Microsoft Word Kurzweil, Ray The Singularity Is Near doc

The Age of Spiritual Machines
that molecular computing would be the sixth major 
computing paradigm, the idea was still controversial. There has been so much progress in the past five years that there 
has been a sea change in attitude among experts, and this is now a mainstream view. We already have proofs of 
concept for all of the major requirements for three-dimensional molecular computing: single-molecule transistors, 
memory cells based on atoms, nanowires, and methods to self-assemble and self-diagnose the trillions (potentially 
trillions of trillions) of components. 
Contemporary electronics proceeds from the design of detailed chip layouts to photolithography to the 
manufacturing of chips in large, centralized factories. Nanocircuits are more likely to be created in small chemistry 
flasks, a development that will be another important step in the decentralization of our industrial infrastructure and will 
maintain the law of accelerating returns through this century and beyond. 
The Computational Capacity of the Human Brain 
It may seem rash to expect fully intelligent machines in a few decades, when the computers have barely 
matched insect mentality in a half-century of development. Indeed, for that reason, many long-time artificial 
intelligence researchers scoff at the suggestion, and offer a few centuries as a more believable period. But 
there are very good reasons why things will go much faster in the next fifty years than they have in the last 
fifty. . . . Since 1990, the power available to individual AI and robotics programs has doubled yearly, to 30 
MIPS by 1994 and 500 MIPS by 1998. Seeds long ago alleged barren are suddenly sprouting. Machines read 
text, recognize speech, even translate languages. Robots drive cross-country, crawl across Mars, and trundle 
down office corridors. In 1996 a theorem-proving program called EQP running five weeks on a 50 MIPS 
computer at Argonne National Laboratory found a proof of a Boolean algebra conjecture by Herbert Robbins 
that had eluded mathematicians for sixty years. And it is still only Spring. Wait until Summer. 
—H
ANS 
M
ORAVEC
,
“W
HEN 
W
ILL 
C
OMPUTER 
H
ARDWARE 
M
ATCH 
T
HE 
H
UMAN 
B
RAIN
?”
1997 

---

What is the computational capacity of a human brain? A number of estimates have been made, based on replicating the 
functionality of brain regions that have been reverse engineered (that is, the methods understood) at human levels of 
performance. Once we have an estimate of the computational capacity for a particular region, we can extrapolate that 
capacity to the entire brain by considering what portion of the brain that region represents. These estimates are based 
on functional simulation, which replicates the overall functionality of a region rather than simulating each neuron and 
interneuronal connection in that region. 
Although we would not want to rely on any single calculation, we find that various assessments of different 
regions of the brain all provide reasonably close estimates for the entire brain. The following are order-of-magnitude 
estimates, meaning that we are attempting to determine the appropriate figures to the closest multiple of ten. The fact 
that different ways of making the same estimate provide similar answers corroborates the approach and indicates that 
the estimates are in an appropriate range. 
The prediction that the Singularity—an expansion of human intelligence by a factor of trillions through merger 
with its nonbiological form—will occur within the next several decades does not depend on the precision of these 
calculations. Even if our estimate of the amount of computation required to simulate the human brain was too 
optimistic (that is, too low) by a factor of even one thousand (which I believe is unlikely), that would delay the 
Singularity by only about eight years.
34
A factor of one million would mean a delay of only about fifteen years, and a 
factor of one billion would be a delay of about twenty-one years.
35
Hans Moravec, legendary roboticist at Carnegie Mellon University, has analyzed the transformations performed 
by the neural image-processing circuitry contained in the retina.
36
The retina is about two centimeters wide and a half 
millimeter thick. Most of the retina's depth is devoted to capturing an image, but one fifth of it is devoted to image 
processing, which includes distinguishing dark and light, and detecting motion in about one million small regions of 
the image. 
The retina, according to Moravec's analysis, performs ten million of these edge and motion detections each 
second. Based on his several decades of experience in creating robotic vision systems, he estimates that the execution 
of about one hundred computer instructions is required to re-create each such detection at human levels of 
performance, meaning that replicating the image-processing functionality of this portion of the retina requires 1,000 
MIPS. The human brain is about 75,000 times heavier than the 0.02 grams of neurons in this portion of the retina, 
resulting in an estimate of about 10
14
(100 trillion) instructions per second for the entire brain.
37 
Another estimate comes from the work of Lloyd Watts and his colleagues on creating functional simulations of 
regions of the human auditory system, which I discuss further in chapter 4.
38
One of the functions of the software 
Watts has developed is a task called "stream separation," which is used in teleconferencing and other applications to 
achieve telepresence (the localization of each participant in a remote audio teleconference), To accomplish this, Watts 
explains, means "precisely measuring the time delay between sound sensors that are separated in space and that both 
receive the sound." The process involves pitch analysis, spatial position, and speech cues, including language-specific 
cues. "One of the important cues used by humans for localizing the position of a sound source is the Interaural Time 
Difference (ITD), that is, the difference in time of arrival of sounds at the two ears."
39
Watts's own group has created functionally equivalent re-creations of these brain regions derived from reverse 
engineering. He estimates that 10
11
cps are required to achieve human-level localization of sounds. The auditory cortex 
regions responsible for this processing comprise at least 0.1 percent of the brain's neurons. So we again arrive at a 
ballpark estimate of around 10
14
cps 
°
10
3
). 
Yet another estimate comes from a simulation at the University of Texas that represents the functionality of a 
cerebellum region containing 10
4
neurons; this required about 10
8
cps, or about 10
4
cps per neuron. Extrapolating this 
over an estimated 10
11
neurons results in a figure of about 10
15
cps for the entire brain. 
We will discuss the state of human-brain reverse engineering later, but it is clear that we can emulate the 
functionality of brain regions with less computation than would be required to simulate the precise nonlinear operation 
of each neuron and all of the neural components (that is, all of the complex interactions that take place inside each 

---

neuron). We come to the same conclusion when we attempt to simulate the functionality of organs in the body. For 
example, implantable devices are being tested that simulate the functionality of the human pancreas in regulating 
insulin levels.
40
These devices work by measuring glucose levels in the blood and releasing insulin in a controlled 
fashion to keep the levels in an appropriate range. While they follow a method similar to that of a biological pancreas, 
they do not, however, attempt to simulate each pancreatic islet cell, and there would be no reason to do so. 
These estimates all result in comparable orders of magnitude (10
14
to 10
15
cps). Given the early stage of human-
brain reverse engineering, I will use a more conservative figure of 10
16
cps for our subsequent discussions. 
Functional simulation of the brain is sufficient to re-create human powers of pattern recognition, intellect, and 
emotional intelligence. On the other hand, if we want to "upload" a particular person's personality (that is, capture all 
of his or her knowledge, skills, and personality, a concept I will explore in greater detail at the end of chapter 4), then 
we may need to simulate neural processes at the level of individual neurons and portions of neurons, such as the soma 
(cell body), axon (output connection), dendrites (trees of incoming connections), and synapses (regions connecting 
axons and dendrites). For this, we need to look at detailed models of individual neurons. The "fan out" (number of 
interneuronal connections) per neuron is estimated at 10
3
. With an estimated 10
11
neurons, that's about 10
14
connections. With a reset time of five milliseconds, that comes to about 10
16
synaptic transactions per second. 
Neuron-model simulations indicate the need for about 10
3
calculations per synaptic transaction to capture the 
nonlinearities (complex interactions) in the dendrites and other neuron regions, resulting in an overall estimate of 
about 10
19
cps for simulating the human brain at this level.
41
We can therefore consider this an upper bound, but 10
14
to 10
16
cps to achieve functional equivalence of all brain regions is likely to be sufficient. 
IBM's Blue Gene/L supercomputer, now being built and scheduled to be completed around the time of the 
publication of this book, is projected to provide 360 trillion calculations per second (3.6 
°
10
14 
cps).
42
This figure is 
already greater than the lower estimates described above. Blue Gene/L will also have around one hundred terabytes 
(about 10
15
bits) of main storage, more than our memory estimate for functional emulation of the human brain (see 
below). In line with my earlier predictions, supercomputers will achieve my more conservative estimate of 10
16
cps for 
functional human-brain emulation by early in the next decade (see the "Supercomputer Power" figure on p. 71).