Microsoft Word Kurzweil, Ray The Singularity Is Near doc

Human Memory Capacity. 
How does computational capacity compare to human memory capacity? It turns out that 
we arrive at similar time-frame estimates if we look at human memory requirements. The number of "chunks" of 
knowledge mastered by an expert in a domain is approximately 10
5
for a variety of domains. These chunks represent 
patterns (such as faces) as well as specific knowledge. For example, a world-class chess master is estimated to have 
mastered about 100,000 board positions. Shakespeare used 29,000 words but close to 100,000 meanings of those 
words. Development of expert systems in medicine indicate that humans can master about 100,000 concepts in a 
domain. If we estimate that this "professional" knowledge represents as little as 1 percent of the overall pattern and 
knowledge store of a human, we arrive at an estimate of 10
7
chunks. 
Based on my own experience in designing systems that can store similar chunks of knowledge in either rule-based 
expert systems or self-organizing pattern-recognition systems, a reasonable estimate is about 10
6
bits per chunk 
(pattern or item of knowledge), for a total capacity of 10
13
(10 trillion) bits for a human's functional memory. 
According to the projections from the ITRS road map (see RAM chart on p. 57), we will be able to purchase 10
13
bits of memory for one thousand dollars by around 2018. Keep in mind that this memory will be millions of times 
faster than the electrochemical memory process used in the human brain and thus will be far more effective. 
Again, if we model human memory on the level of individual interneuronal connections, we get a higher estimate. 
We can estimate about 10
4
bits per connection to store the connection patterns and neurotransmitter concentrations. 
With an estimated 10
14
connections, that comes to 10
18
(a billion billion) bits. 
Based on the above analyses, it is reasonable to expect the hardware that can emulate human-brain functionality to 
be available for approximately one thousand dollars by around 2020. As we will discuss in chapter 4, the software that 
will replicate that functionality will take about a decade longer. However, the exponential growth of the price-
performance, capacity, and speed of our hardware technology will continue during that period, so by 2030 it will take 
a village of human brains (around one thousand) to match a thousand dollars' worth of computing. By 2050, one 
thousand dollars of computing will exceed the processing power of all human brains on Earth. Of course, this figure 
includes those brains still using only biological neurons. 
While human neurons are wondrous creations, we wouldn't (and don't) design computing circuits using the same 
slow methods. Despite the ingenuity of the designs evolved through natural selection, they are many orders of 
magnitude less capable than what we will be able to engineer. As we reverse engineer our bodies and brains, we will 
be in a position to create comparable systems that are far more durable and that operate thousands to millions of times 
faster than our naturally evolved systems. Our electronic circuits are already more than one million times faster than a 
neuron's electrochemical processes, and this speed is continuing to accelerate. 
Most of the complexity of a human neuron is devoted to maintaining its life-support functions, not its information-
processing capabilities. Ultimately, we will be able to port our mental processes to a more suitable computational 
substrate. Then our minds won't have to stay so small. 
The Limits of Computation 

---

If a most efficient supercomputer works all day to compute a weather simulation problem, what is the 
minimum amount of energy that must be dissipated according to the laws of physics? The answer is actually 
very simple to calculate, since it is unrelated to the amount of computation. The answer is always equal to 
zero. 
—E
DWARD 
F
REDKIN
,
P
HYSICIST
45 
We've already had five paradigms (electromechanical calculators, relay-based computing, vacuum tubes, discrete 
transistors, and integrated circuits) that have provided exponential growth to the price-performance and capabilities of 
computation. Each time a paradigm reached its limits, another paradigm took its place. We can already see the outlines 
of the sixth paradigm, which will bring computing into the molecular third dimension. Because computation underlies 
the foundations of everything we care about, from the economy to human intellect and creativity, we might well 
wonder: are there ultimate limits to the capacity of matter and energy to perform computation? If so, what are these 
limits, and how long will it take to reach them? 
Our human intelligence is based on computational processes that we are learning to understand. We will 
ultimately multiply our intellectual powers by applying and extending the methods of human intelligence using the 
vastly greater capacity of nonbiological computation. So to consider the ultimate limits of computation is really to ask: 
what is the destiny of our civilization? 
A common challenge to the ideas presented in this book is that these exponential trends must reach a limit, as 
exponential trends commonly do. When a species happens upon a new habitat, as in the famous example of rabbits in 
Australia, its numbers grow exponentially for a while. But it eventually reaches the limits of that environment's ability 
to support it. Surely the processing of information must have similar constraints. It turns out that, yes, there are limits 
to computation based on the laws of physics. But these still allow for a continuation of exponential growth until 
nonbiological intelligence is trillions of trillions of times more powerful than all of human civilization today, 
contemporary computers included. 
A major factor in considering computational limits is the energy requirement. The energy required per MIPS for 
computing devices has been falling exponentially, as shown in the following figure.
46 

---

However, we also know that the number of MIPS in computing devices has been growing exponentially. The 
extent to which improvements in power usage have kept pace with processor speed depends on the extent to which we 
use parallel processing. A larger number of less-powerful computers can inherently run cooler because the 
computation is spread out over a larger area. Processor speed is related to voltage, and the power required is 
proportional to the square of the voltage. So running a processor at a slower speed significantly reduces power 
consumption. If we invest in more parallel processing rather than faster single processors, it is feasible for energy 
consumption and heat dissipation to keep pace with the growing MIPS per dollar, as the figure "Reduction in Watts 
per MIPS" shows. 
This is essentially the same solution that biological evolution developed in the design of animal brains. Human 
brains use about one hundred trillion computers (the interneuronal connections, where most of the processing takes 
place). But these processors are very low in computational power and therefore run relatively cool. 
Until just recently Intel emphasized the development of faster and faster single-chip processors, which have been 
running at increasingly high temperatures. Intel is gradually changing its strategy toward parallelization by putting 
multiple processors on a single chip. We will see chip technology move in this direction as a way of keeping power 
requirements and heat dissipation in check.
47