The Singularity Is Near_ When Humans Transcend Biology Part 25

Bell is correct that trying to compare the brain's design to a conventional computer would be frustrating. The brain does not follow a typical top-down (modular) design. It uses its probabilistic fractal type of organization to create processes that are chaotic-that is, not fully predictable. There is a well-developed body of mathematics devoted to modeling and simulating chaotic systems, which are used to understand phenomena such as weather patterns and financial markets, that is also applicable to the brain.

Bell makes no mention of this approach. He argues why the brain is dramatically different from conventional logic gates and conventional software design, which leads to his unwarranted conclusion that the brain is not a machine and cannot be modeled by a machine. While he is correct that standard logic gates and the organization of conventional modular software are not the appropriate way to think about the brain, that does not mean that we are unable to simulate the brain on a computer. Because we can describe the brain's principles of operation in mathematical terms, and since we can model any mathematical process (including chaotic ones) on a computer, we are able to implement these types of simulations. Indeed, we're making solid and accelerating progress in doing so.

Despite his skepticism Bell expresses cautious confidence that we will understand our biology and brains well enough to improve on them. He writes: "Will there be a transhuman age? For this there is a strong biological precedent in the two major steps in biological evolution. The first, the incorporation into eukaryotic bacteria of prokaryotic symbiotes, and the second, the emergence of multicellular life-forms from colonies of eukaryotes....I believe that something like [a transhumanist age] may happen."

The Criticism from Microtubules and Quantum Computing

Quantum mechanics is mysterious, and consciousness is mysterious.Q.E.D.: Quantum mechanics and consciousness must be related.-CHRISTOF KOCH, MOCKING ROGER PENROSE'S THEORY OF QUANTUM COMPUTING IN NEURON TUBULES AS THE SOURCE OF HUMAN CONSCIOUSNESS21

Over the past decade Roger Penrose, a noted physicist and philosopher, in conjunction with Stuart Hameroff, an anesthesiologist, has suggested that fine structures in the neurons called microtubules perform an exotic form of computation called "quantum computing." As I discussed, quantum computing is computing using what are called qubits, which take on all possible combinations of solutions simultaneously. The method can be considered to be an extreme form of parallel processing (because every combination of values of the qubits is tested simultaneously). Penrose suggests that the microtubules and their quantum-computing capabilities complicate the concept of re-creating neurons and reinstantiating mind files.22 He also hypothesizes that the brain's quantum computing is responsible for consciousness and that systems, biological or otherwise, cannot be conscious without quantum computing. He also hypothesizes that the brain's quantum computing is responsible for consciousness and that systems, biological or otherwise, cannot be conscious without quantum computing.

Although some scientists have claimed to detect quantum wave collapse (resolution of ambiguous quantum properties such as position, spin, and velocity) in the brain, no one has suggested that human capabilities actually require a capacity for quantum computing. Physicist Seth Lloyd said:

I think that it is incorrect that micro tubules perform computing tasks in the brain, in the way that [Penrose] and Hameroff have proposed. The brain is a hot, wet place. It is not a very favorable environment for exploiting quantum coherence. The kinds of superpositions and a.s.sembly/disa.s.sembly of microtubules for which they search do not seem to exhibit quantum entanglement....The brain clearly isn't a cla.s.sical, digital computer by any means. But my guess is that it performs most of its tasks in a "cla.s.sical" manner. If you were to take a large enough computer, and model all of the neurons, dendrites, synapses, and such, [then] you could probably get the thing to do most of the tasks that brains perform. I don't think that the brain is exploiting any quantum dynamics to perform tasks.23

Anthony Bell also remarks that "there is no evidence that large-scale macroscopic quantum coherences, such as those in superfluids and superconductors, occur in the brain."24 However, even if the brain does do quantum computing, this does not significantly change the outlook for human-level computing (and beyond), nor does it suggest that brain uploading is infeasible. First of all, if the brain does do quantum computing this would only verify that quantum computing is feasible. There would be nothing in such a finding to suggest that quantum computing is restricted to biological mechanisms. Biological quantum-computing mechanisms, if they exist, could be replicated. Indeed, recent experiments with small-scale quantum computers appear to be successful. Even the conventional transistor relies on the quantum effect of electron tunneling.

Penrose's position has been interpreted to imply that it is impossible to perfectly replicate a set of quantum states, so therefore perfect downloading is impossible. Well, how perfect does a download have to be? If we develop downloading technology to the point where the "copies" are as close to the original as the original person is to him- or herself over the course of one minute, that would be good enough for any conceivable purpose yet would not require copying quantum states. As the technology improves, the accuracy of the copy could become as close as the original to within ever briefer periods of time (one second, one millisecond, one microsecond).

When it was pointed out to Penrose that neurons (and even neural connections) were too big for quantum computing, he came up with the tubule theory as a possible mechanism for neural quantum computing. If one is searching for barriers to replicating brain function it is an ingenious theory, but it fails to introduce any genuine barriers. However, there is little evidence to suggest that microtubules, which provide structural integrity to the neural cells, perform quantum computing and that this capability contributes to the thinking process. Even generous models of human knowledge and potential are more than accounted for by current estimates of brain size, based on contemporary models of neuron functioning that do not include microtubule-based quantum computing. Recent experiments showing that hybrid biological! nonbiological networks perform similarly to all-biological networks, while not definitive, are strongly suggestive that our microtubuleless models of neuron functioning are adequate. Lloyd Watts's software simulation of his intricate model of human auditory processing uses orders of magnitude less computation than the networks of neurons he is simulating, and again there is no suggestion that quantum computing is needed. I reviewed other ongoing efforts to model and simulate brain regions in chapter 4, while in chapter 3 I discussed estimates of the amount of computation necessary to simulate all regions of the brain based on functionally equivalent simulations of different regions. None of these a.n.a.lyses demonstrates the necessity for quantum computing in order to achieve human-level performance.

Some detailed models of neurons (in particular those by Penrose and Hameroff) do a.s.sign a role to the microtubules in the functioning and growth of dendrites and axons. However, successful neuromorphic models of neural regions do not appear to require microtubule components. For neuron models that do consider microtubules, results appear to be satisfactory by modeling their overall chaotic behavior without modeling each microtubule filament individually. However, even if the Penrose-Hameroff tubules are an important factor, accounting for them doesn't change the projections I have discussed above to any significant degree. According to my model of computational growth, if the tubules multiplied neuron complexity by even a factor of one thousand (and keep in mind that our current tubuleless neuron models are already complex, including on the order of one thousand connections per neuron, multiple nonlinearities, and other details), this would delay our reaching brain capacity by only about nine years. If we're off by a factor of one million, that's still a delay of only seventeen years. A factor of a billion is around twenty-four years (recall that computation is growing by a double exponential).25

The Criticism from the Church-Turing Thesis

Early in the twentieth century mathematicians Alfred North Whitehead and Bertrand Russell published their seminal work, Principia Mathematica Principia Mathematica, which sought to determine axioms that could serve as the basis for all of mathematics.26 However, they were unable to prove conclusively that an axiomatic system that can generate the natural numbers (the positive integers or counting numbers) would not give rise to contradictions. It was a.s.sumed that such a proof would be found sooner or later, but in the 1930s a young Czech mathematician, Kurt G.o.del, stunned the mathematical world by proving that within such a system there inevitably exist propositions that can be neither proved nor disproved. It was later shown that such unprovable propositions are as common as provable ones. G.o.del's incompleteness theorem, which is fundamentally a proof demonstrating that there are definite limits to what logic, mathematics, and by extension computation can do, has been called the most important in all mathematics, and its implications are still being debated. However, they were unable to prove conclusively that an axiomatic system that can generate the natural numbers (the positive integers or counting numbers) would not give rise to contradictions. It was a.s.sumed that such a proof would be found sooner or later, but in the 1930s a young Czech mathematician, Kurt G.o.del, stunned the mathematical world by proving that within such a system there inevitably exist propositions that can be neither proved nor disproved. It was later shown that such unprovable propositions are as common as provable ones. G.o.del's incompleteness theorem, which is fundamentally a proof demonstrating that there are definite limits to what logic, mathematics, and by extension computation can do, has been called the most important in all mathematics, and its implications are still being debated.27 A similar conclusion was reached by Alan Turing in the context of understanding the nature of computation. When in 1936 Turing presented the Turing machine (described in chapter 2) as a theoretical model of a computer, which continues today to form the basis of modern computational theory, he reported an unexpected discovery similar to G.o.del's.28 In his paper that year he described the concept of unsolvable problems-that is, problems that are well defined, with unique answers that can be shown to exist, but that we can also show can never be computed by a Turing machine. In his paper that year he described the concept of unsolvable problems-that is, problems that are well defined, with unique answers that can be shown to exist, but that we can also show can never be computed by a Turing machine.

The fact that there are problems that cannot be solved by this particular theoretical machine may not seem particularly startling until you consider the other conclusion of Turing's paper: that the Turing machine can model any computational process. Turing showed that there are as many unsolvable problems as solvable ones, the number of each being the lowest order of infinity, the so-called countable infinity (that is, counting the number of integers). Turing also demonstrated that the problem of determining the truth or falsity of any logical proposition in an arbitrary system of logic powerful enough to represent the natural numbers was one example of an unsolved problem, a result similar to G.o.del's. (In other words, there is no procedure guaranteed to answer this question for all such propositions.) Around the same time Alonzo Church, an American mathematician and philosopher, published a theorem that examined a similar question in the context of arithmetic. Church independently came to the same conclusion as Turing.29 Taken together, the works of Turing, Church, and G.o.del were the first formal proofs that there are definite limits to what logic, mathematics, and computation can do. Taken together, the works of Turing, Church, and G.o.del were the first formal proofs that there are definite limits to what logic, mathematics, and computation can do.

In addition, Church and Turing also advanced, independently, an a.s.sertion that has become known as the Church-Turing thesis. This thesis has both weak and strong interpretations. The weak interpretation is that if a problem that can be presented to a Turing machine is not solvable by one, then it is not solvable by any machine. This conclusion follows from Turing's demonstration that the Turing machine could model any algorithmic process. It is only a small step from there to describe the behavior of a machine as following an algorithm.

The strong interpretation is that problems that are not solvable on a Turing machine cannot be solved by human thought, either. The basis of this thesis is that human thought is performed by the human brain (with some influence by the body), that the human brain (and body) comprises matter and energy, that matter and energy follow natural laws, that these laws are describable in mathematical terms, and that mathematics can be simulated to any degree of precision by algorithms. Therefore there exist algorithms that can simulate human thought. The strong version of the Church-Turing thesis postulates an essential equivalence between what a human can think or know and what is computable.

It is important to note that although the existence of Turing's unsolvable problems is a mathematical certainty, the Church-Turing thesis is not a mathematical proposition at all. It is, rather, a conjecture that, in various disguises, is at the heart of some of our most profound debates in the philosophy of mind.30 The criticism of strong AI based on the Church-Turing thesis argues the following: since there are clear limitations to the types of problems that a computer can solve, yet humans are capable of solving these problems, machines will never emulate the full range of human intelligence. This conclusion, however, is not warranted. Humans are no more capable of universally solving such "unsolvable" problems than machines are. We can make educated guesses to solutions in certain instances and can apply heuristic methods (procedures that attempt to solve problems but that are not guaranteed to work) that succeed on occasion. But both these approaches are also algorithmically based processes, which means that machines are also capable of doing them. Indeed, machines can often search for solutions with far greater speed and thoroughness than humans can.

The strong formulation of the Church-Turing thesis implies that biological brains and machines are equally subject to the laws of physics, and therefore mathematics can model and simulate them equally. We've already demonstrated the ability to model and simulate the function of neurons, so why not a system of a hundred billion neurons? Such a system would display the same complexity and lack of predictability as human intelligence. Indeed, we already have computer algorithms (for example, genetic algorithms) with results that are complex and unpredictable and that provide intelligent solutions to problems. If anything, the Church-Turing thesis implies that brains and machines are essentially equivalent.

To see machines' ability to use heuristic methods, consider one of the most interesting of the unsolvable problems, the "busy beaver" problem, formulated by Tibor Rado in 1962.31 Each Turing machine has a certain number of states that its internal program can be in, which correspond to the number of steps in its internal program. There are a number of different 4-state Turing machines that are possible, a certain number of 5-state machines, and so on. In the "busy beaver" problem, given a positive integer Each Turing machine has a certain number of states that its internal program can be in, which correspond to the number of steps in its internal program. There are a number of different 4-state Turing machines that are possible, a certain number of 5-state machines, and so on. In the "busy beaver" problem, given a positive integer n n, we construct all the Turing machines that have n n states. The number of such machines will always be finite. Next we eliminate those states. The number of such machines will always be finite. Next we eliminate those n n-state machines that get into an infinite loop (that is, never halt). Finally, we select the machine (one that does halt) that writes the largest number of 1s on its tape. The number of 1s that this Turing machine writes is called the busy beaver of n n. Rado showed that there is no algorithm-that is, no Turing machine-that can compute this function for all n ns. The crux of the problem is sorting out those n n-state machines that get into infinite loops. If we program a Turing machine to generate and simulate all possible n n-state Turing machines, this simulator itself gets into an infinite loop when it attempts to simulate one of the n n-state machines that gets into an infinite loop.

Despite its status as an unsolvable problem (and one of the most famous), we can determine the busy-beaver function for some n ns. (Interestingly, it is also an unsolvable problem to separate those n ns for which we can determine the busy beaver of n n from those for which we cannot.) For example, the busy beaver of 6 is easily determined to be 35. With seven states, a Turing machine can multiply, so the busy beaver of 7 is much bigger: 22,961. With eight states, a Turing machine can compute exponentials, so the busy beaver of 8 is even bigger: approximately 10 from those for which we cannot.) For example, the busy beaver of 6 is easily determined to be 35. With seven states, a Turing machine can multiply, so the busy beaver of 7 is much bigger: 22,961. With eight states, a Turing machine can compute exponentials, so the busy beaver of 8 is even bigger: approximately 1043. We can see that this is an "intelligent" function, in that it requires greater intelligence to solve for larger n ns.

By the time we get to 10, a Turing machine can perform types of calculations that are impossible for a human to follow (without help from a computer). So we were able to determine the busy beaver of 10 only with a computer's a.s.sistance. The answer requires an exotic notation to write down, in which we have a stack of exponents, the height of which is determined by another stack of exponents, the height of which is determined by another stack of exponents, and so on. Because a computer can keep track of such complex numbers, whereas the human brain cannot, it appears that computers will prove more capable of solving unsolvable problems than humans will.

The Criticism from Failure Rates

Jaron Lanier, Thomas Ray, and other observers all cite high failure rates of technology as a barrier to its continued exponential growth. For example, Ray writes:

The most complex of our creations are showing alarming failure rates. Orbiting satellites and telescopes, s.p.a.ce shuttles, interplanetary probes, the Pentium chip, computer operating systems, all seem to be pushing the limits of what we can effectively design and build through conventional approaches....Our most complex software (operating systems and telecommunications control systems) already contains tens of millions of lines of code. At present it seems unlikely that we can produce and manage software with hundreds of millions or billions of lines of code.32

First, we might ask what alarming failure rates Ray is referring to. As mentioned earlier, computerized systems of significant sophistication routinely fly and land our airplanes automatically and monitor intensive care units in hospitals, yet almost never malfunction. If alarming failure rates are of concern, they're more often attributable to human error. Ray alludes to problems with Intel microprocessor chips, but these problems have been extremely subtle, have caused almost no repercussions, and have quickly been rectified.

The complexity of computerized systems has indeed been scaling up, as we have seen, and moreover the cutting edge of our efforts to emulate human intelligence will utilize the self-organizing paradigms that we find in the human brain. As we continue our progress in reverse engineering the human brain, we will add new self-organizing methods to our pattern recognition and AI toolkit. As I have discussed, self-organizing methods help to alleviate the need for unmanageable levels of complexity. As I pointed out earlier, we will not need systems with "billions of lines of code" to emulate human intelligence.

It is also important to point out that imperfection is an inherent feature of any complex process, and that certainly includes human intelligence.

The Criticism from "Lock-In"

Jaron Lanier and other critics have cited the prospect of a "lock-in," a situation in which old technologies resist displacement because of the large investment in the infrastructure supporting them. They argue that pervasive and complex support systems have blocked innovation in such fields as transportation, which have not seen the rapid development that we've seen in cornputation.33 The concept of lock-in is not the primary obstacle to advancing transportation. If the existence of a complex support system necessarily caused lock-in, then why don't we see this phenomenon affecting the expansion of every aspect of the Internet? After all, the Internet certainly requires an enormous and complex infrastructure. Because it is specifically the processing and movement of information that is growing exponentially, however, one reason that an area such as transportation has reached a plateau (that is, resting at the top of an S-curve) is that many if not most of its purposes have been satisfied by exponentially growing communication technologies. My own organization, for example, has colleagues in different parts of the country, and most of our needs that in times past would have required a person or a package to be transported can be met through the increasingly viable virtual meetings (and electronic distribution of doc.u.ments and other intellectual creations) made possible by a panoply of communication technologies, some of which Lanier himself is working to advance. More important, we will see advances in transportation facilitated by the nanotechnology-based energy technologies I discussed in chapter 5. However, with increasingly realistic, high-resolution full-immersion forms of virtual reality continuing to emerge, our needs to be together will increasingly be met through computation and communication.

As I discussed in chapter 5, the full advent of MNT-based manufacturing will bring the law of accelerating returns to such areas as energy and transportation. Once we can create virtually any physical product from information and very inexpensive raw materials, these traditionally slow-moving industries will see the same kind of annual doubling of price-performance and capacity that we see in information technologies. Energy and transportation will effectively become information technologies.

We will see the advent of nanotechnology-based solar panels that are efficient, lightweight, and inexpensive, as well as comparably powerful fuel cells and other technologies to store and distribute that energy. Inexpensive energy will in turn transform transportation. Energy obtained from nanoengineered solar cells and other renewable technologies and stored in nanoengineered fuel cells will provide clean and inexpensive energy for every type of transportation. In addition, we will be able to manufacture devices-including flying machines of varying sizes-for almost no cost, other than the cost of the design (which needs to be amortized only once). It will be feasible, therefore, to build inexpensive small flying devices that can transport a package directly to your destination in a matter of hours without going through intermediaries such as shipping companies. Larger but still inexpensive vehicles will be able to fly people from place to place with nanoengineered microwings.

Information technologies are already deeply influential in every industry. With the full realization of the GNR revolutions in a few decades, every area of human endeavor will essentially comprise information technologies and thus will directly benefit from the law of accelerating returns.

The Criticism from Ontology: Can a Computer Be Conscious?

Because we do not understand the brain very well we are constantly tempted to use the latest technology as a model for trying to understand it. In my childhood we were always a.s.sured that the brain was a telephone switchboard. ("What else could it be?") I was amused to see that Sherrington, the great British neuroscientist, thought that the brain worked like a telegraph system. Freud often compared the brain to hydraulic and electromagnetic systems. Leibniz compared it to a mill, and I am told some of the ancient Greeks thought the brain functions like a catapult. At present, obviously, the metaphor is the digital computer.-JOHN R. SEARLE, "MINDS, BRAINS, AND SCIENCE"

Can a computer-a nonbiological intelligence-be conscious? We have first, of course, to agree on what the question means. As I discussed earlier, there are conflicting perspectives on what may at first appear to be a straightforward issue. Regardless of how we attempt to define the concept, however, we must acknowledge that consciousness is widely regarded as a crucial, if not essential, attribute of being human.34 John Searle, distinguished philosopher at the University of California at Berkeley, is popular among his followers for what they believe is a staunch defense of the deep mystery of human consciousness against trivialization by strong-AI "reductionists" like Ray Kurzweil. And even though I have always found Searle's logic in his celebrated Chinese Room argument to be tautological, I had expected an elevating treatise on the paradoxes of consciousness. Thus it is with some surprise that I find Searle writing statements such as,

"human brains cause consciousness by a series of specific neurobiological processes in the brain";"The essential thing is to recognize that consciousness is a biological process like digestion, lactation, photosynthesis, or mitosis";"The brain is a machine, a biological machine to be sure, but a machine all the same. So the first step is to figure out how the brain does it and then build an artificial machine that has an equally effective mechanism for causing consciousness"; and"We know that brains cause consciousness with specific biological mechanisms."35

So who is being the reductionist here? Searle apparently expects that we can measure the subjectivity of another ent.i.ty as readily as we measure the oxygen output of photosynthesis.

Searle writes that I "frequently cite IBM's Deep Blue as evidence of superior intelligence in the computer." Of course, the opposite is the case: I cite Deep Blue not to belabor the issue of chess but rather to examine the dear contrast it ill.u.s.trates between the human and contemporary machine approaches to the game. As I pointed out earlier, however, the pattern-recognition ability of chess programs is increasing, so chess machines are beginning to combine the a.n.a.lytical strength of traditional machine intelligence with more humanlike pattern recognition. The human paradigm (of self-organizing chaotic processes) offers profound advantages: we can recognize and respond to extremely subtle patterns. But we can build machines with the same abilities. That, indeed, has been my own area of technical interest.

Searle is best known for his Chinese Room a.n.a.logy and has presented various formulations of it over twenty years. One of the more complete descriptions of it appears in his 1992 book, The Rediscovery of the Mind The Rediscovery of the Mind:

I believe the best-known argument against strong AI was my Chinese room argument ... that showed that a system could instantiate a program so as to give a perfect simulation of some human cognitive capacity, such as the capacity to understand Chinese, even though that system had no understanding of Chinese whatever. Simply imagine that someone who understands no Chinese is locked in a room with a lot of Chinese symbols and a computer program for answering questions in Chinese. The input to the system consists in Chinese symbols in the form of questions; the output of the system consists in Chinese symbols in answer to the questions. We might suppose that the program is so good that the answers to the questions are indistinguishable from those of a native Chinese speaker. But all the same, neither the person inside nor any other part of the system literally understands Chinese; and because the programmed computer has nothing that this system does not have, the programmed computer, qua computer, does not understand Chinese either. Because the program is purely formal or syntactical and because minds have mental or semantic contents, any attempt to produce a mind purely with computer programs leaves out the essential features of the mind.36

Searle's descriptions ill.u.s.trate a failure to evaluate the essence of either brain processes or the nonbiological processes that could replicate them. He starts with the a.s.sumption that the "man" in the room doesn't understand anything because, after all, "he is just a computer," thereby illuminating his own bias. Not surprisingly Searle then concludes that the computer (as implemented by the man) doesn't understand. Searle combines this tautology with a basic contradiction: the computer doesn't understand Chinese, yet (according to Searle) can convincingly answer questions in Chinese. But if an ent.i.ty-biological or otherwise-really doesn't understand human language, it will quickly be unmasked by a competent interlocutor. In addition, for the program to respond convincingly, it would have to be as complex as a human brain. The observers would long be dead while the man in the room spends millions of years following a program many millions of pages long.

Most important, the man is acting only as the central processing unit, a small part of a system. While the man may not see it, the understanding is distributed across the entire pattern of the program itself and the billions of notes he would have to make to follow the program. I understand English, but none of my neurons do I understand English, but none of my neurons do. My understanding is represented in vast patterns of neurotransmitter strengths, synaptic clefts, and interneuronal connections. Searle fails to account for the significance of distributed patterns of information and their emergent properties.

A failure to see that computing processes are capable of being-just like the human brain-chaotic, unpredictable, messy, tentative, and emergent is behind much of the criticism of the prospect of intelligent machines that we hear from Searle and other essentially materialist philosophers. Inevitably Searle comes back to a criticism of "symbolic" computing: that orderly sequential symbolic processes cannot re-create true thinking. I think that's correct (depending, of course, on what level we are modeling an intelligent process), but the manipulation of symbols (in the sense that Searle implies) is not the only way to build machines, or computers.

So-called computers (and part of the problem is the word "computer," because machines can do more than "compute") are not limited to symbolic processing. Nonbiological ent.i.ties can also use the emergent self-organizing paradigm, which is a trend well under way and one that will become even more important over the next several decades. Computers do not have to use only 0 and 1, nor do they have to be all digital. Even if a computer is all digital, digital algorithms can simulate a.n.a.log processes to any degree of precision (or lack of precision). Machines can be ma.s.sively parallel. And machines can use chaotic emergent techniques just as the brain does.

The primary computing techniques that we have used in pattern-recognition systems do not use symbol manipulation but rather self-organizing methods such as those described in chapter 5 (neural nets, Markov models, genetic algorithms, and more complex paradigms based on brain reverse engineering). A machine that could really do what Searle describes in the Chinese Room argument would not merely be manipulating language symbols, because that approach doesn't work. This is at the heart of the philosophical sleight of hand underlying the Chinese Room. The nature of computing is not limited to manipulating logical symbols. Something is going on in the human brain, and there is nothing that prevents these biological processes from being reverse engineered and replicated in nonbiological ent.i.ties.

Adherents appear to believe that Searle's Chinese Room argument demonstrates that machines (that is, nonbiological ent.i.ties) can never truly understand anything of significance, such as Chinese. First, it is important to recognize that for this system-the person and the computer-to, as Searle puts it, "give a perfect simulation of some human cognitive capacity, such as the capacity to understand Chinese," and to convincingly answer questions in Chinese, it must essentially pa.s.s a Chinese Turing test. Keep in mind that we are not talking about answering questions from a fixed list of stock questions (because that's a trivial task) but answering any unantic.i.p.ated question or sequence of questions from a knowledgeable human interrogator.

Now, the human in the Chinese Room has little or no significance. He is just feeding things into the computer and mechanically transmitting its output (or, alternatively, just following the rules in the program). And neither the computer nor the human needs to be in a room. Interpreting Searle's description to imply that the man himself is implementing the program does not change anything other than to make the system far slower than real time and extremely error p.r.o.ne. Both the human and the room are irrelevant Both the human and the room are irrelevant. The only thing that is significant is the computer (either an electronic computer or the computer comprising the man following the program).

For the computer to really perform this "perfect simulation," it would indeed have to understand Chinese. According to the very premise it has "the capacity to understand Chinese," so it is then entirely contradictory to say that "the programmed computer ... does not understand Chinese."

A computer and computer program as we know them today as we know them today could not successfully perform the described task. So if we are to understand the computer to be like to day's computers, then it cannot fulfill the premise. The only way that it could do so would be if it had the depth and complexity of a human. Turing's brilliant insight in proposing his test was that convincingly answering any possible sequence of questions from an intelligent human questioner in a human language really probes all of human intelligence. A computer that is capable of accomplishing this-a computer that will exist a few decades from now-will need to be of human complexity or greater and will indeed understand Chinese in a deep way, because otherwise it would never be convincing in its claim to do so. could not successfully perform the described task. So if we are to understand the computer to be like to day's computers, then it cannot fulfill the premise. The only way that it could do so would be if it had the depth and complexity of a human. Turing's brilliant insight in proposing his test was that convincingly answering any possible sequence of questions from an intelligent human questioner in a human language really probes all of human intelligence. A computer that is capable of accomplishing this-a computer that will exist a few decades from now-will need to be of human complexity or greater and will indeed understand Chinese in a deep way, because otherwise it would never be convincing in its claim to do so.

Merely stating, then, that the computer "does not literally understand Chinese" does not make sense, for it contradicts the entire premise of the argument. To claim that the computer is not conscious is not a compelling contention, either. To be consistent with some of Searle's other statements, we have to conclude that we really don't know if it is conscious or not. With regard to relatively simple machines, including to day's computers, while we can't state for certain that these ent.i.ties are not conscious, their behavior, including their inner workings, doesn't give us that impression. But that will not be true for a computer that can really do what is needed in the Chinese Room. Such a machine will at least seem seem conscious, even if we cannot say definitively whether it is or not. But just declaring that it is obvious that the computer (or the entire system of the computer, person, and room) is not conscious is far from a compelling argument. conscious, even if we cannot say definitively whether it is or not. But just declaring that it is obvious that the computer (or the entire system of the computer, person, and room) is not conscious is far from a compelling argument.

In the quote above Searle states that "the program is purely formal or syntactical," But as I pointed out earlier, that is a bad a.s.sumption, based on Searle's failure to account for the requirements of such a technology. This a.s.sumption is behind much of Searle's criticism of AI. A program that is purely formal or syntactical will not be able to understand Chinese, and it won't "give a perfect simulation of some human cognitive capacity."

But again, we don't have to build our machines that way. We can build them in the same fashion that nature built the human brain: using chaotic emergent methods that are ma.s.sively parallel. Furthermore, there is nothing inherent in the concept of a machine that restricts its expertise to the level of syntax alone and prevents it from mastering semantics. Indeed, if the machine inherent in Searle's conception of the Chinese Room had not mastered semantics, it would not be able to convincingly answer questions in Chinese and thus would contradict Searle's own premise.

In chapter 4 I discussed the ongoing effort to reverse engineer the human brain and to apply these methods to computing platforms of sufficient power. So, like a human brain, if we teach a computer Chinese, it will understand Chinese. This may seem to be an obvious statement, but it is one with which Searle takes issue. To use his own terminology, I am not talking about a simulation per se but rather a duplication of the causal powers of the ma.s.sive neuron cl.u.s.ter that const.i.tutes the brain, at least those causal powers salient and relevant to thinking.

Will such a copy be conscious? I don't think the Chinese Room tells us anything about this question.

It is also important to point out that Searle's Chinese Room argument can be applied to the human brain itself. Although it is clearly not his intent, his line of reasoning implies that the human brain has no understanding. He writes: "The computer ... succeeds by manipulating formal symbols. The symbols themselves are quite meaningless: they have only the meaning we have attached to them. The computer knows nothing of this, it just shuffles the symbols." Searle acknowledges that biological neurons are machines, so if we simply subst.i.tute the phrase "human brain" for "computer" and "neurotransmitter concentrations and related mechanisms" for "formal symbols," we get:

The [human brain] ... succeeds by manipulating [neurotransmitter concentrations and related mechanisms]. The [neurotransmitter concentrations and related mechanisms] themselves are quite meaningless: they have only the meaning we have attached to them. The [human brain] knows nothing of this, it just shuffles the [neurotransmitter concentrations and related mechanisms].

Of course, neurotransmitter concentrations and other neural details (for example, interneuronal connection and neurotransmitter patterns) have no meaning in and of themselves. The meaning and understanding that emerge in the human brain are exactly that: an emergent emergent property of its complex patterns of activity. The same is true for machines. Although "shuffling symbols" does not have meaning in and of itself, the emergent patterns have the same potential role in nonbiological systems as they do in biological systems such as the brain. Hans Moravec has written, "Searle is looking for understanding in the wrong places....[He] seemingly cannot accept that real meaning can exist in mere patterns. property of its complex patterns of activity. The same is true for machines. Although "shuffling symbols" does not have meaning in and of itself, the emergent patterns have the same potential role in nonbiological systems as they do in biological systems such as the brain. Hans Moravec has written, "Searle is looking for understanding in the wrong places....[He] seemingly cannot accept that real meaning can exist in mere patterns.37 Let's address a second version of the Chinese Room. In this conception the room does not include a computer or a man simulating a computer but has a room full of people manipulating slips of paper with Chinese symbols on them-essentially, a lot of people simulating a computer. This system would convincingly answer questions in Chinese, but none of the partic.i.p.ants would know Chinese, nor could we say that the whole system really knows Chinese-at least not in a conscious way. Searle then essentially ridicules the idea that this "system" could be conscious. What are we to consider conscious, he asks: the slips of paper? The room?

One of the problems with this version of the Chinese Room argument is that it does not come remotely close to really solving the specific problem of answering questions in Chinese. Instead it is really a description of a machinelike process that uses the equivalent of a table lookup, with perhaps some straightforward logical manipulations, to answer questions. It would be able to answer a limited number of canned questions, but if it were to answer any arbitrary question that it might be asked, it would really have to understand Chinese in the same way that a Chinese-speaking person does. Again, it is essentially being asked to pa.s.s a Chinese Turing test, and as such, would have to be as clever, and about as complex, as a human brain. Straightforward table lookup algorithms are simply not going to achieve that.

If we want to re-create a brain that understands Chinese using people as little cogs in the re-creation, we would really need billions of people simulating the processes in a human brain (essentially the people would be simulating a computer, which would be simulating human brain methods). This would require a rather large room, indeed. And even if extremely efficiently organized, this system would run many thousands of times slower than the Chinese-speaking brain it is attempting to re-create.

Now, it's true that none of these billions of people would need to know anything about Chinese, and none of them would necessarily know what is going on in this elaborate system. But that's equally true of the neural connections in a real human brain. None of the hundred trillion connections in my brain knows anything about this book I am writing, nor do any of them know English, nor any of the other things that I know. None of them is conscious of this chapter, nor of any of the things I am conscious of. Probably none of them is conscious at all. But the entire system of them-that is, Ray Kurzweil-is conscious. At least I'm claiming that I'm conscious (and so far, these claims have not been challenged).

So if we scale up Searle's Chinese Room to be the rather ma.s.sive "room" it needs to be, who's to say that the entire system of billions of people simulating a brain that knows Chinese isn't conscious? Certainly it would be correct to say that such a system knows Chinese. And we can't say that it is not conscious any more than we can say that about any other brain process. We can't know the subjective experience of another ent.i.ty (and in at least some of Searle's other writings, he appears to acknowledge this limitation). And this ma.s.sive multibillion-person "room" is an ent.i.ty. And perhaps it is conscious. Searle is just declaring ipso facto that it isn't conscious and that this conclusion is obvious. It may seem that way when you call it a room and talk about a limited number of people manipulating a small number of slips of paper. But as I said, such a system doesn't remotely work.

Another key to the philosophical confusion implicit in the Chinese Room argument is specifically related to the complexity and scale of the system. Searle says that whereas he cannot prove that his typewriter or tape recorder is not conscious, he feels it is obvious that they are not. Why is this so obvious? At least one reason is because a typewriter and a tape recorder are relatively simple ent.i.ties.

But the existence or absence of consciousness is not so obvious in a system that is as complex as the human brain-indeed, one that may be a direct copy of the organization and "causal powers" of a real human brain. If such a "system" acts human and knows Chinese in a human way, is it conscious? Now the answer is no longer so obvious. What Searle is saying in the Chinese Room argument is that we take a simple "machine" and then consider how absurd it is to consider such a simple machine to be conscious. The fallacy has everything to do with the scale and complexity of the system. Complexity alone does not necessarily give us consciousness, but the Chinese Room tells us nothing about whether or not such a system is conscious.

Kurzweil's Chinese Room. I have my own conception of the Chinese Room-call it Ray Kurzweil's Chinese Room. I have my own conception of the Chinese Room-call it Ray Kurzweil's Chinese Room.

In my thought experiment there is a human in a room. The room has decorations from the Ming dynasty, including a pedestal on which sits a mechanical typewriter. The typewriter has been modified so that its keys are marked with Chinese symbols instead of English letters. And the mechanical linkages have been cleverly altered so that when the human types in a question in Chinese, the typewriter does not type the question but instead types the answer to the question. Now, the person receives questions in Chinese characters and dutifully presses the appropriate keys on the typewriter. The typewriter types out not the question, but the appropriate answer. The human then pa.s.ses the answer outside the room.

So here we have a room with a human in it who appears from the outside to know Chinese yet clearly does not. And clearly the typewriter does not know Chinese, either. It is just an ordinary typewriter with its mechanical linkages modified. So despite the fact that the man in the room can answer questions in Chinese, who or what can we say truly knows Chinese? The decorations?

Now, you might have some objections to my Chinese Room.