Today's Computers, Intelligent Machines and Our Future, Hans Moravec, Stanford AI Lab, 1978

Today's Computers, Intelligent Machines and Our Future

Hans Moravec
Stanford University
July 21, 1976
this version
1978

Introduction:

The unprecedented opportunities for experiments in complexity
presented by the first modern computers in the late 1940's raised
hopes in early computer scientists (eg. John von Neumann and Alan
Turing) that the ability to think, our greatest asset in our dealings
with the world, might soon be understood well enough to be
duplicated. Success in such an endeavor would extend mankind's mind in
the same way that the development of energy machinery extended his
muscles.

In the thirty years since then computers have become vastly
more capable, but the goal of human performance in most areas seems as
elusive as ever, in spite of a great deal of effort. The last ten
years, in particular, has seen thousands of people years devoted
directly to the problem, referred to as Artificial Intelligence or
AI. Attempts have been made to develop computer programs which do
mathematics, computer programming and common sense reasoning, are able
to understand natural languages and interpret scenes seen through
cameras and spoken language heard through microphones and to play
games humans find challenging.

There has been some progress. Samuel's checker program can
occasionally beat checker champions. Chess programs regularly play at
good amateur level, and in March 1977 a chess program from
Northwestern University, running on a CDC Cyber-176 (which is about 20
times as fast as previous computers used to play chess) won the
Minnesota Open Championship, against a slate of class A and expert
players. A ten year effort at MIT has produced a system, Mathlab,
capable of doing symbolic algebra, trigonometry and calculus
operations better in many ways than most humans experienced in those
fields. Programs exist which can understand English sentences with
restricted grammar and vocabulary, given the letter sequence, or
interpret spoken commands from hundred word vocabularies. Some can do
very simple visual inspection tasks, such as deciding whether or not a
screw is at the end of a shaft. The most difficult tasks to automate,
for which computer performance to date has been most disappointing,
are those that humans do most naturally, such as seeing, hearing and
common sense reasoning.

A major reason for the difficulty has become very clear to me
in the course of my work on computer vision. It is simply that the
machines with which we are working are still a hundred thousand to a
million times too slow to match the performance of human nervous
systems in those functions for which humans are specially wired. This
enormous discrepancy is distorting our work, creating problems where
there are none, making others impossibly difficult, and generally
causing effort to be misdirected.

In the early days of AI the thought that existing machines
might be much too small was widespread, but people hoped that clever
mathematics and advancing computer technology could soon make up the
difference. The idea that available compute power might still be
vastly inadequate has since been swept under the rug, due to wishful
thinking and a feeling that there was nothing to be done about it
anyway and that voicing such an opinion could cause AI to be
considered impractical, resulting in reduced funding. This attitude
has had some bad effects, one of them being that AI research has been
centered on computers less powerful than absolutely necessary.

The first section of this essay discusses natural
intelligence. It notes two major branches of the animal kingdom in
which intelligence evolved independently, and suggests that it is
easier to construct than is sometimes assumed.

The second part compares the information processing ability of
present computers with intelligent nervous systems. The factor of one
million is derived in two different ways.

Section three examines the development of electronics, and
concludes the state of the art can provide more power than is now
available, and that the one million gap could be closed in ten years.

Part four introduces some hardware and software aspects of a
system which would be able to make use of the advancing technology,
providing a means for achieving human equivalence, perhaps by the next
decade.

Part five considers the implications of the emergence of
intelligent machines, and concludes that they are the final step in a
revolution in the nature of life. Classical evolution based on DNA,
random mutations and natural selection may be completely replaced by
the much faster process of intelligence mediated cultural and
technological evolution.

Section 1: The Natural History of Intelligence

Product lines:
Natural evolution has produced a continuum of complexities of
behavior, from the mechanical simplicity of viruses to the magic of
mammals. In the higher animals most of the complexity resides in the
nervous system.

Evolution of the brain began in early multi-celled animals a
billion years ago with the development of cells capable of
transmitting electrochemical signals. Because neurons are more
localized than hormones they allow a greater variety of signals in a
given volume. They also provide evolution with a more uniform medium
for experiments in complexity.

The advantages of implementing behavioral complexity in neural
nets seem to have been overwhelming, since all modern animals more
than a few cells in size have them [animal refs.].

Two major branches in the animal kingdom, vertebrates and
mollusks, contain species which can be considered intelligent. Both
stem from one of the earliest multi-celled organisms, an animal
something like a hydra made of a double layer of cells and possessing
a primitive nerve net.

Most mollusks are intellectually unimpressive sessile
shellfish, but one branch, the cephalopods, possesses high mobility,
large brains and imaging eyes. These structures evolved independently
of the corresponding equipment in vertebrates and there are
fascinating differences. The optic nerve connects to the back of the
retina, so there is no blind spot. The brain is annular, forming a
ring encircling the esophagus. The circulatory system, also
independently evolved, has three blood pumps, a systemic heart pumping
oxygenated blood to the tissues and two gill hearts, each pumping
venous blood to one gill. The oxygen carrier is a green copper
compound called hemocyanin, evolved from an earlier protein that also
became hemoglobin.

These animals have some unique abilities. Shallow water
octopus and squid are covered by a million individually controlled
color changing effectors called chromatophores, whose functions are
camouflage and communication. The capabilities of this arrangement
have been demonstrated by a cuttlefish accurately imitating a
checkerboard it was placed upon, and an octopus in flight which
produced a pattern like the seaweed it was traversing, coruscating
backward along the length of its body, diverting the eye from the true
motion. Deep sea squid have photophores capable of generating large
quantities of multicolored light. Some are as complex as eyes,
containing irises and lenses [squid]. The light show is modulated by
emotions in major and subtle ways. There has been little study of
these matters, but this must provide means of social interaction.
Since they also have good vision, there is the potential for high
bandwidth communication.

Cephalopod intelligence has not been extensively investigated,
but a few controlled experiments indicate rapid learning in small
octopus [Boycott]. The Cousteau film in the references shows an
octopus' response to a problem requiring a two stage solution. A
fishbowl containing a lobster is sealed with a cork and dropped into
the water near it. The octopus is attracted, and spends a long while
alternately probing the container in various ways and returning to its
lair in iridescent frustration. On the final iteration it exits its
little hole in the ground and unhesitatingly wraps three tentacles
around the bowl, and one about the cork, and pulls. The cork shoots
to the surface and the octopus eats. The Time-Life film contains a
similar sequence, with a screw top instead of a cork! If small
octopus have almost mammalian behavior, what might giant deep sea
squid be capable of? The behavior of these large brained, apparently
shy, animals has virtually never been observed.

Birds are more closely related to humans than are cephalopods,
their common ancestor with us being a 300 million year old early
reptile. Size-limited by the dynamics of flying, some birds have
reached an intellectual level comparable to the highest mammals.

Crows and ravens are notable for frequently outwitting people.
Their intuitive number sense (ability to perceive the cardinality of a
set without counting) extends to seven, as opposed to three or four in
us. Such a sense is useful for keeping track of the number of eggs in
a nest. Experiments have shown [Stettner] that most birds are more
capable of high order "reversal" and "learning set" learning than all
mammals except the higher primates. In mammals these abilities
increase with increasing cerebral cortex size. In birds the same
functions depend on areas not present in mammalian brains, forebrain
regions called the "Wulst" and the hyperstriatum. The cortex is small
and relatively unimportant. Clearly this is another case of
independent evolution of similar mental functions. Penguins, now
similar to seals in behavior and habitat, might be expected to become
fully aquatic, and evolve analogously to the great whales.

The cetaceans are related to us through a small 30 million
year old primitive mammal. Some species of dolphin have body and brain
masses identical to ours, and archaeology reveals they have been this
way several times as long. They are as good as us at many kinds of
problem solving, and perhaps at language. The references contain many
anecdotes, and describe a few controlled experiments, showing that
dolphins can grasp and communicate complex ideas. Killer whales have
brains seven times human size, and their ability to formulate plans is
better than the dolphins', on whom they occasionally feed. Sperm
whales, though not the largest animals, have the world's largest
brains. There may be intelligence mediated conflict with large squid,
their main food supply.

Elephants have brains about five times human size, matriarchal
tribal societies, and complex behavior. Indian domestic elephants
usually learn 500 commands, limited by the range of tractor-like tasks
their owners need done, and form voluntary mutual benefit
relationships with their trainers, exchanging labor for baths. They
can solve problems such as how to sneak into a plantation at night to
steal bananas, after having been belled (answer: stuff mud into the
bells). And they remember for decades. Inconvenience and cost has
prevented more elephant research.

The apes are our cousins. Chimps and gorillas can learn to
use tools and to communicate with human sign languages at a retarded
level. As chimps have one third, and gorillas one half, human
brainsize, similar results should be achievable with the larger
brained, but less human-like animals. Though no other species has
managed to develop a technological culture, it may be that some of
them can be made partners in ours, accelerating its evolution with
their unique capabilities.

Nervous System Size and Intelligence

A feature shared by all living organisms whose behavior is
complex enough to indicate near-human intelligence is a nervous system
of a hundred billion neurons. Imaging vision requires a billion
neurons. A million brain cells usually permits fast and interesting,
but stereotyped, behavior as in a bee. A thousand is adequate for slow
moving animals with minimal sensory input, such as slugs and worms. A
hundred runs most sessile animals. The portions of nervous systems
for which tentative wiring diagrams have been obtained (eg. much of
the brain of the large neuroned sea slug, Aplysia, the flight
controller of the locust and the early stages of some vertebrate
visual systems) reveal that the neurons are configured into efficient,
clever, assemblies. This should not be surprising, as unnecessary
redundancy means unnecessary metabolic load, a distinct selective
disadvantage.

Time before present Representative Creatures Significant events

FIGURE: Highlights in the evolution of terrestrial intelligence.
The distance along the edge of the tree is proportional
to the square root of the time from the present. This
seems to space things nicely.

Evolution has stumbled on many ways of speeding up its own
progress, since species that adapt more quickly have a selective
advantage. Most of these speedups, such as sex and dying of old age,
are refinements of one of the oldest, the encoding of genetic
information in the easily mutated and modular DNA molecule. In the
last few million years the genetically evolved ability of animals,
especially mammals, to learn a significant fraction of their behavior
after birth has provided a new medium for growth of complexity.
Modern man, though perhaps not the most individually intelligent
animal on the planet, is the species in which this cultural evolution
seems to have had the greatest effect, making human culture the most
potent force on the earth's surface.

Our cultural and technological evolution has proceeded by
massive interchange of ideas and information, trial and error guided
by the ability to predict the outcome of simple situations, and other
techniques mediated by our intelligence. The process is self
reinforcing because its consequences, such as improved communication
methods and increased wealth and population, allow more experiments
and faster cross fertilization among different lines of inquiry. Many
of its techniques have not been available to biological evolution.
The effect is that present day global civilization is developing
capabilities orders of magnitude faster. Of course biological
evolution has had a massive head start.

Although cultural evolution has developed methods beyond those
of its genetic counterpart, the overall process is essentially the
same. It involves trying large numbers of possibilities, selecting
the best ones, and combining successes from different lines of
investigation. This requires time and other finite resources.

Finding the optimum assembly of particular type of component
which achieves a desired function usually requires examination of a
number of possibilities exponential in the number of components in the
solution. With fixed resources this implies a design time rising
exponentially with complexity. Alternatively the resources can be
used in stages, to design subassemblies, which are then combined into
larger units, and so on, until the desired result is achieved. This
can be much faster since the effort rises exponentially with the
incremental size of each stage and linearly with the number of stages,
with an additional small term, for overall planning, exponential in
the number of stages. The resulting construct will probably use more
of the basic component and be less efficient than an optimal design.

Biological evolution is affected by these considerations as
much as our technology. If a device is so difficult to design that our
technology cannot build it, then neither should we expect to find it
in the biological world. Conversely, if we find some naturally evolved
thing, we can rest assured that designing an equally good one one is
not an impossibly difficult task. Presumably there is a way of using
the physics of the universe to construct entities functionally
equivalent to human beings, but vastly smaller and more efficient.
Terrestrial evolution has not had the time or space to develop such
things. But by building within the sequence atoms, amino acids,
proteins, cells, organs, animal (often concurrently), it produced a
technological civilization out of inanimate matter in only two billion
years.

Harangue:

The existence of several examples of intelligence designed
under these constraints should give us great confidence that we can
achieve the same in a time span similar to that of other technological
accomplishments.

The situation is analogous to the history of heavier than air
flight, where birds, bats and insects clearly demonstrated the
possibility before our culture mastered it. Flight without adequate
power to weight ratio is heartbreakingly difficult (vis. Langley's
steam powered aircraft or current attempts at man powered flight),
whereas with enough power (on good authority!) a shingle will fly.
Refinement of the aerodynamics of lift and turbulence is most
effectively tackled after some experience with suboptimal aeroplanes.
After the initial successes our culture was able to far surpass
biological flight in a few decades.

Although there are known brute force solutions to most AI
problems, current machinery makes their implementation impractical.
Instead we are forced to expend our human resources trying to find
computationally less intensive answers, even where there is no
evidence that they exist. This is honorable scientific endeavor, but,
like trying to design optimal airplanes from first principles, a slow
way to get the job done.

With more processing power, competing presently impractical
schemes could be compared by experiment, with the outcomes often
suggesting incremental or revolutionary improvements. Computationally
expensive highly optimizing compilers would permit efficient code
generation at less human cost. The expanded abilities of existing
systems such as Mathlab, the symbolic mathematics system from MIT,
which can be used as a desk calculator for doing algebra and
trigonometry as well as arithmetic, along with new experimental
results, would accelerate theoretical development. Gains made this way
would improve the very systems being used, causing more speedup. The
intermediate results would be inefficient kludges busily contributing
to their own improvement. The end result is systems as efficient and
clever as any designed by more theoretical approaches, but sooner,
because more of the labor has been done by machines.

With enough power anything will fly. The next section examines
how much is needed.

Section 2: Measuring Processing Power

During the past ten years Digital Equipment Corporation's
PDP-10 has become the standard computer for AI and related research,
partly because it was designed with advanced techniques, such as time
sharing and unusual computer languages, in mind.

When first introduced, the PDP-10 was considered a large
machine. By today's standards it is medium size. The PDP-10 dealt
with in this section is the KA model, the standard until very
recently. The very largest scientific computers, heavily used in
physics, chemistry and other fields, made by companies such as Control
Data Corp. and IBM, are about 100 times the speed of the KA. When it
was new a KA system cost about half a million dollars. Large
computers sell for around 10 million.

Low level vision:

The visual system of a few animals has been studied in some
detail, especially the layers of the optic nerve near the retina. The
neurons comprising these structures are used efficiently to compute
local operations like high pass filtering and edge, curvature,
orientation and motion detection.

Assuming the visual cortex (and possibly the optic nerve
itself) is as computationally intensive as the retina, successive
layers producing increasingly abstracted representations, we can
estimate the total capability. There are a million separate fibers in
a cross section of the human optic nerve. The thickness of the optical
cortex is a thousand times the depth occupied by the neurons which
apply a single simple operation. The eye is capable of processing
images at the rate of ten per second (flicker at higher frequencies is
detected by special operators). This means that the human visual
system evaluates 10,000 million pixel simple operators each second.

A tightly hand coded simple operator, like high pass filtering
by subtraction of a local average, applied to a million pixel picture
takes at least 160 seconds when executed on a PDP-10, not counting
timesharing. Since the computer can evaluate only one at a time, the
effective rate is 1/160 million pixel simple operators per second.

Thus a hand coded PDP-10 falls short of being the equal of the
human visual system by a speed factor of 1.6 million.

It may not be necessary to apply every operator to every
portion of every picture, and a general purpose computer, being more
versatile than the optic nerve, can take advantage of this. I grant
an order of magnitude for this effect, reducing the optic nerve to a
mere 100,000 PDP-10 equivalents.

14| sperm whale *
10 |
| human *
13| chimp *
10 |
| human vision *
12|
10 |
|
11| proposed NASA wind +
10 | tunnel simulator
|
10|
10 |
|
9 | Cray +
10 | bee *
| CDC 7600, IBM 360/195 +
8 |
10 | KL-10 +
CP |
7 | KA-10 +
bit 10 |
--- |
sec 6 |
10 | slug *
|
5 |
10 | * sponge (alive)
|
4 |
10 |
|
3 |
10 | + pocket calculator
|
+---------------------------------------------------------------------------
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10

CE (bits)

FIGURE: Compute Power and Energy of various devices. Scales are logarithmic.
The Cray machine is an extremely fast and large scientific computer.
The NASA simulator would probably be a general purpose computer
100 times as powerful as the biggest existing machines. It has not
been designed yet.

The size of this factor is related to having chosen to
implement our algorithms in machine language. If we had opted to
disassemble a number of PDP-10's and reconfigure the components to do
the computation, far fewer (perhaps only one!) would have been
required. On the other hand if we had run our algorithms in
interpreted Lisp, 10 to 100 times as many would be needed. The
tradeoff is that the design time varies inversely with the execution
efficiency. A good Lisp program to compute a given function is much
easier to produce than an efficient machine language program, or an
equivalent piece of hardware.

As a practical example of the kind of problem this gap poses
in current research, consider my work. The task is to construct a
program which can drive a vehicle sensing the world with a TV camera
through terrain cluttered with obstacles, avoiding the obstacles and
getting to a desired place. The programs are written efficiently and
in the spirit of computing only as much as is actually required to
track objects from one image to the next, and to judge their distance
from the parallax caused by vehicle motion. In spite of this it takes
a large program several minutes of computing to process each frame.
Differences in performance caused by changes in the program can often
be determined only after tens of images have been processed, implying
a run time of hours. This greatly limits experimentation. Also, many
ideas on how to significantly improve performance cannot reasonably be
tried because they slow down the computation by another factor of 10
or 100, increasing typical runs to days and weeks! Many (such as
taking pictures at much smaller intervals than the current two foot
motions) require very little additional programming, and would be
almost certain to improve things.

Entropy measurement:

Is there a quantitative way in which the processing power of a
system, independent of its detailed nature, can be measured? A
feature of things which compute massively is that they change state in
complicated and unexpected ways. The reason for believing that, say,
a stationary rock does little computing is its high predictability.
By this criterion the amount of computing done by a device is in the
mind of the beholder. A machine with a digital display which flashed
1, 2, 3, 4 etc., at fixed intervals would seem highly predictable to
an adult of our culture, but might be justifiably considered to be
doing an interesting, nontrivial and informative computation by a
young child. Information theory provides a measure for this idea. If
a system is in a given state and can change to one of a number of next
states with equal probability, the information in the transition,
which I will call the Compute Energy (CE), is given by

CE = log2 N

where N is the number of next states. The measure is in binary
digits, bits. If we consider the system in the long run, considering
all the states it might ever eventually be in, then this measure
expresses the total potential variety of the system.

A machine which can accomplish a given thing faster is more
powerful than a slower one. A measure for Compute Power is obtained
by dividing the above sum by the time required for a transition.
Thus:
CP = log2 N / t

The units are bits/second.

Slightly more complicated formulas, which give lower values,
apply if the transitions probabilities and times are not all equal.

These measures are highly analogous to the energy and power
capacities of a battery. Some properties follow:

They are linear, i.e. the compute power and energy of a system of two
or more independent machines is the sum of the individual power and
energies;

Speeding up a machine by a factor of n increases the CP by the same
factor;

A completely predictable system has a CP and CE of zero;

A machine with a high short term CP, which can reach a moderate number
of states in a short time, can yet have a low CE, if the total number
of states attainable in the long run is not high.

A representative computer:

For the KA-PDP10, considering one instruction time, we have
(roughly) that in one microsecond this machine is able to execute one
of 2^5 different instructions, involving one of 2^4 accumulators and
one of 2^18 memory locations, most of these combinations resulting in
distinct next sates. This corresponds to a CP of

log2 (2^5 x 2^4 x 2^18) bit / 10^-6 sec = 27 x 10^6 bit/sec

This number is reduced by the fact that that different instruction
sequences can result in the same outcome, and increased slightly by
information flowing in from high speed storage devices connected to
the computer for a net of about 8.5x10^6 bit/sec (details in
[Moravec]).

The CP is also limited by the total compute energy. If we
ignore external devices, this is simply the total amount of memory,
about 36x2^18 = 9.4x10^6 bits. The PDP 10 could execute at its
maximum effectiveness for 9.4/8.5 = 1.1 seconds before reaching a
state which could have been arrived at more quickly another way. The
energy can be extended indefinitely, however, by addition of external
storage devices, such as disks and tapes.

Overall, the processing power of a typical major AI center
computer is at most 10^7 bits/sec. Time sharing reduces this to about
10^6 b/s per user. Programming in a moderately efficient high level
language costs another factor of 10, and running under an interpreter
may result in a per user power of a mere 10,000 bits/sec, if the
source code is efficient.

A typical nervous system:

We now consider the processing ability of animal nervous
systems, using humans as an example. Since the data is even more
scanty than what we assumed about the PDP-10, some not unassailable
assumptions need to be made. The first is that the processing power
resides in the neurons and their interconnections, and not in more
compact nucleic acid or other chemical encodings. There is no
currently widely accepted evidence for the latter, while neural
mechanisms for memory and learning are being slowly revealed. A
second is that the neurons are used reasonably efficiently, as
detailed analysis of small nervous systems and small parts of large
ones reveals (and common sense applied to evolution suggests).
Thirdly, that neurons are fairly simple, and their state can be
represented by a binary variable, "firing" or "not firing", which can
change about once per millisecond. Finally we assume that human
nervous systems contain about 40 billion neurons.

Considering the space of all possible interconnections of
these 40 billion (treating this as the search space available to
natural evolution in its unwitting attempt to produce intelligence, in
the same sense that the space of all possible programs is available to
someone trying to create intelligence in a computer), we note that
there is no particular reason why every neuron should not be able to
change state every millisecond. The number of combinations thus
reachable from a given state is 2^(40x10^9) the binary log of which
gives CE = 40x10^9. This leads to a compute power of

CP = 40 x 10^9 bit / 10^-3 sec = 40 x 10^12 bit/sec

which is about a million times the maximum power of the KA-10.

Keep in mind that much of this difference is due to the high
level of interpretation in the KA, compared to what we assumed for the
nervous system. Rewiring its gates or transistors for each new task
would greatly increase the CP, but also the programming time. If the
processor is made of 100,000 devices which can change state in 100 ns,
the potential CP available through reconfiguration is 10^5 bits/10^-7
sec = 10^12 b/s. The CE would be unaffected. If automatic design and
fabrication methods result in small quantity integrated circuit
manufacture becoming less expensive and more widely practiced, my
calculations may prove overly pessimistic.

Thermodynamic efficiency:

Thermodynamics and information theory provide us with a
minimum amount of energy per bit of information generated at a given
background temperature (the energy required to out shout the thermal
noise). This is approximately the Boltzmann constant,

1.38 x 10^-16 erg/deg variable = 0.96 x 10^-16 erg/deg bit

The reduction is due to the theoretical fact that a "variable", also
known as a degree of freedom, is worth log2 e bits, about 1.44
bits. This measure allows us to estimate the overall energy efficiency
of computing engines. For instance, we determined the computing power
of the brain, which operates at 300 degrees K, to be 40x10^12
bits/sec. This corresponds to a physical power of

40 x 10^12 bit/sec x 300 deg x 0.96 x 10^-16 erg/deg bit =
1.15 erg/sec = 1.15 x 10^-7 watt

The brain runs on approximately 40 watts, so we conclude that
it is 10^-8 times as efficient as the physical limits allow.

Doing the same calculation for the KA10, again at 300 deg, we
see that a CP of 8.5x10^6 bit/sec is worth 2.44x10^-14 watts. Since
this machine needs 10 kilowatts the efficiency is only 10^-18.
Conceivably a ten watt, but otherwise equivalent, KA10 could be
designed today, if care were taken to use the best logic for the
required speed in every assembly. The efficiency would then still be
only 10^-15.

As noted previously, there is a large cost inherent in the
organization of a general purpose computer. We might investigate the
computing efficiency of the logic gates of which it is constructed (as
was, in fact, done with the brain measure). A standard TTL gate can
change state in about 10ns, and consumes 10^-3 watt. The switching
speed corresponds to a CP of 10^8 bit/sec, or a physical power of
2.87x10^-13 watt. So the efficiency is 10^-10, only one hundred times
worse than a vertebrate neuron.

The newer semiconductor logic families are even better. C-MOS
is twice as efficient as TTL, and Integrated Injection Logic is 100
times better, putting it on a par with neurons.

Experimental superconducting Josephson junction logic operates
at 4 deg K, switches in 10^-11 sec, and uses 10^-7 watts per gate.
This implies a physical compute power of 3.5x10^-12 watt, and an
efficiency of 7x10^-5, 1000 times better than neurons. At room
temperature it requires a refrigerator that consumes 100 times as much
energy as the logic, to pump the waste heat uphill from 4 degrees to
300. Since the background temperature of the universe is about 4
degreees, this can probably eventually be done away with.

It is thus likely that there exist ways of interconnecting
gates made with known techniques which would result in behavior
effectively equivalent to that of human nervous systems. Using a
million I^2L gates, or 10 thousand Josephson junction gates, and a
trillion bits of slower bulk storage, all running at full speed, such
assemblies would consume as little as, or less than, the power needed
to operate a brain of the conventional type.

Past performance indicates that the amount of human and
electronic compute power available is inadequate to design such an
assembly within the next few years. The problem is much reduced if
the components used are suitable large subassemblies. Statements of
good high level computer languages are the most effective such
modularizations yet discovered, and are probably the quickest route to
human equivalence, if the necessary raw processing power can be
accessed through them. This section has indicated that a million
times the power of typical existing machines is required. The next
suggests this should be available at reasonable cost in about ten
years.

Section 3: The Growth of Processing Power

The references below present, among other things, the
following data points on a price curve:

Transistor price
.0001c .01c 1c $1 $100
+---+---+---+---+---+---+---+---+---+---+ Year
| |
+- O -+ 1950
| # | 1951 $100 transistor
| # | 1952 transistor hearing aid
| # | 1953
| # | 1954
+- # -+ 1955 transistor radios
| # | 1956
| O | 1957 $10 transistor
| # | 1958
| # | 1959
+- O -+ 1960 $1 transistor
| # | 1961
| # | 1962 $100,000 small computer (IBM 1620)
| # | 1963
| O | 1964
+- # -+ 1965 $0.08 transistor (IC)
| # | 1966 $1000 4 func calculator
| # | 1967 $6000 scientific calc.
| # | 1968 $10,000 small computer (PDP 8)
| # | 1969
+- O -+ 1970 $200 4 func calculator
| # | 1971
| # | 1972 1K RAMS (1 c/bit)
| # | 1973
| # | 1974 $1000 small computer (PDP 11)
+- O -+ 1975 4K RAMS (.1 c/bit)
| # | 1976 $5 4 func calc (.05 c/trans)
| | 1977
| | 1978
| | 1979
+---+---+---+---+---+---+---+---+---+---+

The numbers indicate a remarkably stable evolution. The price
per electronic switch has declined by a steady factor of ten every
five years, if speed and reliability gains are included. Occasionally
there is a more precipitous drop, when a price threshold which opens a
mass market is reached. This makes for high incentives, stiff price
competition and mass production economies. It happened in the early
sixties with transistor radios, and is going on now for pocket
calculators and digital wristwatches. It is begining for
microcomputers, as these are incorporated into consumer products such
as stoves, washing machines, televisions and sewing machines, and soon
cars. During such periods the price can plummet by a factor of 100 in
a five year period. Since the range of application for cheap
processors is larger than for radios and calculators, the explosion
will be more pronounced.

The pace of these gains is in no danger of slackening in the
forseeable future. In the next decade the current period may seem to
be merely the flat portion of an exponential rise. On the immediate
horizon are the new semiconductor techniques, I^2L, and super fast
D-MOS, CCD for large sensors and fast bulk memory, and magnetic
bubbles for mass storage. The new 16K RAM designs use a folded
(thicker) cell structure to reduce the area required per bit, which
can be interpreted as the first step towards 3 dimensional
integration, which could vastly increase the density of circuitry.
The use of V-MOS, an IC technique that vertically stacks the elements
of a MOS transistor is expanding. In the same direction, electron
beam and X-ray lithography will permit smaller circuit elements.

In the longer run we have ultra fast and efficient Josephson
junction logic, of which small IC's exist in an IBM lab, optical
communication techniques, currently being incorporated into
intermediate distance telephone links, and other things now just
gleams in the eye of some fledgling physicist or engineer.

My favorite fantasies include the "electronics" of super-dense
matter, either made of muonic atoms, where the electrons are replaced
by more massive negative particles or of atoms constructed of magnetic
monopoles which (if they exist) are very massive and affect each other
more strongly than electric charges. The electronics and chemistry of
such matter, where the "electron" orbitals are extremely close to the
nucleus, would be more energetic, and circuitry built of it should be
astronomically faster and smaller, and probably hotter. Mechanically
it should exhibit higher strength to weight ratios. The critical
superconducting transition field strengths and temperatures would be
higher. For monopoles there is the possibility of combination
magnetic electric circuitry which can contain, among many other
goodies, DC transformers, where an electric current induces a monopole
current at right angles to it, which in turn induces another electric
current. One might also imagine quantum DC transformers, matter
composed of a chainlike mesh of alternating orbiting electric and
magnetic charges.

I interpret these things to mean that the cost of computing
will fall by a factor of 100 during the next 5 years, as a consequence
of the processor explosion, and by the usual factor of 10 in the 5
years after that. As an approximation to what is available today,
note that in large quantities an LSI-11 sells for under $500. This
provides a moderately fast 16 bit processor with 4K of memory.
Another $500 could buy an additional 32K of memory, if we bought in
quantity. The result would be a respectable machine, somewhat less
powerful than the KA-10, for $1000. At the crude level of
approximation employed in the previous section, a million machines of
this type should permit human equivalence. A million dollars would
provide a thousand of them today (a much better buy, in terms of raw
processing power, than a million dollar large processor). In ten
years a million dollars should provide the equivalent of a million
such machines, in the form of a smaller number of faster processors,
putting human equivalence within reach.

A roomful of isolated small computers is unlikely to prove
very useful for our purposes. The next section suggests how to make
them work together.

Section 4: Mega Processing

The following discussion is based on an interconnection system
for computers described in a more technical version of this essay
[Moravec], based on Batcher sorting nets, which has approximately the
following properties:

Every processor may send a fairly long message to any other processor
about every quarter of a microsecond. The messages from all the
processors are emitted in synchronized waves. A wave takes one
microsecond to filter through the interconnection net, causing there
to be four waves in the net at one time.

Each message includes a priority number introduced by the sending
computer.

The network delivers to each processor the message with the highest
priority addressed to it, if any. The processor sending each
delivered message receives an acknowledgement, the processors whose
messages were blocked by higher priority ones receive notices of
failure.

The amount of network logic per processor is small, and grows as the
square of the log of the number of processors. This low growth rate
ensures that even in a system of a million processors the cost of the
interconnection is no greater than the cost of the processors.

A major feature of this scheme is its flexibility. It can
function as any of the fixed interconnection patterns of current
experimental multiprocessors, or as a hexagonal mesh, or a 7
dimensional cubic lattice, should that be desired, or the tree
organization being considered in a Stanford proposal. It can simulate
programmed pipeline machines, where numbers stream between units that
combine and transform them. What is more, it can do all of these
things simultaneously, since messages within one isolated subset of
the processors have no effect on messages in a disjoint subset. This
permits a very convenient kind of "time" sharing, where individual
users get and return processors as their demands change.

Such mimicry fails to take advantage of the ability to
reconfigure the interconnection totally every message wave. There are
many applications, such as searching a tree of possibilities in
reasoning or game playing where this could be used very effectively.
Several existing programming languages can be extended to make this
capability conveniently available to programmers.

Conventional programming languages consist of strings of
commands and conditional commands to be obeyed by the computer. This
type of programming can be extended to make reasonably convenient use
of a parallel computer by providing means by which the programmer can
specify that several strings of such commands can be carried out
concurrently, and by providing large data objects such as arrays which
are manipulated by operations that work on all the elements of the
objects simultaneously. The high bandwidth of the communications net
is required to transmit data manipulation commands to multiple parts
of large structures (by a chain letter technique), and to pass program
segments from processor to processor. [Moravec] contains many more
details, and also suggests what may be a more elegant solution.

We will probably want the first versions of such a system to
be able to serve several independent users simultaneously. The
system's resources would be managed by the system monitor, a program
running on several machines which maintains a pool of free processors,
and parcels them out on request, and which also handles file system
requests (bulk storage would be connected to a handful of the
processors), and allocation of other devices.

Processes belonging to a single user will be initiated by a
particular master machine, probably the one connected to his console.
This master can create a tree of subprocesses, possibly
intercommunicating, running on different machines. It should be
possible, for example, to do vision by having one subset configured as
an array processor for efficient implementation of retina-like
processing, while another is running an Algol/APL for the less
structured analytic geometry needed to interpret the image, and yet a
third is operating a Lisp system doing abstract reasoning about the
scene. Many existing systems permit this kind of organization, but
they are hampered by having an absurdly small amount of computing
power.

How is a system of this kind initialized, and how does one
abort an out of control process taking place in part of it without
affecting the rest? A possibility is to have an "executive" class of
messages (perhaps signalled by a particular bit in the data portion),
which user jobs are not permitted to emit. Reception of such messages
might cause resetting of the processor, loading of memory locations
within it, and starting execution at a requested locations. A single
externally controllable machine can be used to get things going,
fairly quickly if it emits a self replicating chain letter.

Now consider reliability. The system can obviously tolerate
any reasonable number of inoperable processors, by simply declaring
them unavailable for use. Failures in the communication net are much
more serious, and under most situations will require the system to
stop operating normally. It is possible to write diagnostic programs
which can track down defective comparator elements or broken data
wires. If something should happen to the clock signals to a given
level it would be necessary to wheel out an oscilloscope. If
reliability were a critical issue it would be possible to include a
duplicate net, to run things the while other was being debugged.

Section 5: The Future

Suppose my projections are correct, and the hardware
requirements for human equivalence are available in 10 years for about
the current price of a medium large computer. Suppose further that
software development keeps pace (and it should be increasingly easy,
because big computers are great programming aids), and machines able
to think as well as humans begin to appear in 10 years. If the cost
of electronics continues to plummet beyond then (and the existence of
increasingly cheaper and better robot labor, in addition to scientific
and engineering improvements, should ensure that), an additional 15
years should bring human equivalence into the pocket calculator price
range. I also assume that sensors and effectors for such devices will
be able to match human performance, since even today's technology is
able to supercede it in many areas. What then?

Well, even if these machines are only as clever as human
beings, they will have enormous advantages over humans in competitive
situations. Their production and upkeep is vastly less expensive, so
more of them can be put to work with given resources. They can be
easily specialized for given tasks, and be programmed to work
tirelessly. Because we are not constrained to use any particular type
of component in building them, versions can be designed to work
efficiently in environments in which sustaining humans is very
expensive, such as deep in the oceans, and more importantly in
boundless outer space. Most significantly of all, they can be put to
work as programmers and engineers, with the task of optimizing the
software and hardware which make them what they are. The successive
generations of machines produced this way will be increasingly smarter
and more cost effective. Of course, there is no reason to assume that
human equivalence represents any sort of upper bound. When pocket
calculators can out-think humans, what will a really big computer be
like? Regardless of how benevolent these machines are made, homo
sapiens will simply be outclassed.

Societies and economies are as surely subject to evolutionary
pressures as biological organisms. Failing social systems eventually
wither and die, and are replaced by more successful competitors, and
those that can sustain the most rapid expansion dominate sooner or
later.

I expect the human race to expand into space in the near
future, and O'Neill's habitats for people will be part of this. But
as soon as machines are able to match human performance, the economics
against human colonies become very persuasive. Just as it was much
cheaper to send Pioneer to Jupiter and Viking to Mars than men to the
Moon, so it will be cheaper to build orbiting power stations with
robot rather than human labor. A machine can be designed to live in
free space and love it, drinking in unattenuated sunlight and
tolerating hard radiation. And instead of expensive pressurized,
gravitied, decorated human colonies, the machines could be put to work
converting lunar material into orbiting automatic factories. The
doubling time for a machine society of this type would be much shorter
than for human habitats, and the productive capability would expand
correspondingly faster.

The first societies in space will be composed of co-operating
humans and machines, but as the capabilities of the self-improving
machine component grow, the human portion will function more and more
as a parasitic drag. Communities with a higher ratio of machines to
people will be able to expand faster, and will become the bulk of the
intelligent activity in the solar system. In the long run the sheer
physical inability of humans to keep up with these rapidly evolving
progeny of our minds will ensure that the ratio of people to machines
approaches zero, and that a direct descendant of our culture, but not
our genes, inherits the universe.

This may not be as bad as it sounds, since the machine society
can, and for its own benefit probably should, take along with it
everything we consider important, up to and including the information
in our minds and genes. Real live human beings, and a whole human
community, could then be reconstituted if an appropriate circumstance
ever arose. Since biology has committed us to personal death anyway,
with whatever immortality we can hope for residing only in our
children and our culture, shouldn't we be happy to see that culture
become as capable as possible? In fact, attempting to hobble its
growth is an almost certain recipe for long term suicide. The universe
is one random event after another. Sooner or later an unstoppable
virus deadly to humans will evolve, or a major asteroid will collide
with the earth, or the sun will go nova, or we will be invaded from
the stars by a culture that didn't try to slow down its own evolution,
or any number of other things. The bigger, more diverse and competent
our offspring are, the more capable they will be of detecting and
dealing with the problems that arise.

For the egomaniacs among us there is another possibility. The
main problem in keeping up with the machines is that we evolve by the
old DNA + nucleated cell + sex + personal death method, while our
machines evolve by the new improved intelligence + language + culture
+ science + technology technique, which is so very much faster that
our biology seems to stand still in comparison. If we could somehow
transfer our evolution to the faster form, we should be able hold our
own.

At first thought genetic engineering might seem to be the key.
Successive generations of human beings could be designed by
engineering mathematics and on the basis of computer simulations just
like airplanes and computers are now. But this is just like building
robots out of proteins instead of metal and plastic. Being made of
protein is in fact a major drawback. That stuff is stable only in a
narrow temperature and pressure range, sensitive to all sorts of high
energy disturbances, and so on, and rules out many construction
techniques and components. Is there some way to retain our essential
humanness, at least temporarily until we think of something better,
while transferring ourselves to a more malleable form?

Imagine the following process (meant to suggest a variety of
ways such a thing could be done). You are in an operating theater,
and a brain surgeon (probably a machine) is in attendance. On a table
next to yours is a potentially human equivalent computer, dormant now
for lack of a program to run. Your skull, but not your brain, is
under the influence of a local anaesthetic. You are fully conscious.
Your brain case is opened, and the surgeon peers inside. Its
attention is directed at a small clump of about 100 neurons somewhere
near the surface. It examines, non-destructively, the three
dimensional structure and chemical makeup of that clump with neutron
tomography, phased array radio encephalography, etc., and derives all
the relevant parameters. It then writes a program which can simulate
the behavior of the clump as a whole, and starts it running on a small
portion of the computer next to you. It then carefully runs very fine
wires from the computer to the edges of the neuron assembly, to
provide the simulation with the same inputs the neurons are getting.
You and it both check out the accuracy of the simulation. After you
are satisfied, it carefully inserts tiny relays between the edges of
the clump and the rest of the brain, and runs another set of wires
from the relays to the computer. Initially these simply transmit the
clump's signals through to the brain, but on command they can connect
the simulation instead. A button which activates the relays when
pressed is placed in your hand. You press it, release it and press it
again. There should be no difference. As soon as you are satisfied,
the simulation connection is established firmly, and the now
unconnected clump of neurons is removed.

The process is repeated over and over for adjoining clumps,
until the entire brain has been dealt with. Occasionally several
clump simulations are combined into a single equivalent but more
efficient program. Though you have not lost consciousness, or even
your train of thought, your mind has been removed from the brain and
transferred to the machine. A final step is the disconnection of the
your old sensory and motor system, to be replaced by higher quality
ones in your new home. This last part is no different than the
installation of functioning artificial arms, legs, pacemakers,
kidneys, ears and hearts and eyes being done or contemplated now.

Advantages become apparent as soon as the process is complete.
Somewhere in your machine is a control labelled "speed". It was
initially set to "slow", to enable the simulations to remain
synchronized with the rest of your old brain, but now the setting is
changed to "fast". You can communicate, react and think at a thousand
times your former rate. But this is only a minor first step.

Major possibilities stem from the fact that the machine has a
port which enables the changing program that is you to be read out,
non-destructively, and also permits new portions of program to be read
in. This allows you to conveniently examine, modify, improve and
extend yourself in ways currently completely out of the question. Or,
your entire program can be copied into a similar machine, resulting in
two thinking, feeling versions of you. Or a thousand, if you want. And
your mind can be moved to computers better suited for given
environments, or simply technologically improved, far more
conveniently than the difficult first transfer. The program can also
be copied to a dormant information storage medium, such as magnetic
tape. In case the machine you inhabit is fatally clobbered, a copy of
this kind can be read into an unprogrammed computer, resulting in
another you, minus the memories accumulated since the copy was made.
By making frequent copies, the concept of personal death could be made
virtually meaningless. Another plus is that since the essence of you
is an information packet, it can be sent over information
channels. Your program can be read out, radioed to the moon, say, and
infused there into a waiting computer. This is travel at the speed of
light. The copy that is left behind could be shut down until the trip
is over, at which time the program representing you with lunar
experiences is radioed back, and transfused into the old body. But
what if the original were not shut down during the trip? There would
then be two separate versions of you, with different memories for the
trip interval.

When the organization of the programs making up humans is
adequately understood, it should become possible to merge two sets of
memories. To avoid confusion, they would be carefully labelled as to
which had happened where, just as our current memories are usually
labelled with the time of the events they record. This technique
opens another vast realm of possibilities. Merging should be possible
not only between two versions of the same individual but also between
different persons. And there is no particular reason why mergings
cannot be selective, involving some of the other person's memories,
and not others. This is a very superior form of communication, in
which memories, skills, attitudes and personalities can be rapidly and
effectively shared.

The amount of memory storage an individual will typically
carry will certainly be greater than humans make do with today, but
the growth of knowledge will insure the impracticability of everybody
lugging around all the world's knowledge. This implies that
individuals will have to pick and choose what their minds contain at
any one time. There will often be knowledge and skills available from
others superior to a person's own. The incentive to substitute those
talents for native ones will be overwhelming most of the time. This
will result in a gradual erosion of individuality, and formation of an
incredibly potent community mind.

A pleasant possibility presents itself. Why should the mind
transferral process be limited to human beings? Earthly life contains
several species with brains as large as or larger than man's, from
dolphins, our cephalic equals, to elephants and the large whales, and
perhaps giant squid, with brains up to twenty times as big. If the
technical problem of translation can be overcome, and it may be quite
difficult for squid, in particular, since their minds are evolved
entirely independently, then our culture could be fused with theirs,
with each component used according to its value. In fact, a synthesis
of all terrestrial life is desirable with the simpler organisms
contributing only the information in their DNA, if that's all they
have. In this way all the knowledge generated by terrestrial
biological and cultural evolution will be retained in the data banks,
available whenever needed. This is a far more secure form of storage
than the present one, where genes and ideas are lost as species become
extinct and individuals die.

We now have a picture of a super-consciousness, the synthesis
of terrestrial life, and perhaps jovian and martian life as well,
constantly improving and extending itself, spreading outwards from the
solar system, converting non-life into mind. There may be other such
bubbles expanding from elsewhere. What happens when we meet another?
Well, it's presumptuous of me to say at this tender stage of the
evolution, but fusion of us with them is certainly a possibility,
requiring only a translation scheme between the data representations.
This process, possibly occuring now elsewhere, might convert the
entire universe into an extended thinking entity.

References

Section 1: The Natural History of Intelligence

[animal]
JERISON, Harry J., RIOPELLE, A.J., ed.
"Paleoneurology and the Evolution of Mind", "Animal Problem Solving",
Scientific American, Vol. 234, No. 1, Penguin Books, 1967.
January 1976, 90-101.
GOODRICH, Edwin S.,
BITTERMAN, M. E., "Studies on the Structure and Development of Vertebrates",
"The Evolution of Intelligence", Dover Publications Inc., New York, 1958.
Scientific American, Vol. 212, No. 1,
January 1965, 92-100. BUCHSBAUM, Ralph,
"Animals without Backbones",
GRIFFIN, Donald R., ed. The University of Chicago Press, 1948.
"Animal Engineering",
W.H. Freeman and Company, FARAGO, Peter, and Lagnado, John,
San Francisco, June 1974. "Life in Action"
Alfred A. Knopf, New York, 1972.
BURIAN, Z. and Spinar, Z.V.,
"Life Before Man", BONNER, John Tyler,
American Heritage Press, 1972. "Cells and Societies",
Princeton University Press, Princeton, 1955.

[squid]
COUSTEAU, Jacques-Yves and Diole, Philippe, BOYCOTT, Brian B.,
"Octopus and Squid", "Learning in the Octopus",
Doubleday & Company, Garden City, N.Y., 1973. Scientific American, Vol. 212, No. 3,
(also a televised film of the same name) March 1965, 42-50.

"The Octopus", LANE, Frank W.,
a televised film, Time-Life films. "The Kingdom of the Octopus",
Worlds of Science Book, Pyramid Publications Inc.
October, 1962.
[bird]
BAKKER, Robert T., STETTNER, Laurence Jay and Matyniak, Kenneth A.
"Dinosaur Renaissance", "The Brain of Birds",
Scientific American, Vol. 232, No. 4, Scientific American, Vol. 218, No. 6,
April 1975. June 1968, 64-76.

[whale]
LILLY, John. C., STENUIT, Robert,
"The Mind of the Dolphin" & "The Dolphin, Cousin to Man",
"Man and Dolphin", Bantam Books, New York, 1972.
Doubleday and Company, New York, 1967.

COUSTEAU, Jacques-Yves and Diole, Philippe, "Whales and Dolphins",
"The Whale", A BBC produced film shown
Doubleday & Company, Garden City, N.Y., 1972. in the NOVA television series

FICHTELIUS, Karl-Erik and Sjolander, Sverre, McINTYRE, Joan, ed.
"Smarter than Man?", "Mind in the Waters",
Ballantine Books, New York, 1974. Charles Scribner's Sons, San Francisco, 1974.

[elephant]
RENSCH, Bernhard,
"The Intelligence of Elephants",
Scientific American,
February 1957, 44.

[primate]
"The First Words of Washoe", PFEIFFER, John,
Televised film shown in the NOVA series "The Human Brain",
Worlds of Science Books, Pyramid Publications Inc.,
LeGROS CLARK, W.E., New York, 1962.
"History of the Primates",
The University of Chicago Press, Chicago 1966.

Section 2: Measuring Processing Power

WILLOWS, A.O.D., HUBEL, David H.,
"Giant Brain Cells in Mollusks", "The Visual Cortex of the Brain",
Scientific American, Vol. 224, No. 2, Scientific American,
February 1971, 69-75. November 1963, 54-62.

KANDEL, Eric R., WILLIAMS, Peter L., Warwick, Roger
"Nerve Cells and Behavior", "Functional Neuroanatomy of Man",
Scientific American, Vol. 223, No. 1, W.B. Saunders Company, Philadelphia, 1975.
July 1970, 57-70.
"PDP-10 Reference Handbook",
AGRANOFF, Bernard W., Digital Equipment Corporation, Maynard Mass., 1971.
"Memory and Protein Synthesis",
Scientific American, Vol. 216, No. 6, TRIBUS, Myron and McIrvine, Edward C.,
June 1967, 115-122. "Energy and Information",
Scientific American, Vol. 224, No. 3,
KENNEDY, Donald, September 1971, 179-188.
"Small Systems of Nerve Cells",
Scientific American, Vol. 216, No. 5, GLASSTONE, Samuel, Lewis, David,
May 1967, 44-52. "Elements of Physical Chemistry",
D. Van Nostrand Co. Inc., New York, 1960.
BAKER, Peter F.,
"The Nerve Axon", MILLER, Richard T.,
Scientific American, Vol. 214, No. 3, "Super Switch", 284, in Science Year annual 1975,
March 1966, 74-82. Field Enterprises Educational Corp., 1975.

LANDAUER, Rolf,
"Fundamental Limitations in the Computational
Process", IBM, Yorktown Heights N.Y. 1976

Section 3: The Growth of Processing Power

McWHORTER, Eugene W. TIEN, P. K.
"The Small Electronic Calculator", "Integrated Optics",
Scientific American, Vol. 234, No. 3, Scientific American, Vol. 230, No. 4,
March 1976, 88-98. April 1974, 28-35.

HODGES, David A., HITTINGER, William C.
"Trends in Computer Hardware Technology", Metal-Oxide-Semiconductor Technology",
Computer Design, Vol. 15, No. 2, Scientific American, Vol. 229, No. 2,
February 1976, 77-85. August 1973, 48-57.

SCRUPSKI, Stephen E., et al., BOBECK, Andrew H. and Scovil, H. E. D.
"Technology Update", "Magnetic Bubbles",
Electronics, Vol. 48, No. 21, McGraw-Hill, Scientific American, Vol. 224, No. 6,
October 16, 1975, 74-127. June 1971, 78-90.

VACROUX, Andre G. HEATH, F. G.
"Microcomputers", "Large-Scale Integration in Electronics",
Scientific American, Vol. 232, No. 5, Scientific American, Vol. 222, No. 3,
May 1975, 32-40. February 1970, 22-31.

TURN, Rien RAJCHMAN, Jan A.
"Computers in the 1980's", "Integrated Computer Memories",
Columbia University Press, Scientific American, Vol. 217, No. 1,
Rand Corporation, 1974. July 1967, 18-31.

HITTINGER, William C. and Sparks, Morgan
"Microelectronics",
Scientific American, Vol. 213, No. 5,
November 1965, 56-70.

Section 4: Mega Processing

BATCHER, K.E. KNUTH, D.E.
"Sorting Networks and their Applications", "Sorting and Searching",
1968 Spring Joint Computer Conf. Proceedings The Art of Computer Programming, Vol. 3
April 1968, 307-314. Addison-Wesley, 1973.

VAN VOORHIS, David C.
"An Economical Construction for Sorting Networks",
1974 National Computer Conference Proceedings
April 1974, 921-927.

General Reference

MORAVEC, Hans P.
"The Role of Raw Power in Intelligence",
Stanford AI Memo (to be published) available from the author