Human Culture - A Genetic Takeover Underway


 Hans Moravec
 Robotics Institute
 Carnegie-Mellon University
 Pittsburgh, PA  15213

 July 1988


	This is the end. Our genes, engaged for four billion years in
a relentless, spiralling arms race with one other, have finally
outsmarted themselves. They've produced a weapon so powerful it will
vanquish the losers and winners alike. I do not mean nuclear devices
-- {\it their} widespread use would merely delay the immensely more
interesting demise that's been engineered.

	You may be surprised to encounter an author who cheerfully
concludes the human race is in its last century, and goes on to
suggest how to help the process along. Surely, though, the surprise is
more in the timing than in the fact itself. The evolution of species
is a firmly established idea, and the accelerating pace of cultural
change has been a daily reality for a century. During those hundred
years many projections of future life, serious and fictional, have
been published. Most past futurism has kept separate the changes
anticipated in the external world, and those expected in our bodies
and minds. While our environment and our machinery could be rapidly
engineered through industrious invention, alterations in ourselves
were paced by the much slower Darwinian processes of mutation and
selection.

	In the late twentieth century the barriers of complexity that
divided the engineers of inanimate matter from the breeders of living
things have been crumbling. In the future presented in this book the
human race itself is swept away by the tide of cultural change, not to
oblivion, but to a future that, from our vantage point, is best
described by the word ``supernatural.'' Though the ultimate
consequences are unimaginable, the process itself is quite palpable,
and many of the intermediate steps are predictable. This book reflects
that progression--from uncontroversial history of relevant
technologies, to modest near-term projections, to speculative glimpses
of the distant future (discerning the fuzzy boundaries between them is
up to the reader). The underlying theme is the maturation of our
machines from the simple devices they still are, to entities as
complex as ourselves, to something transcending everything we know, in
whom we can take pride when they refer to themselves as our
descendants.

	As humans, we are half-breeds: part nature, part nurture. The
cultural half is built, and depends for its existence on the
biological foundation. But there is a tension between the two. Often
expressed as the drag of the flesh on the spirit, the problem is that
cultural development proceeds much faster than biological
evolution. Many of our fleshly traits are out of step with the
inventions of our minds. Yet machines, as purely cultural entities, do
not share this dilemma of the human condition.  Unfettered, they are
visibly overtaking us. Sooner or later they will be able to manage
their own design and construction, freeing them from the last vestiges
of their biological scaffolding, the society of flesh and blood humans
that gave them them birth.  There may be ways for human minds to share
in this emancipation.

	Free of the arbitrary limits of our biological evolution, the
children of our minds will yet be constrained by the physics and the
logic of the universe. Present knowledge hints at the motives, shapes
and effects of post-biological life, but there may be ways to
transcend even the most apparently fundamental barriers.

\section{The Slippery Slope to Genetic Takeover}

	The trouble began about a 100 million years ago when some gene
lines hit upon a way to make animals with the ability to learn
behaviors from their elders during life, rather than inheriting them
at conception.  It was accelerated 10 million years ago when our
ancestors began to rely on tools like bones and sticks and stones. It
was massively compounded with the coming of fire and complex
languages, perhaps 1 million years ago.  By the time our species
appeared, maybe 100 thousand years ago, the genes' job was done;
cultural evolution, the juggernaut that they had unwittingly
constructed, was rolling. Within the last ten thousand years human
culture produced the agricultural revolution and subsequently large
scale bureaucratic government, written language, taxes, and leisure
classes. In the last thousand years this change blossomed into a host
of inventions such as movable type printing that accelerated the
process. With the industrial revolution two hundred years ago, we
entered the final phase. Bit by bit, ever more rapidly, cultural
evolution discovered economically attractive artificial substitutes
for human body functions, as well as totally new abilities. One
hundred years ago we invented practical calculating machines that
could duplicate some small, but vexing, functions of the human mind.
Since then the mental power of calculating devices has risen a
thousandfold every twenty years.

	We are very near to the time when {\it no} essential human
function will lack an artificial counterpart. The embodiment of this
convergence of cultural developments is the intelligent robot, a
machine that can think and act as a human, however inhuman it may be
in physical or mental detail. Such machines could carry on our
cultural evolution, including their own increasingly rapid
self-improvement, without us, and without the genes that built us. It
will be then that our DNA will be out of a job, having passed the
torch, and lost the race, to a new kind of competition. The genetic
information carrier, in the new scheme of things, will be exclusively
knowledge, passed from mind to artificial mind.

	A. G. Cairns-Smith, a chemist contemplating the beginnings of
life on the early earth, calls this kind of internal coup a {\it
genetic takeover}. He suggests that it has happened at least once
before.  In Cairns-Smith's convincingly argued theory, presented most
accessibly in {\bf Seven Clues to Origin of Life}, the first organisms
were microscopic crystals of clay that reproduced by the common
processes of crystal growth and fracture, and carried genetic
information as patterns of crystal defects.  These defects influence
the physical properties of a clay, and its action as a chemical
catalyst, and so partially control that clay's immediate surroundings.
In a Darwinian process of reproduction, mutation and selection, some
crystal species stumbled on a way to harness nearby carbon compounds
as construction materials and machinery, and even as {\it external
repositories for genetic information}.  The carbon machinery was so
effective that organisms using it to ever greater extent won out,
resulting eventually in carbon based organisms with no vestiges of the
original crystalline genetics. Life as we know it had begun.

	How should you and I, products of both an organic and a
cultural heritage, feel about the coming rift between the two? We owe
our existence to organic evolution, but do we owe it any loyalty?  Our
minds and genes share many common goals during life, but even then
there is a tension between time and energy spent acquiring,
developing, and spreading ideas, and effort expended towards
biological reproduction (as any parent of teenagers will attest). As
death nears, the dichotomy widens; too many aspects of mental
existence simply cannot be passed on. The problem is partly one of
timescales: humans already live extraordinarily long compared to other
animals, no doubt to better teach their young, but the lifespan is a
compromise with the genes' evolutionary imperative to experiment, the
better to adapt. Things are a little askew because this deal was
forged long ago, when the cultural life was simpler.  The amount to
teach and learn has ballooned recently and, all other things being
equal, we'd likely be better off with a somewhat longer lifespan.  But
what would be the optimal lifespan if our genes' specialized needs
were no longer a factor?

	A sexually produced body is a finalized evolutionary
experiment.  Further genetic adaptation is precluded until offspring
are produced through a genetic bottleneck, and then the experiment is
over. A mind, however, is a conduit for ideas, and can evolve and
adapt without such abrupt beginnings and endings. In principle it
could cope successfully indefinitely. It is true that human minds,
tuned for mortality, undergo a maturation from impressionable
plasticity to self assured rigidity, and this makes them unpromising
material for immortality. But there are adaptable entities on earth
with indefinite life spans: living species and some human
institutions.  Their secret is a balance between continuity and
experimentation. Death of individual organisms plays a central role in
successful species. Old experiments are cleared away, making room for
new ones, in a genteel, prearranged way, or by relentless
life-and-death competitions.  In human institutions turnover in
skilled personnel and alteration of the company rules play the same
role. The point is that the larger unit, the species or the
organization, can adapt indefinitely (perhaps beyond recognition in
the long run) without losing its identity, as its design and
components are altered bit by bit.

	A thinking machine could probably be designed from the ground
up to have this same kind of flexibility.  Mental genes could be
created, imported, tested in combinations, and added and deleted to
keep the thinking current. The testing is of central importance: it
steers the evolution. If the machine makes too many bad decisions in
these tests, it will fail totally, in the old fashioned, Darwinian,
way.

	And so the world of the children of our minds will be as
different from our own as the world of living things is different from
the lifelessness than preceded it. The consequences of unfettered
thought are quite unimaginable.  We're going to try to imagine some of
them anyway.

\section{Machines Who Think (Weakly)}

	Later I will argue that robots with human intelligence will be
common within fifty years. By comparison, the best of today's machines
have minds more like those of insects. This in itself is a recent
gaint leap from far more modest beginnings. While mechanical
imitations of life have been with us for at least several hundred
years, the earliest machines, powered by running water, falling
weights, or springs copied the motions of living things, often
charmingly, but could not respond to the world around them.  They
could only {\it act}. The development of electrical, electronic and
radio technology early in this century made possible machines that
reacted to light, sound, and other subtle cues, and also provided a
means of invisible remote control.  These possibilities inspired a
number of entertaining demonstration robots, as well as thoughts and
stories about future human-like mechanisms, but only simple
connections between the sensors and motors were possible at
first. These machines could {\it sense} and {\it act}, but hardly
think.

	Analog computers were designed during World War II for
controlling anti-aircraft guns, for navigation, and for precision
bombing. Some of their developers noticed a similarity between the
operation of the devices and the regulatory systems in living things,
and these researchers were inspired to build machines that acted as if
they were alive. Norbert Wiener of MIT coined the term ``cybernetics''
for this unified study of control and communication in animals and
machines. Its practitioners combined new theory on feedback regulation
with post war electronics and early knowledge of living nervous
systems to build machines that responded like simple animals, and were
able to learn. The rudiments of {\it thought} had arrived.

	The field thrived less than two decades. Among its highlights
was a series of electronic turtles built during the 1950s by W. Grey
Walter, a British psychologist. With subminiature tube electronic
brains, and rotating phototube eyes, microphone ears and contact
switch feelers, the first versions could locate their ``recharging
hutch'' when their batteries ran low, and otherwise avoid trouble
while wandering about.  Groups of them exhibited complex social
behavior by responding to each other's control lights and touches. A
later machine with the same senses, could be conditioned to associate
one stimulus with another, and could learn, by repeated experience,
that, for instance, a loud noise would be followed by a kick to its
shell. Once educated, the turtle would avoid a noise as it had before
responded to a kick. The associations were slowly accumulated as
electrical charges in capacitors.

	The swan song of the cybernetics effort may have been the
Johns Hopkins University ``Beast.'' Built by a group of brain
researchers in the early 1960s, it wandered the halls, guided by sonar
and a specialized photocell eye that searched for the distinctive
black cover plate of wall outlets, where it would plug itself in, to
feed. It inspired a number of imitators. Some used special circuits
connected to TV cameras instead of photocells, and were controlled by
assemblies of (then new) transistor digital logic gates. Some added
new motions such as ``shake to untangle arm'' to the repertoire of
basic actions.

	Cybernetics was laid low by a relative. The war's many small
analog computers, which had inspired cybernetics, had a few, much
larger, digital cousins.  The first automatic digital computers, giant
autonomous calculators, were completed toward the end of the war and
used for codebreaking, calculating artillery tables, and atomic bomb
design.  Less belligerently, they provided unprecedented opportunities
for experiments in complexity, and raised the hope in some pioneers
like Alan Turing and John von Neumann that the ability to think
rationally, our most unique asset in dealing with the world, could be
captured in a machine. Our minds might be amplified just as our
muscles had been by the energy machines of the industrial
revolution. Programs to reason and to play intellectual games like
chess were designed, for instance by Claude Shannon and by Turing in
1950, but the earliest computers were too puny and too expensive for
this kind of use. A few poor checker playing programs did appear on
the first commercial machines in the early 1950s, and equally poor
chess programs showed up in latter half of that decade, along with a
better checker player.  In 1957 Allen Newell, Herbert Simon, and John
Shaw demonstrated the {\it Logic Theorist}, the first program able to
reason about arbitrary matters, by starting with axioms and applying
rules of inference to prove theorems.

	In 1960 John McCarthy coined the term ``Artificial
Intelligence'' for the effort to make computers think. By 1965 the
first students of McCarthy, Marvin Minsky, Newell, and Simon had
produced programs that proved theorems in geometry, solved problems
from intelligence tests, algebra books, and calculus exams, and they
played chess all with the proficiency of an average college
freshman. Each program could handle only one narrow problem type, but
for first efforts they were very encouraging-- so encouraging that
most involved felt that another decade of progress would surely
produce a genuinely intelligent machine. In later chapters I will
explain the nature of their understandable miscalculation.

	Now, thirty years later, computers are thousands of times as
powerful, but they don't seem much smarter. In the past three decades
progress in artificial intelligence has slowed from the heady sprint
of a handful of enthusiasts to the plodding trudge of growing throngs
of workers.  Even so, modest successes have maintained flickering
hope. So-called ``expert systems,'' programs encoding the decision
rules of human experts in narrow domains such as diagnosis of
infections, factory scheduling, or computer system configuration, are
earning their keep in the business world.  A fifteen-year effort at
MIT has gathered knowledge about algebra, trigonometry, calculus, and
related fields into a program called MACSYMA; this wonderful program
manipulates symbolic formulas and helps to solve otherwise forbidding
problems. Several chess playing programs are now officially rated as
chess masters, and excellent performance has been achieved in other
games like backgammon. Other semi-intelligent programs can understand
simplified typewritten English about restricted subjects, make
elementary deductions in the course of answering questions, and
interpret spoken commands chosen from thousand-word repertoires. Some
can do simple visual inspection tasks, such as deciding whether a part
is in its desired location.

	Unfortunately for humanlike robots, computers are at their
worst trying to do the things most natural to humans, like seeing,
hearing, manipulating, language, and common sense. This
dichotomy--machines doing well things humans find hard, while doing
poorly what's easy for us--is a giant clue to the nature of the
intelligent machine problem.

\section{Machines Who See (Dimly) and Act (Clumsily)}

	In the mid 1960s Minsky's students at MIT began to connect
television camera eyes and mechanical robot arms to their computers,
giving eyes and hands to computer minds, for machines that could see,
plan, and act.  By 1965 they had created programs that could find and
remove children's blocks, painted white, from a black tabletop. This
was a difficult and impressive accomplishment, requiring a controlling
program as complex as any of the then current pure reasoning
programs. Yet, while the reasoning programs, unencumbered by robot
appendages, matched college freshmen in fields like calculus, Minsky's
hand-eye system could be bested by a toddler. Nevertheless, hand-eye
experiments continued at MIT and elsewhere, gradually developing the
field which now goes by the name ``robotics,'' a term coined in
science fiction stories by Isaac Asimov. As with mainstream artificial
intelligence programs, robotics has progressed at an agonizingly slow
rate over the last twenty years.

	Not all robots, nor all people, idle away their lives in
universities.  Many must work for a living. Even before the industrial
revolution, before any kind of thought was mechanized, partially
automatic machinery, powered by wind or flowing water, was put to work
grinding grain and cutting lumber. The beginnings of the industrial
revolution in the eighteenth century were marked by the invention of a
plethora of devices that could substitute for manual labor in a
powerful, precise, and thoroughly inhuman way. Powered by turning
shafts driven by water or steam, these machines pumped, pounded, cut,
spun, wove, stamped, moved materials and parts and much else,
consistently and tirelessly. Once in a while something ingeniously
different appeared: the Jacquard loom, invented in 1801, could weave
intricate tapestries specified by a string of punched cards (a human
operator provided power and the routine motions of the weaving
shuttle). By the early twentieth century electronics had given the
machinery limited senses; it could now stop when something went wrong,
or control the temperature, thickness, even consistency, of its
workpieces.  Still, each machine did one job and one job only. This
meant that, as technical developments occurred with increasing
rapidity, the product produced by the machine often became obsolete
before the machine had paid back its design and construction costs, a
problem which had become particularly acute by the end of World War
II.

	In 1954 the inventor George Devol filed a patent for a new
kind of industrial machine, the programmable robot arm, whose
movements would be controlled by a stream of punched cards, and whose
task could thus be altered simply by changing its program cards. In
1958, with Joseph Engelberger, Devol founded a company named Unimation
(a contraction of "universal" and "automation") to build such
machines. The punched cards soon gave way to a magnetic memory,
thereby allowing the robot to be programmed simply by leading it by
the hand through its required paces once.  The first industrial robot
began work in a General Motors plant in 1961. To this day most large
robots seen welding, spray painting, and moving pieces of cars are
still of this type.

	Only when the cost of small computers dropped to less than
\$10,000 did robotics research conducted in universities begin to
influence the robot industry. The first industrial vision systems,
usually coupled with a new class of small robot arms, appeared in the
late 1970s, and now play a modest, but quietly booming, role in the
assembly and inspection of small devices like calculators, printed
circuit boards, and automobile water pumps. Indeed, industrial needs
have strongly influenced university research. What was once a
negligible number of smart robot projects has swelled to the
hundreds. And while cybernetics may be relatively dormant, its stodgy
parent, control theory, has grown massively since the war to meet the
profitable needs of the aerospace industry; moreover, the applications
developed for controlling air- and spacecraft and weapons are once
again finding their way into robots. The goal of humanlike
performance, though highly diluted by a myriad of approaches and short
term goals, has acquired a relentless, Darwinian, vigor.  As a story,
it becomes bewildering in its diversity and interrelatedness. Let's
move on to the sparser world of robots that rove.

\section{Machines Who Explore (Haltingly)}

	In the next section I will try to convince you that mobility
is a key to developing fully intelligent machines, an argument that
begins with the observation that {\it reasoning}, as such, is only the
thinnest veneer of human thought, effective only because it is
supported by much older and much more powerful and diverse unconscious
mental machinery. This opinion may have been common among the
cybernetics researchers, many of whose self-contained experiments were
animal-like and mobile.  It is not yet widespread in the artificial
intelligence research community, where experiments are typically
encumbered by huge, immobile mainframe computers, and dedicated to
mechanizing pure reasoning. Nevertheless, a small number of mobile
robots have appeared in the artificial intelligence laboratories.

	Stanford Research Institute's ``Shakey,'' was a mobile robot
built by the researchers who believed that reasoning was the essence
of intelligence, and in 1970 it was the first mobile robot to be
controlled by programs that reasoned.  Five feet tall, equipped with a
television camera, it was remote controlled by a large
computer. Inspired by the first wave of successes in AI research, its
designers sought to apply logic-based problem solving methods to a
real world task.  Controlling the movement of the robot, and
interpreting its sensory data, were treated as secondary tasks and
relegated to junior programmers.  MIT's ``blocks world'' vision
methods were used, and a robot environment was constructed in which
the robot moved through several rooms bounded by clean walls, seeing,
and sometimes pushing, large, uniformly painted blocks and wedges.
Shakey's most impressive performance, executed piecemeal over a period
of days, was to solve a so called ``monkey and bananas'' problem.
Told to push a particular block that happened to be resting on a
larger one, the robot constructed and acted on a plan that included
finding a wedge that could serve as a ramp, pushing it against the
large block, driving up the ramp, and delivering the requested push.

	The environment was contrived, and the problem staged, but it
provided a motivation, and a test, for a clever reasoning program
called STRIPS (the STanford Research Institute Problem Solver) that,
given a task for the robot, assembled a plan out of the little actions
the robot could take. Each little action had preconditions (e.g., to
push a block, it must be in front of us) and probable consequences
(e.g., after we push a block, it is moved). The state of the robot's
world was represented in sentences of mathematical logic, and
formulating a plan was like proving a theorem, with the initial state
of the world being the axioms, and primitive actions being the rules
of inference. One complication was immediately evident: the outcome of
a primitive action is not always what one expects (as, for instance,
when the block does not budge). Shakey had a limited ability to handle
such glitches by occasionally observing parts of the world, and
adjusting its internal description and replanning its actions if the
conditions were not as it had assumed.

	Shakey's specialty was {\it reasoning} - its rudimentary
vision and motion software worked only in starkly simple surroundings.
At about the same time, on a much lower budget, a mobile robot that
was to specialize in {\it seeing} and {\it moving} in natural settings
was born at Stanford University's Artificial Intelligence
Project. John McCarthy founded the Project in 1963 with the then
plausible goal of building a fully intelligent machine in a
decade. (The Project was renamed the Stanford AI Laboratory, or SAIL,
as the decade drew nigh and plausibility drifted away.)  Reflecting
the priorities of early AI research, McCarthy worked on reasoning, and
delegated to others the design of ears, eyes, and hands for the
anticipated artificial mind . SAIL's hand-eye group soon overtook the
MIT robotics group in visible results, and was seminal in the later
industrial smart robot explosion. A modest investment in mobility was
added when Les Earnest, SAIL's technically astute chief administrator,
learned of a vehicle abandoned by Stanford's mechanical engineering
department after a short stint as a simulated remote controlled lunar
rover.  At SAIL it became the Stanford Cart, the first mobile robot
controlled by a large computer that did {\it not} reason , and the
first testbed for computer vision in the cluttered, haphazardly
illuminated, world most animals inhabit. The progeny of two PhD
theses, it slowly navigated raw indoor and outdoor spaces guided by TV
images processed by programs quite different from those in the blocks
world.

	In the mid 1970s NASA began planning for a robot Mars mission
to follow the successful Viking landings. Scheduled for launch in
1984, it was to include two vehicles roving the Martian surface. Mars
is so far away, even by radio, that simple remote control was
unattractve; the delay between sending a command and seeing its
consequence could be as long as forty minutes. Much greater distances
would be possible if the robot could travel safely on its own much of
the time. Toward this end Caltech's Jet Propulsion Laboratory,
designer of most of NASA's robot spacecraft, which until then used
quite safe and simple automation, initiated an intelligent robotics
project. Pulling together methods, hardware, and people from
university robotics programs, it built a large wheeled test platform
called the Robotics Research Vehicle, or RRV, a contraption that
carried cameras, a laser rangefinder, a robot arm, and a full
electronics rack, all connected by a long cable to a big computer. By
1977 it could struggle through short stretches of rock-littered
parking lot to pick up a certain rock and rotate it for the
cameras. But in 1978 the project was halted when the Mars 1984 mission
was cancelled and removed from NASA's budget. (Of course, Mars hasn't
gone away, and the JPL is considering a visit there at the end of the
millenium.)

	The best supporter of artificial intelligence research is the
Department of Defense's Advanced Research Project Agency (DARPA).
Founded after the 1957 humiliation of Sputnik to fund far out projects
as insurance against future unwelcome technological surprises, it
became the world's first government agency to foster AI
investigations. In 1981 managers in DARPA decided that robot
navigation was sufficiently advanced to warrant a major effort to
develop autonomous vehicles able to travel large distances overland
without a human operator, perhaps into war zones or other hazardous
areas. The number of mobile robot projects jumped dizzyingly, in
universities and at defense contractors, as funding for this project
materialized. Even now, several new, truck-sized, robots are
negotiating test roads around the country--and the dust is still
settling.

	On a more workaday level, it is not a trivial matter that
fixed robot arms in factories must have their work delivered to
them. An assembly line conveyor belt is one solution, but managers of
increasingly automated factories in the late 1970s and early 1980s
found belts, whose routes are difficult to change, too
restrictive. Their robots could be rapidly reprogrammed for different
jobs, but the material flow routes could not.  Several large companies
worldwide dealt with the problem by building what they called
Automatically Guided Vehicles, AGVs, that navigated by sensing signals
transmitted by wires buried along their route. Looking like fork lifts
or large bumper cars, they can be programmed to travel from place to
place and be loaded and unloaded by robot arms. Some recent variants
carry their own robotic arms. Burying the route wires in concrete
factory floors is expensive, and alternative methods of navigation are
being sought. As with robot arms, the academic and industrial efforts
have merged, and a bewildering number of directions and ideas are
being energetically pursued.

	The history presented so far is highly sanitized, and
describes only a few major actors in the newly united field of
robotics. The reality is a turbulent witch's brew of approaches,
motivations, and, as yet, unconnected problems. The practitioners are
large and small groups around the world of electrical, mechanical,
optical, and all other kinds of engineers, physicists, mathematicians,
biologists, chemists, medical technologists, computer scientists,
artists, and inventors. Computer scientists and biologists are
collaborating on the development of machines that see.  Physicists and
mathematicians can be found improving sonar and other senses.
Mechanical engineers have built machines that walk on legs, and others
that grasp with robot hands of nearly human dexterity. These are all
fledgling efforts, and the ground rules are not yet worked out. Each
group represents a different set of backgrounds, desires, and skills;
communication among groups is often difficult. There are no good
general texts in the field, nor even a generally agreed upon
outline. Continuing diversity and rapid change make it likely that
this situation will continue for many years. In spite of the chaos,
however, I maintain that the first mass offering from the cauldron
will probably be served within a decade. And what leaps out of the
brew in fifty years is the subject of the rest of this book. Before
concluding this chapter, I'll foreshadow some of the contents in the
cauldron by returning to notions raised at the outset.

\section{Mobility and Intelligence}

	I've been hinting that robot research, especially the mobile
robot variety, has a significance much greater than the sum of its
many applications, and is, indeed, the safest route to full
intelligent machines.  I'll offer more detailed evidence later, but
briefly the argument goes like this.

       Computers were created to do arithmetic faster and better than
people. AI attempts to extend this superiority to other mental arenas.
Some mental activities require little data, but others depend on
voluminous knowledge of the world.  Robotics was pursued in AI labs
partly to automate the acquisition of world knowledge.  It was soon
noticed that the acquisition problem was less tractable than the
mental activities it was to serve.  While computers often exhibited
adult level performance in difficult mental tasks, robotic controllers
were incapable of matching even infantile perceptual skills.

       In hindsight the dichotomy is not surprising.  Animal genomes
have been engaged in a billion year arms race among themselves, with
survival often awarded to the quickest to produce a correct action
from inconclusive perceptions.  We are all prodigous olympians in
perceptual and motor areas, so good that we make the hard look easy.
Abstract thought, on the other hand, is a small new trick, perhaps
less than a hundred thousand years old, not yet mastered.  It just
looks hard when we do it.

	How hard and how easy?  Average humans beings can be beaten at
arithmetic by a one operation per second machine, in logic problems by
100 operations per second, at chess by 10,000 operations per second,
in some narrow "expert systems" areas by a million operations.
Robotic performance can not yet provide this same standard of
comparison, but a calculation based on retinal processes and their
computer visual equivalents suggests that a {\it billion} ($10^{9}$)
operations per second are required to do the job of the retina, and
$10$ {\it trillion} ($10^{13}$) to match the bulk of the human brain.

       Truly expert human performance may depend on mapping a problem
into structures originally constructed for perceptual and motor tasks
- so it can be internally visualized, felt, heard or perhaps smelled
and tasted.  Such transformations give the trillion operations per
second engine a purchase on the problem.  The same perceptual-motor
structures may also be the seat of ``common sense'', since they
probably contain a powerful model of the world - developed to solve
the merciless life and death problems of rapidly jumping to the right
conclusion from the slightest sensory clues.

	Decades of steady growth trends in computer power suggest that
trillion operation per second computers will be common in twenty to
forty years.  Can we expect to program them to mimic the ``hard''
parts of human thought in the same way that current AI program capture
some of the easy parts?  It is unlikely that introspection of
conscious thought can carry us very far - most of the brain is not
instrumented for introspection, the neurons are occupied efficiently
solving the problem at hand, as in the retina.  Neurobiologists are
providing some very helpful instrumentation extra-somatically, but not
fast enough for the forty year timetable.

	Another approach is to attempt to parallel the evolution of
animal nervous systems by seeking situations with selection criteria
like those in their history.  By solving similar incremental problems,
we may be driven, step by step, through the same solutions (helped,
where possible, by biological peeks at the ``back of the book'').
That animals started with small nervous systems gives confidence that
small computers can emulate the intermediate steps, and mobile robots
provide the natural external forms for recreating the evolutionary
tests we must pass.  Followers of this ``bottom up'' route to AI may
one day meet those pursuing the traditional ``top down'' route half
way.  Fully intelligent machines will result when the metaphorical
golden spike is driven uniting the two efforts.

\begin{figure}
\vspace{6in}
\caption{The evolution of terrestrial intelligence.}
\end{figure}

	The parallel between the evolution of intelligent living
organisms and the development of robots is a strong one. Many
real-world constraints that shaped life by favoring one kind of change
over another in the contest for survival also affect the viability of
robot characteristics.  To a large extent the incremental paths of
development pioneered by living things are being followed by their
technological imitators. Given this, there are lessons to be learned
from the diversity of life. One is the observation made earlier, that
mobile organisms tend to evolve the mental characteristics that form
the bedrock of human intelligence, immobile ones do not.  Plants are
an example of the latter case; vertebrates an example of the
former. An especially dramatic contrast is provided in an invertebrate
phylum, the molluscs. Many are shellfish like clams and oysters that
move little and have small nervous systems and behaviors more like
plants than like animals. Yet they have relatives, the cephalopods,
like octopus and squid, that are mobile and have independently
developed many of the characteristics of vertebrates, including
imaging eyes, large nervous systems and very interesting behavior,
including major problem solving abilities.

	 Two billion years ago our unicelled ancestors parted genetic
company with the plants. By dint of energetics and heritage, large
plants now live their lives fixed in place. Awesomely effective in
their own right, the plants have no apparent inclinations toward
intelligence--negative evidence that supports my thesis that mobility
is a parent of this trait. Animals bolster the argument on the
positive side, except for the immobile minority like sponges and clams
that support it on the negative.

	A billion years ago, before brains or eyes were invented, when
the most complicated animals were something like hydras (i.e., double
layers of cells with a primitive nerve net), our progenitors split
with invertebrates. Now both clans have "intelligent" members. Most
mollusks are sessile shellfish, but octopus and squid are highly
mobile, with big brains and excellent eyes.  Evolved independently of
us, they are quite different in detail. The optic nerve connects to
the back of the retina, so there is no blind spot. The brain is
annular, a ring around the esophagus. The green blood is circulated by
a systemic heart oxygenating the tissues and two gill hearts moving
depleted blood. Hemocyanin, a copper doped protein related to
hemoglobin and chlorophyll, carries the oxygen. Octopus and their
relatives are swimming light-shows, their surfaces covered by a
million individually controlled color changing cells. A cuttlefish
placed on a checkerboard can imitate the pattern, a fleeing octopus
can make deceiving seaweed shapes coruscate backward along its
body. Photophores of deep sea squid, some with irises and lenses,
generate bright multicolored light.  Since they also have good vision,
there is a potential for rich communication.

	Martin Moynihan, a biologist at the University of Indiana, in
{\bf Communication and Noncommunication by Cephalopods} identifies
several dozen distinct symbolic displays , many apparently expressing
strong emotions.  Their behavior is mammallike.  Octopus are reclusive
and shy; squid are occasionally aggressive. Small octopus can learn to
solve problems like how to open a container of food.  Giant squid,
with large nervous systems, have hardly ever been observed except as
corpses. They might be as clever as whales.

	 Birds are vertebrates, related to us through a 300 million
year old, probably not very bright, early reptile. Size-limited by the
dynamics of flying, some are intellectually comparable to the highest
mammals. The intuitive number sense of crows and ravens, for example,
extends to seven, compared to three or four for us. Birds outperform
all mammals except higher primates and the whales in ``learning set''
tasks, where the idea is to generalize from specific instances . In
mammals generalization depends on cerebral cortex size. In birds
forebrain regions called the Wulst and the hyperstriatum are critical
, while the cortex is small and unimportant.

	Our last common ancestor with the whales was a primitive
shrew-like mammal alive 100 million years ago. Some dolphin species
have body and brain masses identical to ours, and have had them for
more generations . They are as good as us at many kinds of problem
solving , and can grasp and communicate complex ideas.  Killer whales
have brains five times human size, and their ability to formulate
plans is better than the dolphins', whom they occasionally eat.  Sperm
whales, though not the largest animals, have the world's largest
brains.  Intelligence may be an important part of their struggle with
large squid, their main food. Elephant brains are three times human
size. Elephants form matriarchal tribal societies and exhibit complex
behavior. Indian domestic elephants learn over 500 commands, 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 do have long
memories. Apes are our 10 million year cousins.  Chimps and gorillas
can learn to use tools and to communicate in human sign languages at a
retarded level. Chimps have one third, and gorillas one half, human
brain size.

	Animals exhibiting near-human behavior have hundred-billion
neuron nervous systems. Imaging vision alone requires a billion. The
most developed insects have a million brain cells, while slugs and
worms make do with fewer than one hundred thousand, and sessile
animals with a few thousand.  The portions of nervous systems for
which tentative wiring diagrams have been obtained, including several
nerve clumps of the large neuroned sea slugs, and leeches, and the
early stages of vertebrate vision, reveal neurons configured into
efficient, clever, assemblies.

	The twenty year old modern robotics effort can hardly hope to
rival the billion year history of large life on earth in richness of
example or profundity of result. Nevertheless, the evolutionary
pressures that shaped life are already palpable in the robotics
labs. The following is a thought experiment that reflects this
situation. We wish to make robots execute general tasks such as ``go
down the hall to the third door, go in, look for a cup and bring it
back.'' This desire has created a pressing need--a computer language
in which to specify complex tasks for a rover, and a hardware and
software system to embody it. Sequential control languages
successfully used with industrial manipulators might seem a good
starting point. Paper attempts at defining the structures and
primitives required for the mobile application revealed that the
linear control structure of these state-of-the-art robot arm
controlling languages was inadequate for a rover. The essential
difference is that a rover, in its wanderings, is regularly
``surprised'' by events it cannot anticipate, but with which it must
deal. This requires that contingency routines be activated in
arbitrary order, and run concurrently, each with its own access to the
needed sensors, effectors, and internal state of the machine, and a
way of arbitrating their differences. As conditions change the
priority of the modules changes, and control may be passed from one to
another.

	Suppose that we ask a future robot to go down the hall to the
third door, go in, look for a cup and bring it back. This will be
implemented as a process that looks very much like a program written
for the arm control languages (that in turn look very much like Algol,
or Basic), except that the door recognizer routine would probably be
activated separately. Consider the following caricature of such a
program.

 module GO-FETCH-CUP
 wake up DOOR-RECOGNIZER with instructions
      ( on FINDING-DOOR add 1 to DOOR-NUMBER
    	 record DOOR-LOCATION )

 record START-LOCATION
 set DOOR-NUMBER to 0
 while DOOR-NUMBER<3 WALL-FOLLOW
 FACE-DOOR
 if DOOR-OPEN then GO-THROUGH-OPENING
              else OPEN-DOOR-AND-GO-THROUGH
 set CUP-LOCATION to result of LOOK-FOR-CUP
 TRAVEL to CUP-LOCATION
 PICKUP-CUP at CUP-LOCATION
 TRAVEL to DOOR-LOCATION
 FACE-DOOR
 if DOOR-OPEN then GO-THROUGH-OPENING
              else OPEN-DOOR-AND-GO-THROUGH
 TRAVEL to START-LOCATION
 end

	So far so good. We activate our program, and the robot
obediently begins to trundle down the hall counting doors. It
correctly recognizes the first one. The second door, unfortunately, is
decorated with garish posters, and the lighting in that part of the
corridor is poor, and our experimental door recognizer fails to detect
it. The wall follower, however, continues to operate properly and the
robot continues on down the hall, its door count short by one. It
recognizes door 3, the one we had asked it to go through, but thinks
it is only the second, so continues. The next door is recognized
correctly, and is open. The program, thinking it is the third one,
faces it and proceeds to go through. This fourth door, sadly, leads to
the stairwell, and the poor robot, unequipped to travel on stairs, is
in mortal danger. Fortunately there is a process in our concurrent
programming system called {\tt DETECT-CLIFF}. This program is always
running and checks ground position data posted on the blackboard by
the vision processes and also requests sonar and infrared proximity
checks on the ground. It combines these, perhaps with an a-priori
expectation of finding a cliff set high when operating in dangerous
areas, to produce a number that indicates the likelihood there is a
drop-off in the neighborhood. A companion process {\tt
DEAL-WITH-CLIFF}, also running continuously, but with low priority,
regularly checks this number and adjusts its own priority on the basis
of it. When the cliff probability variable becomes high enough, the
priority of {\tt DEAL-WITH-CLIFF} will exceed the priority of the
current process in control, {\tt GO-FETCH-CUP} in our example, and
{\tt DEAL-WITH-CLIFF} takes over control of the robot. A properly
written {\tt DEAL-WITH-CLIFF} will then proceed to stop or greatly
slow down the movement of the robot, to increase the frequency of
sensor measurements of the cliff, and to back away slowly from it when
it has been reliably identified and located.

	Now there's a curious thing about this sequence of actions. A
person seeing them, not knowing about the internal mechanisms of the
robot, might offer the interpretation ``First the robot was determined
to go through the door, but then it noticed the stairs and became so
frightened and preoccupied it forgot all about what it had been
doing''. Knowing what we do about what really happened in the robot,
we might be tempted to berate this poor person for using such sloppy
anthropomorphic concepts as determination, fear, preoccupation, and
forgetfulness in describing the actions of a machine. We could berate
the person, but it would be wrong.  The robot came by the emotions and
foibles indicated as honestly as any living animal; the observed
behavior is the correct course of action for a being operating with
uncertain data in a dangerous and uncertain world. An octopus in
pursuit of a meal can be diverted by hints of danger in just the way
the robot was. An octopus also happens to have a nervous system that
evolved entirely independently of our own vertebrate version. Yet most
of us feel no qualms about ascribing qualities like passion, pleasure,
fear, and pain to the actions of the animal. We have in the behavior
of the vertebrate, the mollusc, and the robot a case of convergent
evolution. The needs of the mobile way of life have conspired in all
three instances to create an entity that has modes of operation for
different circumstances, and that changes quickly from mode to mode on
the basis of uncertain and noisy data prone to misinterpretation. As
the complexity of the mobile robots increases, their similarity to
animals and humans will become even greater.

	Hold on a minute, you say. There may be some resemblance
between the robot's reaction to a dangerous situation and an animal's,
but surely there are differences. Isn't the robot more like a startled
spider, or even a bacterium, than like a frightened human being?
Wouldn't it react over and over again in exactly the same way, even if
the situation turned out not to be dangerous? You've caught me. I
think the spider's nervous system is an excellent match for robot
programs possible today. We passed the bacterial stage in the 1950s
with light-seeking electronic turtles. This does not mean that
concepts like thinking and consciousness are ruled out. In the book
{\bf Animal Thinking}, the animal ethologist D. G. Griffiths reviews
evidence that much animal behavior, including that of insects, can be
explained economically in terms of consciousness: an internal model of
the self and surroundings, that, however crudely, allows consideration
of alternative actions. But there are differences of degree.

\section{Other Emotions}

	 When tickled, the sea slug Aplysia withdraws its delicate
gills into its body. If the tickling is repeated often, Aplysia
gradually learns to ignore the nuisance, and the gills remain
deployed. If, later, tickles are followed by harsh stimuli, such as
contact with a strong acid, the withdrawal reflex returns with a
vengeance. Either way, the modified behavior is remembered for
hours. Aplysia has been studied so thoroughly in the last few decades
that the neurons involved in the reflex are well known, and the
learning has recently been traced to chemical changes in single
synapses on these neurons. Larger networks of neurons can adapt in
more elaborate ways, for instance by learning to associate specific
pairs of stimuli with one another. Such mechanisms tune a nervous
system to the body it inhabits, and to its environment. Vertebrates
owe much of their behavioral flexibility to an elaboration of this
arrangement, systems that can be activated from many locations that
encourage and discourage future repetitions of recent
behaviors. Though their neural architecture is not understood, their
effect is self evident in the subjective sensations we call pleasure
and pain.

	 A unified conditioning mechanism has obvious advantages in
guiding an animal through a changing world. It seems to me that it
also conveys a long term evolutionary advantage by providing a
``cheap'' means of entry into fundamental new behaviors. A new need or
danger can be accommodated through a modest mutation of the sensory
neural wiring, the connection of a detector for the condition to a
pleasure or pain site. The standard conditioning mechanism will then
ensure that animals with the mutation learn to seek the conditions
that meet the need, or to avoid the danger, even if the required
behavior is complex. Without the learning mechanism a much more
specific sensor to motor connection would have to be discovered.

	We are deep in the realm of speculation now, but the
generalpleasure/pain learning mechanism may provide an explanation for
abstract emotions. Let's suppose that altruism, for instance of a
mother toward her offspring, can enhance the long term survival of the
altruist's genes even though it has a negative effect on the
individual altruist.  Feeding the young may leave the mother exhausted
and hungry, and defending them may involve her in risk of injury.  In
a successful animal hunger and injury would surely be wired to
register as pain. Wouldn't the conditioning mechanisms we've just
described then eventually suppress a mother's ministrations on behalf
of her young?

	Activities whose beneficial or detrimental effects act only
across the generations can be conditioned just as readily those with
more immediate effects, if detectors for them are wired strongly to
pleasure and pain sites. For instance, mother love is encouraged if
the sight, feel, sound or smell of the offspring triggers pleasure,
and if absence of the young is painful. To the extent that
conditioning stimuli have subjective manifestations other than the
pain or pleasure sensation itself, such long range causes are likely
to {\it feel} different from more immediate ones like skin pain or
hunger. Most of the immediate concerns are associated with some part
of the body and can be usefully mapped there in the organism's
conscious map of self and world. Multigenerational imperatives, on the
other hand, cannot be so simply related to the physically apparent
world. This may help explain the ethereal or spiritual associations
people often assign to them.  Certainly they deserve respect, being
the distillation of perhaps tens of millions of years of life or death
trials, the wisdom of many lifetimes.

\section{What If?}

	 Elaboration of the internal world model in higher animals
made possible another twist. A rich world model allows its possessor
to examine in detail alternative situations, past, future, or merely
hypothetical.  Dangers avoided can yet be brooded over, and what {\it
might} have happened can be imagined. If the mental simulation is
accurate enough, such brooding can produce useful warnings, or point
out missed opportunities. These lessons of the imagination are most
effective if their consequences are tied to the conditioning
mechanism, just as with real events. Such a connection is particularly
easy to explain if, as we elaborate below, the most powerful aspects
of reason are due to world knowledge powerfully encoded in the sensory
and motor systems. The same wiring that conditions in real situations
would be activated for imaginary ones.

	The ability to imagine must be a key component in
communication among higher animals (and between you and me). Messages
trigger mental scenarios that then provide conditioning (i.e.,
learning). Communication that fails to engage the emotions is not very
educational in this sense, and a waste of time. Imagine circuitry for
detecting time well spent and time wasted wired to the conditioning
centers. It's not too far fetched to think that these correspond to
the subjective emotions of ``interesting'' and ``boring''. Humans seem
to have cross wiring that allows elaborate imagining, for instance
about future rewards, to make interesting activities that might
normally be boring. How else can one explain the existence of
intellectuals? Indeed, the conventional view of intelligence, and the
bulk of work in artificial intelligence, centers on this final
twist. While I believe that it is important, it is only a tiny part of
the whole story, and often overrated.

\section{Coming Soon}

	 A few of the ideas above have been explored in machinery. I
mentioned earlier that W. Grey Walter built electronic turtles which
demonstrated learning by association, represented as charges on a
matrix of capacitors.  Arthur Samuel at IBM wrote a checker playing
program that adjusted evaluation parameters to improve its play, and
was able to learn simply by playing game after game against itself
overnight. Frank Rosenblatt of Cornell invented networks, called
``perceptrons,'' of artificial neurons that could be trained to do
simple tasks by properly timed punish and reward signals. These
approaches of the 1950s and 1960s fell out of fashion in the 1970s,
but modern variations of it are again in vogue.

	 Among the natural traits in the immediate roving robot
horizon is parameter adjustment learning. A precise mechanical arm in
a rigid environment can usually be ``tuned'' for optimal control once,
permanently.  A mobile robot bouncing around in the muddy world, on
the other hand, is likely to continuously suffer insults like dirt
buildup, tire wear, frame bends, and small mounting bracket slips that
ruin precise adjustments. Some of the programs that drive our robots
through obstacle courses now have a camera calibration phase. The
robot is parked with its camera ``eye'' facing a precisely painted
grid of spots. A program notes where the spots appear in the camera's
images and figures a correction for camera distortions, so that later
programs can make precise visual angle measurements. The driving
program is highly sensitive to miscalibrations, and we are working on
a method that will continuously calibrate the cameras just from the
images perceived on normal trips through clutter. With such a
procedure in place, a bump that slightly shifts one of the robot's
cameras will no longer cause systematic errors in its
navigation. Animals seem to tune most of their nervous systems with
processes of this kind, and such accommodation may be a precursor to
more general kinds of learning.

	 Perhaps more controversially, I see the beginnings of
awareness in the minds of our machines. The more advanced control
programs use data from the robot's sensors to maintain
representations, at varying levels of abstraction and precision, of
the world around the robot, of the robot's position within that world,
and of the robot's internal condition. The programs that plan actions
for the robot manipulate these ``world models'' to weigh alternative
future moves. The world models can also be stored from time to time,
and examined later, as a basis for learning. A verbal interface keyed
to these programs would be able to meaningfully answer questions like
``Where are you?'' (``I'm in an area of about twenty square meters,
bounded on three sides, and there are three small objects in front of
me'') and "Why did you do that?" (``I turned right because I didn't
think I could fit through the opening on the left ''). Our programs
usually present such information from their world models in the form
of pictures on computer screens--a direct window into their minds.

\section{When?}

	How does computer speed compare with human thought?  The
answer has been changing.

	The first electronic computers were constructed in the mid
1940s to solve problems too large for unaided humans.  {\it Colossus},
one of a series of ultrasecret British machines, broke the German {\it
Enigma} code, greatly influencing the course of the European war, by
scanning through code keys tens of thousands of times faster than
humanly possible.  In the US {\it Eniac} computed antiaircraft
artillery tables for the Army, and later did calculations for the
atomic bomb, at similar speeds.  Such feats earned the early machines
the popular appellation {\it Giant Brains}.

	In the mid 1950s computers more than ten times faster than
Eniac appeared in many larger Universities. They did numerical
scientific calculations nearly a million times faster than humans. A
few visionaries took the Giant Brains metaphor seriously and began to
write programs for them to solve intellectual problems going beyond
mere calculation.  The first such programs were encouragingly
successful.  Computers were soon solving logic problems, proving
theorems in Euclidean geometry, playing checkers, even doing well in
IQ test analogy problems. The performance level and the speed in each
of these narrow areas was roughly equivalent to that of a college
freshman who had recently learned the subject.  The automation of
thought had made a great leap, but paradoxically the term ``Giant
Brains'' seemed less appropriate.

	In the mid 1960s a few centers working in this new area of
{\it Artificial Intelligence} added another twist: mechanical eyes,
hands and ears to provide real world interaction for the thinking
programs.  By then computers were a thousand times faster than Eniac,
but programs to do even simple things like clearing white blocks from
a black tabletop turned out to be very difficult to write, and
performed hundreds of times more slowly, and much less reliably, than
a human.  Slightly more complicated tasks took much longer, and many
seemingly trivial things, like identifying a few simple objects in a
jumble, still cannot be done acceptably at all twenty years later,
even given hours of computer time.  Forty years of research and a
millionfold increase in computer power has reduced the image of
computers from Giant Brains to mental midgets.  Is this silly, or
what?

\section{Easy and Hard}

	The human evolutionary record provides a clue to the
paradox. While our sensory and muscle control systems have been in
development for almost a billion years, and common sense reasoning has
been honed for perhaps a million, really high level, deep, thinking is
little more than a parlor trick, culturally developed over a few
thousand years, which a few humans, operating largely against their
natures, can learn. As with Samuel Johnson's dancing dog, what is
amazing is not how well it is done, but that it is done at all.

	Computers can challenge humans in intellectual areas, where
humans are evolutionary novices, because they can be programmed to
carry on much less wastefully.  Arithmetic is an extreme example, a
function learned by humans with great difficulty, but instinctive to
computers. A 1987 home computer can add a million large numbers in a
second, astronomically faster than a person, and with no errors. Yet
the 100 billion neurons in a human brain, if reorganized by a mad
neurosurgeon into adders using switching logic design principles,
could sum one hundred thousand times faster than the computer.

\section{Retina and Computer}

	The retina is the best studied piece of the vertebrate nervous
system. Though located at the back of the eyeball, some distance from
the bulk of the brain, it is really an elongated extension of the
brain.  Its separation makes it comparatively easy to study, even in
living animals. Removed from the body, it can be kept functioning for
hours, with its inputs and outputs highly accessible.  Transparent,
and thinner than a sheet of paper, it is ideal for light and electron
microscopic examination, when stained with dyes to make specific
neurons visible. It consists of a layer of light-detecting photocells
connected to a network of neurons that respond to contrast and motion
and more specific features in the image received by the eye. These
preprocessed data are then passed by the optic nerve to larger neural
centers in the brain.

	It is a peculiar feature of the vertebrate retina that light
must pass through the neural network to get to the photocells. This
is, no doubt, an unfortunate design choice, now locked in, made early
in the evolutionary history of the eye. The independently evolved
retinas of octopus and squid sensibly have their photoreceptors up
front.  The awkward position of the vertebrate retinal nerve net
greatly limits its size. On the other hand, there is strong selection
pressure to enhance its function. The retinal cells are in a unique
position to rapidly and comprehensively abstract the essentials from
an image, and good vision was a key survival tool for our ancestors:
life and death alternatives often depended on small differences in
visual speed or acuity. The product of this evolutionary adaptation is
bound to be a little atypical of the rest of the brain, where space is
larger, and payoffs for small improvements are more dilute.

	Retinal neurons, as I noted, form a thin sheet. Although nerve
tissue is usually gray or white, the retinal neurons, and their
supporting glial cells, are clear.  Retinal neurons are smaller than
most found in the brain.  Though the rest of the brain is too poorly
understood to be sure, the same pressures make it likely that the
retina is wired more precisely than neural centers with gentler
criteria, and that the retinal neurons are used more effectively. A
supporting fact is that the retinal neurons communicate among
themselves almost exclusively by smoothly varying voltages rather than
pulses, though their computations are ultimately encoded as pulses on
the optic nerve. This continuous mode works over only small distances
in the wet environment, but at that range is faster and more
precise. The retina may thus be representative of the most efficient
neural structures in vertebrates.

	So what does the retina actually do? A rough and ready answer
is available. Five cell types do most of the work. Photocells
(subdivided into cone cells, which together discriminate colors, and
rod cells which don't) intercept the light. Horizontal, bipolar, and
amacrine cells, working with continuous voltages, process the
image. Ganglion cells, whose axons form the optic nerve, combine
outputs from the other cells and produce pulsed signals that go into
the brain. After adapting to a particular overall light level,
clusters of photocells create a voltage proportional to the amount of
light striking them. This signal is received by two classes of
neurons, the horizontal cells and the bipolar cells. The horizontal
cells, whose thousands of fibers cover large circular fields of
photocells, produce a kind of average of their areas. If the voltages
of all the horizontal cells were mapped onto a TV screen, a blurry
version of the retinal image would be displayed. The bipolar cells, on
the other hand, are wired only to small areas, and would provide a
sharp picture on our imaginary TV. Some of the bipolar cells also
receive inputs from nearby horizontal cells, and then compute a
difference between the small bipolar center areas and the large
horizontal surround. Viewed on our TV, their picture would look much
paler than the original, except at the edges of objects and patterns,
where a distinct bright halo would be seen. The bipolar cell axons
connect to complicated multilayer synapses on the axonless amacrine
cells. Each ganglion cell collects inputs from several of these
amacrine synapses, and produces a pulsed output, which travels up its
long axon. Each amacrine cell connects to several bipolar and ganglion
cells, and some of the juctions appear to both send and receive
signals.  Some amacrine cells enhance the ``center surround''
response, others detect changes in brightness in parts of the
image. On the TV, some of these would show only objects moving left to
right, while others would reveal other directions of motion. Each
ganglion cell connects to several bipolar and amacrine cells, and
produces pulse streams whose rate is proportional to a computed
feature of the image. Some report on high contrast in specific parts
of the picture, others on various kinds of motion, or combinations of
contrast and motion.

	The TV I'm referring to is not totally imaginary. Sitting next
to me as I write is a TV monitor that often displays images just like
those described.  They come not from an animal's retina, but from the
eye of a robot. The picture from a TV camera on the robot is converted
by electronics into an array of numbers in a computer memory. Programs
in the computer combine these numbers to deduce things about the
robot's surroundings.  Though designed with little reference to
neurobiology, many of the program steps resemble strongly the
operations of the retinal cells--a case of convergent evolution. The
parallel provides a way to measure the net computational power of
neural tissue.

	The human retina has about 100 million photocells, tens of
millions of horizontal, bipolar and amacrine cells, and one million
ganglion cells, each contributing one signal-carrying fiber to the
optic nerve. All this is packaged in a volume a third of a millimeter
thick and less than a centimeter square, $1/100,000$ the volume of the
whole brain.  The photocells interact with their neighbors to enhance
each other's output, and their great multiplicity appears to be a way
to maximize sensitivity; a single photon sometimes produces a
detectable response. The horizontal and bipolar cells and the amacrine
cell synapses each seem to perform a unique computation. The bottom
line, however, is that the million ganglion cell axons each reports on
a specific function computed over a particular patch of photocells.

	To find the computer equivalent for such a function, we'll
first have to match the visual detail of the human eye in the computer
equivalent.  Simply counting photocells in the eye leads to an
overestimate, because they work in groups. External visual acuity
tests are better, but complicated by the fact that the retina has a
small, dense, high resolution center area, the fovea, which can
resolve details more than 10 times as fine as the rest of the
eye. Though it covers less than $1\%$ of the visual field, the fovea
employs perhaps one quarter of the retinal circuitry, and one quarter
of the optic nerve fibers. Under optimal seeing conditions as many as
500 distinct points can be resolved across the width of this central
region. This feat could be matched by a TV camera with 500 separate
picture elements, or ``pixels,'' in the horizontal direction. The
vertical resolution is similar, so our camera would need $500 \times
500$, or one quarter million pixels, in all--which, incidentally, just
happens to be the resolution of a good quality image on a standard
television set. But don't we see more finely than that? Not
really. The $500 \times 500$ array corresponds only to our fovea,
spanning a mere $5\deg$ of our field of view. A standard TV screen
subtends about $5\deg$ when viewed from a distance of 10 meters.  At
that range, the scanning lines and other resolution defects of the TV
image are invisible because the resolution of our eye is no better. At
closer range we can concentrate our fovea on small parts of the TV
image to get greater detail. We have the illusion of seeing the whole
screen this sharply because our unconsciously swiveling eyes rapidly
zip the foveal area from one place to another. Somewhere, in an as yet
mysterious part of our brain, a high resolution image is being
synthesized like a jigsaw puzzle from these fragmentary glimpses.

	So the foveal circuitry in the retina effectively takes a
$500\times 500$ image and processes it to produce 250,000 values, some
being center-surround operations, some being motion detections. One
key question remains. How fast does this happen? Experience with
motion pictures provides a ready answer. When successive frames are
presented at a rate slower than about ten per second, the individual
frames become distinguishable. At faster rates they blend together
into apparently smooth motion. Though the separate frames cannot be
distinguished faster than ten per second, if the light flickers at the
frame rate, as it does in a movie projector and on a TV screen, the
flicker itself is detectable until it reaches a frequency of about 50
flashes per second. Presumably in the 10-50 cycle range the simplest
brightness change detectors are triggered, but the more complicated
neuron chains do not have time to react.

	In our lab we have often programmed computers to do
center-surround operations on images from TV toting robots, and once
or twice we have written motion detectors. To get the speed up, we
have spent much programming effort and mathematical trickery to do the
job as efficiently as possible. Despite our best efforts, 10 frame per
second processing rates have been out of reach because our computers
are too slow. In a rough sense, with an efficient program a
center-surround calculation applied to each pixel in a $500 \times
500$ image takes about 25 million calculations, which breaks down to
about 100 calculations for each center-surround value produced. A
motion-detecting operator can be applied at a similar cost.
Translated to the retina, this means that each ganglion cell reports
on the computer equivalent of 100 calculations every tenth of a
second, and thus represents 1000 calculations per second. The whole
million-fiber optic nerve then carries the answers to a billion
calculations per second.

	If the retina's processing can be matched by a billion
computer calculations per second, what can we say about the entire
brain? The brain has about 1000 times as many neurons as the retina,
but its volume is 100,000 times as large. The retina's evolutionarily
pressed neurons are smaller and more tightly packed than average. By
multiplying the computational equivalent of the retina by a compromise
value of 10,000 for the ratio of brain complexity to retina
complexity, I rashly conclude that the whole brain's job might be done
by a computer performing 10 trillion ($10^{13}$) calculations per
second. This is about a million times faster than the medium size
machines that now drive my robots, and one thousand times more than
today's fastest supercomputers.

\begin{figure}
\vspace{6in}
\caption{Computing speed and memory of some animals and machines. The
animal figures are for the nervous system only, calculated at 100 bits
per second and 100 bits of storage per neuron.  These are speculative
estimates, but note that a factor of 100 one way or the other would
change the appearance of the graph only slightly.}
\label{brains}
\end{figure}

\section{Intellectual Voyages}

	Interesting computation and thought requires a processing
engine of sufficient computational {\it power} and {\it
capacity}. Roughly, power is the speed of the machine, and capacity is
its memory size.

	Here's a helpful metaphor. Computing is like a sea voyage in a
motorboat. How fast a given journey can be completed depends on the
power of the boat's engine.  The maximum length of any journey is
limited by the capacity of its fuel tank.  The effective speed is
decreased, in general, if the course of the boat is constrained, for
instance to major compass directions.

	Some computations are like a trip to a known location on a
distant shore, others resemble a mapless search for a lost island.
Parallel computing is like having a fleet of small boats - it helps in
searches, and in reaching multiple goals, but not very much in
problems that require a distant sprint. Special purpose machines trade
a larger engine for less rudder control.

	Attaching disks and tapes to a computer is like adding
secondary fuel tanks to the boat. The capacity, and thus the range, is
increased, but if the connecting plumbing is too thin, it will limit
the fuel flow rate and thus the effective power of the engine.

	Extending the metaphor, input/output devices are like boat
sails.  They capture power and capacity in the environment.  Outside
information is a source of variability, and thus power, by our
definition. More concretely, it may contain answers that would
otherwise have to be computed. The external medium can also function
as extra memory, increasing capacity.

	Figure \ref{brains} shows the power and capacity of some
interesting natural and artificial thinking engines.  At its best, a
computer instruction has a few tens of bits of information, and a
million instruction per second computer represents a few tens of
millions of bits/second of power.  The power ratio between nervous
systems and computers is as calculated in the last section: a million
instructions per second is worth about a hundred thousand neurons.  I
also assume that a neuron represents about 100 bits of storage,
suggested by recent evidence of synaptic learning in simple neurvous
systems by Eric Kandel and others. Note that change of a factor of ten
or even one hundred in these ratios would hardly change the graph
qualitatively.  (My forthcoming book {\it Mind Children}, from which
this paper is drawn, offers more detailed technical justifications for
these numbers).

	The figure shows that current laboratory computers are equal
in power approximately to the nervous systems of insects.  It is these
machines that support essentially all the research in artificial
intelligence. No wonder the results to date are so sparse! The largest
supercomputers of the mid 1980s are a match for the 1 gram brain of a
mouse, but at ten million dollars or more apiece they are reserved for
serious work.

\begin{figure}
\vspace{6in}
\caption{A Century of Computing - The cost of calculation has dropped
a thousandfold every twenty years (or halved every two years) since
the late nineteenth century. Before then mechanical calculation was an
unreliable and expensive novelty with no particular edge over hand
calculation. The graph shows a mind boggling {\it trillionfold}
decrease in the cost since then. The pace has actually picked up a
little since the beginning of the century. It once took 30 years to
accumulate a thousandfold improvement; in recent decades it takes only
19.  Human equivalence should be affordable very early in the 21st
century.}
\label{compute}
\end{figure}


\section{The Growth of Processing Power}

	How long before the research medium is rich enough for full
intelligence?

	Although a number of mechanical digital calculators were
devised and built during the seventeenth and eighteenth centuries,
only with the mechanical advances of the industrial revolution did
they become reliable and inexpensive enough to routinely rival manual
calculation. By the late nineteenth century their edge was clear, and
the continuing progress dramatic.  Since then the cost of computing
has dropped a thousandfold every twenty years (figure \ref{compute}).

	The early improvements in speed and reliability came with
advances in mechanics - precision mass produced gears and cams, for
instance, improved springs and lubricants, as well as increasing
design experience and competition among the calculator manufacturers.
Powering calculators by electric motors provided a boost in both speed
and automation in the 1920s, as did incorporating electromagnets and
special switches in the innards in the 1930s.  Telephone relay methods
were used to make fully automatic computers during World War II, but
these were quickly eclipsed by electronic tube computers using radio,
and ultrafast radar, techniques.  By the 1950s computers were an
industry that itself spurred further major component improvements.

	The curve in figure \ref{compute} is not leveling off, and the
technological pipeline is full of developments that can sustain the
pace for the foreseeable future. Success in this enterprise, as in
others, breeds success. Not only is an increasing fraction of the best
human talent engaged in the research, but the ever more powerful
computers themselves feed the process. Electronics is riding this
curve so quickly that it is likely to be the main occupation of the
human race by the end of the century.

	The price decline is fueled by miniaturization, which supplies
a double whammy. Small components both cost less and operate more
quickly.  Charles Babbage, who in 1834 was the first person to
conceive the idea of an automatic computer, realized this. He wrote
that the speed of his design, which called for hundreds of thousands
of mechanical components, could be increased in proportion if ``as the
mechanical art achieved higher states of perfection'' his palm sized
gears could be reduced to the scale of clockwork, or further to
watchwork. (I fantasize an electricityless world where the best minds
continued on Babbage's course.  By now there would be desk and pocket
sized mechanical computers containing millions of microscopic gears,
computing at thousands of revolutions per second.)

	To a remarkable extent the cost per pound of machinery has
remained constant as its intricacy increased.  This is as true of
consumer electronics as of computers (merging categories in the
1980s). The radios of the 1930s were as large and as expensive as the
televisions of the 1950s, the color televisions of the 1970s, and the
home computers of the 1980s. The volume required to amplify or switch
a single signal dropped from the size of a fist in 1940, to that of a
thumb in 1950, to a pencil eraser in 1960, to a salt grain in 1970, to
a small bacterium in 1980. In the same period the basic switching
speed rose a millionfold, and the cost declined by the same huge
amount.

	Predicting the detailed future course is impossible for many
reasons.  Entirely new and unexpected possibilities are encountered in
the course of basic research. Even among the known, many techniques
are in competition, and a promising line of development may be
abandoned simply because some other approach has a slight edge. I'll
content myself with a short list of some of what looks promising
today.

	In recent years the widths of the connections within
integrated circuits have shrunk to less than one micron, perilously
close to the wavelength of the light used to ``print'' the circuitry.
The manufacturers have switched from visible light to shorter
wavelength ultraviolet, but this gives them only a short respite.
X-rays, with much shorter wavelengths, would serve longer, but
conventional X-ray sources are so weak and diffuse that they need
uneconomically long exposure times. High energy particle physicists
have an answer.  Speeding electrons curve in magnetic fields, and
spray photons like mud from a spinning wheel.  Called synchotron
radiation for the class of particle accelerator where it became a
nuisance, the effect can be harnessed to produce powerful beamed
X-rays.  The stronger the magnets, the smaller can be the
synchotron. With liquid helium cooled superconductiong magnets an
adequate machine can fit into a truck, otherwise it is the size of a
small building. Either way, synchotrons are now an area of hot
interest, and promise to shrink mass-produced circuitry into the
sub-micron region. Electron and ion beams are also being used to write
submicron circuits, but present systems affect only small regions at a
time, and must be scanned slowly across a chip.  The scanned nature
makes computer controlled electron beams ideal, however, for
manufacturing the ``masks'' that act like photographic negatives in
circuit printing.

	Smaller circuits have less electronic ``inertia'' and switch
both faster and with less power. On the negative side, as the number
of electrons in a signal drops it becomes more prone to thermal
jostling. This effect can be countered by cooling, and indeed very
fast experimental circuits can now be found in many labs running in
supercold liquid nitrogen, and one supercomputer is being designed
this way. Liquid nitrogen is produced in huge amounts in the
manufacture of liquid oxygen from air, and it is very cheap (unlike
the much colder liquid helium).

	The smaller the circuit, the smaller the regions across which
voltages appear, calling for lower voltages.  Clumping of the
substances in the crystal that make the circuit becomes more of a
problem as they get smaller, so more uniform ``doping'' methods are
being developed. As the circuits become smaller quantum effects become
more pronounced, creating new problems and new
opportunities. Superlattices, mutiple layers of atoms-thick regions of
differently doped silicon made with molecular beams, are such an
opportunity.  They allow the electronic characteristics of the
material to be tuned, and permit entirely new switching methods, often
giving tenfold improvements.

	The first transistors were made of germanium; they could not
stand high temperatures and tended to be unreliable. Improved
understanding of semiconductor physics and ways of growing silicon
crystals made possible faster and more reliable silicon transistors
and integrated circuits.  New materials are now coming into their
own. The most immediate is gallium arsenide. Its lattice impedes
electrons less than silicon, and makes circuits up to ten times
faster. The Cray 3 supercomputer due in 1989 will use gallium arsenide
integrated circuits, packed into a one cubic foot volume, to top the
Cray 2's speed tenfold. Other compounds like indium phosphide and
silicon carbide wait in the wings. Pure carbon in diamond form is an
obvious possibility - it should be as much an improvement over Gallium
Arsenide as that crystal is over Silicon. Among its many superlatives,
perfect diamond is the best solid conductor of heat, an important
property in densely packed circuitry. The vision of an utradense three
dimensional circuit in a gem quality diamond is compelling. As yet no
working circuits of diamond have been reported, but excitment is
mounting as reports of diamond layers up to a millimeter thick grown
from hot methane come from the Soviet Union, Japan and, belatedly, the
United States.

	Farther off the beaten track are optical circuits that use
lasers and non-linear optical effects to switch light instead of
electricity. Switching times of a few picoseconds, a hundred times
faster than conventional circuits, have been demonstrated, but many
practical problems remain. Finely tuned laser has also been used with
light sensitive crystals and organic molecules in demonstration
memories that store up to a trillion bits per square centimeter.

	The ultimate circuits may be superconducting quantum devices,
which are not only extremely fast, but extremely efficient. Various
superconducting devices have been in and out of fashion several times
over the past twenty years. They've had a tough time because the
liquid helium environment they require is expensive, the
heating/cooling cycles are stressful, and especially because rapidly
improving semiconductors have offered such tough competition.

	Underlying these technical advances, and preceding them, are
equally amazing advances in the methods of basic physics. One recent,
unexpected and somewhat unlikely, device is the inexpensive tunnelling
microscope that can reliably see, identify and soon manipulate single
atoms on surfaces by scanning them with a very sharp needle. The tip
is positioned by three piezoelectric crystals microscopically moved by
small voltages. It maintains a gap a few atoms in size by monitoring a
current that jumps across it. The trickiest part is isolating the
system from vibrations. It provides our first solid toehold on the
atomic scale.

	A new approach to miniaturization is being pursued by
enthusiasts in the laboratories of both semiconductor and
biotechnology companies, and elswhere. Living organisms are clearly
machines when viewed at the molecular scale.  Information encoded in
RNA ``tapes'' directs protein assembly devices called ribosomes to
pluck particular sequences of amino acids from their environment and
attach them to the ends of growing chains. Proteins, in turn, fold up
in certain ways, depending on their sequence, to do their jobs. Some
have moving parts acting like hinges, springs, latches triggered by
templates.  Others are primarily structural, like bricks or ropes or
wires.  The proteins of muscle tissue work like ratcheting
pistons. Minor modifications of this existing machinery are the core
of today's biotechnology industry. The visionaries see much greater
possibilities.

	Proteins to do specific jobs can be engineered even without a
perfect model of their physics.  Design guidelines, with safety
margins to cover the uncertainties, can substitute.  The first
generation of artificial molecular machinery would be made of protein
by mechanisms recruited from living cells. Early products would be
simple, like tailored medicines, and experimental, like little
computer circuits.  Gradually a bag of tricks, and computer design
aids, would accumulate to build more complicated machines.  Eventually
it may be possible to build tiny robot arms, and equally tiny
computers to control them, able to grab molecules and hold them,
thermally wriggling, in place.  The protein apparatus could then be
used as machine tools to build a second generation of molecular
devices by assembling atoms and molecules of all kinds. For instance,
carbon atoms might be laid, bricklike, into ultra strong fibers of
perfect diamond. The smaller, harder, tougher machines so produced
would be the second generation molecular machinery.

	The book {\bf Engines of Creation} by Eric Drexler, and a
forthcoming book by Conrad Schneiker, call the entire scheme {\it
nanotechnology}, for the nanometer scale of its parts. By contrast
today's integrated circuit microtechnology has micrometer features, a
thousand times bigger.  Some things are easier at the nanometer scale.
Atoms are perfectly uniform in size and shape, if somewhat fuzzy, and
behave predictably, unlike the nicked, warped and cracked parts in
larger

\section{A Stumble}

	It seemed to me throughout the 1970s (I was serving an
extended sentence as a graduate student at the time) that the
processing power available to AI programs was not increasing very
rapidly. In 1970 most of my work was done on a Digital Equipment
Corp. PDP-10 serving a community of perhaps thirty people. In 1980 my
computer was a DEC KL-10, five times as fast and with five times the
memory of the old machine, but with twice as many users. Worse, the
little remaining speedup seemed to have been absorbed in
computationally expensive convenience features: fancier time sharing
and high level languages, graphics, screen editors, mail systems,
computer networking and other luxuries that soon became necessities.

	Several effects together produced this state of
affairs. Support for university science in general had wound down in
the aftermath of the Apollo moon landings and politics of the Vietnam
war, leaving the universities to limp along with aging equipment. The
same conditions caused a recession in the technical industries -
unemployed engineers opened fast food restaurants instead of designing
computers (the rate of change in figure \ref{compute} does slacken
slightly in the mid 1970s). The initially successful ``problem
solving'' thrust in AI had not yet run its course, and it still seemed
to many that existing machines were powerful enough - if only the
right programs could be found. While spectacular progress in the
research became increasingly difficult, a pleasant synergism among the
growing number of information utilities on the computers created an
attractive diversion for the best programmers - creating more
utilities.

	If the 1970s were the doldrums, the 1980s more than
compensated.  Several salvations had been brewing. The Japanese
industrial successes focused attention worldwide on the importance of
technology, particularly computers and automation, in modern economies
- American industries and government responded with research
dollars. The Japanese stoked the fires, under the influence of a small
group of senior researchers, by boldly announcing a major initiative
towards future computers, the so called ``Fifth Generation'' project,
pushing the most promising American and European research directions.
The Americans responded with more money. Besides this, integrated
circuitry had evolved far enough that an entire computer could fit on
a chip. Suddenly computers were affordable by individuals, and a new
generation of computer customers and manufacturers came into
being. The market was lucrative, the competition fierce, and the
evolution swift, and by the mid 1980s the momentum lost in the
previous decade had been regained, with interest. Artificial
intelligence research is awash in a cornucopia of powerful new
``personal'' workstation computers, and there is talk of applying
supercomputers to the work.

	Even without supercomputers, human equivalence in a research
setting should be possible by around 2010, as suggested by figure
\ref{compute}. Now, the smallest vertebrates, shrews and hummingbirds,
get interesting behavior from nervous systems one ten thousandth the
size of a human's, so I expect fair motor and perceptual competence,
in about a decade.

\section{Faster Yet?}

	Very specialized machines can provide up to one thousand times
the effective performance for a given price in well defined
tasks. Some vision and control problems may be candidates for this
approach. Special purpose machines are not a good solution in the
groping research stage, but may dramatically lower the costs of
intelligent machines when the problems and solutions are well
understood. Some principals in the Japanese Fifth Generation Computer
Project have been quoted as planning ``man capable'' systems in ten
years. I believe this more optimistic projection is unlikely, but not
impossible.

	As the computers become more powerful and as research in this
area becomes more widespread the rate of visible progress should
accelerate. I think artificial intelligence via the ``bottom up''
approach of technological recapitulation of the evolution of mobile
animals is the surest bet because the existence of independently
evolved intelligent nervous systems indicates that there is an
incremental route to intelligence. It is also possible, of course,
that the more traditional ``top down'' approach will achieve its
goals, growing from the narrow problem solvers of today into the much
harder areas of learning, common-sense reasoning and perceptual
acquisition of knowledge as computers become large and powerful
enough, and the techniques are mastered. Most likely both approaches
will make enough progress that they can effectively meet somewhere in
the middle, for a grand synthesis into a true artificial sentience.

	This artificial person will have some interesting
properties. Its high level reasoning abilities should be astonishingly
better than a human's - even today's puny systems are much better in
some areas - but its low level perceptual and motor abilities will be
comparable to ours. Most interestingly it will be highly changeable,
both on an individual basis and from one of its generations to the
next. And it will quickly become cheap.

\footnote{Adapted from the forthcoming book {\bf Mind Children},
	Harvard University Press, Fall 1988. This work has been supported
	in part by the Office of Naval Research under contract N00014-81-K-503.

\end{document}