Редактирование: Transhumanist FAQ Version 3 (раздел)

===What is superintelligence?===

A superintelligent intellect (a superintelligence, sometimes called “ultraintelligence”) is one that has the capacity to radically outperform the best human brains in practically every field, including scientific creativity, general wisdom, and social skills.

Sometimes a distinction is made between weak and strong superintelligence. Weak superintelligence is what you would get if you could run a human intellect at an accelerated clock speed, such as by uploading it to a fast computer [see “What is uploading?”]. If the upload’s clock-rate were a thousand times that of a biological brain, it would perceive reality as being slowed down by a factor of a thousand. It would think a thousand times more thoughts in a given time interval than its biological counterpart.

Strong superintelligence refers to an intellect that is not only faster than a human brain but also smarter in a qualitative sense. No matter how much you speed up your dog’s brain, you’re not going to get the equivalent of a human intellect. Analogously, there might be kinds of smartness that wouldn’t be accessible to even very fast human brains given their current capacities. Something as simple as increasing the size or connectivity of our neuronal networks might give us some of these capacities. Other improvements may require wholesale reorganization of our cognitive architecture or the addition of new layers of cognition on top of the old ones.

However, the distinction between weak and strong superintelligence may not be clear-cut. A sufficiently long-lived human who didn’t make any errors and had a sufficient stack of scrap paper at hand could in principle compute any Turing computable function. (According to Church’s thesis, the class of Turing computable functions is identical to the class of physically computable functions.)

Many but not all transhumanists expect that superintelligence will be created within the first half of this century. Superintelligence requires two things: hardware and software.

Chip-manufacturers planning the next generation of microprocessors commonly rely on a well-known empirical regularity known as Moore’s Law. In its original 1965-formulation by Intel co-founder Gordon Moore, it stated that the number of components on a chip doubled every year. In contemporary use, the “law” is commonly understood as referring more generally to a doubling of computing power, or of computing power per dollar. For the past couple of years, the doubling time has hovered between 18 months and two years.

The human brain’s processing power is difficult to determine precisely, but common estimates range from 1014 instructions per second (IPS) up to 1017 IPS or more. The lower estimate, derived by Carnegie Mellon robotics professor Hans Moravec, is based on the computing power needed to replicate the signal processing performed by the human retina and assumes a significant degree of software optimization. The 1017 IPS estimate is obtained by multiplying the number of neurons in a human brain (~100 billion) with the average number of synapses per neuron (~1,000) and with the average spike rate (~100 Hz), and assuming ~10 instructions to represent the effect on one action potential traversing one synapse. An even higher estimate would be obtained e.g. if one were to suppose that functionally relevant and computationally intensive processing occurs within compartments of a dendrite tree.

Most experts, Moore included, think that computing power will continue to double about every 18 months for at least another two decades. This expectation is based in part on extrapolation from the past and in part on consideration of developments currently underway in laboratories. The fastest computer under construction is IBM’s Blue Gene/L, which when it is ready in 2005 is expected to perform ~2*1014 IPS. Thus it appears quite likely that human-equivalent hardware will have been achieved within not much more than a couple of decades.

How long it will take to solve the software problem is harder to estimate. One possibility is that progress in computational neuroscience will teach us about the computational architecture of the human brain and what learning rules it employs. We can then implement the same algorithms on a computer. In this approach, the superintelligence would not be completely specified by the programmers but would instead have to grow by learning from experience the same way a human infant does. An alternative approach would be to use genetic algorithms and methods from classical AI. This might result in a superintelligence that bears no close resemblance to a human brain. At the opposite extreme, we could seek to create a superintelligence by uploading a human intellect and then accelerating and enhancing it [see “What is uploading?”]. The outcome of this might be a superintelligence that is a radically upgraded version of one particular human mind.

The arrival of superintelligence will clearly deal a heavy blow to anthropocentric worldviews. Much more important than its philosophical implications, however, would be its practical effects. Creating superintelligence may be the last invention that humans will ever need to make, since superintelligences could themselves take care of further scientific and technological development. They would do so more effectively than humans. Biological humanity would no longer be the smartest life form on the block.

The prospect of superintelligence raises many big issues and concerns that we should think deeply about in advance of its actual development. The paramount question is: What can be done to maximize the chances that the arrival of superintelligence will benefit rather than harm us? The range of expertise needed to address this question extends far beyond the community of AI researchers. Neuroscientists, economists, cognitive scientists, computer scientists, philosophers, ethicists, sociologists, science-fiction writers, military strategists, politicians, legislators, and many others will have to pool their insights if we are to deal wisely with what may be the most important task our species will ever have to tackle.

Many transhumanists would like to become superintelligent themselves. This is obviously a long-term and uncertain goal, but it might be achievable either through uploading and subsequent enhancement or through the gradual augmentation of our biological brains, by means of future nootropics (cognitive enhancement drugs), cognitive techniques, IT tools (e.g. wearable computers, smart agents, information filtering systems, visualization software, etc.), neural-computer interfaces, or brain implants.

References:

Moravec, H. Mind Children (Harvard: Harvard University Press, 1989).

Bostrom, N. “How Long Before Superintelligence?” International Journal of Futures Studies. Vol. 2. (1998).