really CLASS 5 problems. He argues that they are problems in which the construction of the fitness landscape itself is so computationally intensive that it is not correct to separate the specification of the fitness landscape from its optimization. Instead, both take place together. This means really that the metaphor of optimization has broken down completely. Whatever evolution is doing cannot, he argues, be conceptualized as extremization on a pre-existing fitness landscape.
Thus, the problem of designing an entity that can solve CLASS 5 problems is at least a CLASS 4 problem, and very likely is a CLASS 5 problem. But is it only this hard, or harder still? Human's can solve some CLASS 4 and 5 problems, but it is not at all obvious that the problems of these kinds that we can solve are comparable to the problems that natural selection has solved in designing us. At the very least, it is likely that the time required to solve the problem of designing us may take a great deal longer than the tine it takes to solve the CLASS 4 and 5 problems we have so far dealt with. It took natural selection 4 billion years to design us. Let us assume that we could do it much faster. How much faster? Let us assume that we could use genetic engineering to engineer an artificial speciation in an animal. Speciation is a process that takes on the order of 100,000 years. Given very optimistic assumptions it is possible to imagine that some years from now this is something we will be able to accomplish in on the order of 100 years. It could certainly not be less than that as we cannot do it faster than the time it takes for several generations to grow to maturity. (Because the interaction of an animal and its environment is a CLASS 5 problem, we are not likely to be able to simulate it reliably enough to replace the phase where we grow the animal and observe what happens.) This would mean that we had the tools to speed up natural selection by a factor of 1,000. Even with this fantastic increase of speed it would still take us a million years to invent something like ourselves, starting from scratch. (Note that this is true even if we skip the pre-cambrian stages of evolution, which begins with creatures whose cell biology and biochemistry is far advanced of what we have so far designed. Note also that many biologists working in parallel won't help as natural selection also works in parallel.)
This is on the order of the lifetime of a species. A problem like this, whose minimum time for solution is on the order of the lifetime of a whole species of creatures that can solve CLASS 5 problems deserves a separate class. So we may call this a CLASS 6 problem.
Is it possible that there is a way to do it much faster, by taking a route that natural selection could not have? One cannot say this is impossible, but all this means is that so little is known about the problem that it is in a class of problems we have no idea how to solve.
To summarize: the claim that optimization of present computer designs could produce something that is "as powerful" as humans requires that there is only one kind of intelligent entity, and they all live in a fixed landscape with a single local extremum. But we are not only not in the same basin of attraction as present day computers, it is not even obvious that the problem of constructing us has anything in common with problems we have so far solved. This is not to deny that someday humans may learn how to solve the problem of designing creatures that can themselves solve CLASS 5 problems. The point is only that there is no rational basis for predicting when or even whether this may happen, as the solution to this problem is not closely related to the kind of optimization problems that human designers have so far learned to solve.