- Hanson, Robin.
The Age of Em.
Oxford: Oxford University Press, 2016.
ISBN 978-0-19-875462-6.
-
Many books, both fiction and nonfiction, have been devoted to the
prospects for and consequences of the advent of artificial
intelligence: machines with a general cognitive capacity which equals
or exceeds that of humans. While machines have already surpassed the
abilities of the best humans in certain narrow domains (for example,
playing games such as chess or go), you can't take a chess playing
machine and expect it to be even marginally competent at a task as
different as driving a car or writing a short summary of a newspaper
story—things most humans can do with a little experience. A
machine with “artificial general intelligence” (AGI) would
be as adaptable as humans, and able with practice to master a wide
variety of skills.
The usual scenario is that continued exponential progress in computing
power and storage capacity, combined with better understanding of how
the brain solves problems, will eventually reach a cross-over point
where artificial intelligence matches human capability. But since
electronic circuitry runs so much faster than the chemical signalling
of the brain, even the first artificial intelligences will be able to
work much faster than people, and, applying their talents to improving
their own design at a rate much faster than human engineers can
work, will result in an “intelligence explosion”, where
the capability of machine intelligence runs away and rapidly approaches
the physical limits of computation, far surpassing human cognition.
Whether the thinking of these super-minds will be any more comprehensible
to humans than quantum field theory is to a goldfish and whether humans
will continue to have a place in this new world and, if so, what it
may be, has been the point of departure for much speculation.
In the present book,
Robin Hanson,
a professor of economics at George Mason University,
explores a very different scenario.
What if the problem of artificial intelligence (figuring out how to
design software with capabilities comparable to the human brain)
proves to be much more difficult than many researchers assume,
but that we continue to experience exponential growth in computing
and our ability to map and understand the fine-scale structure of
the brain, both in animals and eventually humans? Then some time in
the next hundred years (and perhaps as soon as 2050), we may have the
ability to emulate the low-level operation of the brain with
an electronic computing substrate. Note that we need not have any idea
how the brain actually does what it does in order to do this: all we
need to do is understand the components (neurons, synapses,
neurotransmitters, etc.) and how they're connected together, then
build a faithful emulation of them on another substrate. This
emulation, presented with the same inputs (for example, the pulse
trains which encode visual information from the eyes and sound
from the ears), should produce the same outputs (pulse trains which
activate muscles, or internal changes within the brain which encode
memories).
Building an emulation of a brain is much like reverse-engineering an
electronic device. It's often unnecessary to know how the device
actually works as long as you can identify all of the components,
their values, and how they're interconnected. If you re-create that
structure, even though it may not look anything like the original
or use identical parts, it will still work the same as the prototype.
In the case of brain emulation, we're still not certain at what level
the emulation must operate nor how faithful it must be to the
original. This is something we can expect to learn
as more and more detailed emulations of parts of the brain are
built. The
Blue Brain Project
set out in 2005 to emulate one
neocortical
column
of the rat brain. This goal has now been achieved, and work is
progressing both toward more faithful simulation and expanding the
emulation to larger portions of the brain. For a sense of scale,
the human
neocortex consists
of about one million cortical columns.
In this work, the author assumes that emulation of the human brain
will eventually be achieved, then uses standard theories from the
physical sciences, economics, and social sciences to explore the
consequences and characteristics of the era in which emulations will
become common. He calls an emulation an “em”, and the
age in which they are the dominant form of sentient life on Earth
the “age of em”. He describes this future as
“troublingly strange”. Let's explore it.
As a starting point, assume that when emulation becomes possible, we
will not be able to change or enhance the operation of the emulated
brains in any way. This means that ems will have the same memory
capacity, propensity to forget things, emotions, enthusiasms,
psychological quirks and pathologies, and all of the idiosyncrasies of
the individual human brains upon which they are based. They will not
be the cold, purely logical, and all-knowing minds which science
fiction often portrays artificial intelligences to be. Instead, if you
know Bob well, and an emulation is made of his brain, immediately
after the emulation is started, you won't be able to distinguish Bob
from Em-Bob in a conversation. As the em continues to run and has its
own unique experiences, it will diverge from Bob based upon them, but,
we can expect much of its Bob-ness to remain.
But simply by being emulations, ems will inhabit a very different
world than humans, and can be expected to develop their own unique
society which differs from that of humans at least as much as the
behaviour of humans who inhabit an industrial society differs from
hunter-gatherer bands of the Paleolithic. One key aspect of
emulations is that they can be checkpointed, backed up, and copied
without errors. This is something which does not exist in biology,
but with which computer users are familiar. Suppose an em is about to
undertake something risky, which might destroy the hardware running
the emulation. It can simply make a backup, store it in a safe place,
and if disaster ensues, arrange to have to the backup restored onto
new hardware, picking up right where it left off at the time of the
backup (but, of course, knowing from others what happened to its
earlier instantiation and acting accordingly). Philosophers will fret
over whether the restored em has the same identity as the one which
was destroyed and whether it has continuity of consciousness. To
this, I say, let them fret; they're always fretting about something.
As an engineer, I don't spend time worrying about things I can't
define, no less observe, such as “consciousness”,
“identity”, or “the soul”. If I did, I'd
worry about whether those things were lost when undergoing general
anaesthesia. Have the wisdom teeth out, wake up, and get on with your
life.
If you have a backup, there's no need to wait until the em from which
it was made is destroyed to launch it. It can be instantiated on
different hardware at any time, and now you have two ems, whose life
experiences were identical up to the time the backup was made, running
simultaneously. This process can be repeated as many times as you
wish, at a cost of only the processing and storage charges to run the
new ems. It will thus be common to capture backups of exceptionally
talented ems at the height of their intellectual and creative powers
so that as many can be created as the market demands their
services. These new instances will require no training, but be able
to undertake new projects within their area of knowledge at the moment
they're launched. Since ems which start out as copies of a common
prototype will be similar, they are likely to understand one another
to an extent even human identical twins do not, and form clans of
those sharing an ancestor. These clans will be composed of subclans
sharing an ancestor which was a member of the clan, but which diverged
from the original prototype before the subclan parent backup was
created.
Because electronic circuits run so much faster than the chemistry of
the brain, ems will have the capability to run over a wide range of
speeds and probably will be able to vary their speed at will. The
faster an em runs, the more it will have to pay for the processing
hardware, electrical power, and cooling resources it requires. The
author introduces a terminology for speed where an em is assumed to
run around the same speed as a human, a kilo-em a thousand times
faster, and a mega-em a million times faster. Ems can also run
slower: a milli-em runs 1000 times slower than a human and a micro-em
at one millionth the speed. This will produce a variation in
subjective time which is entirely novel to the human experience. A
kilo-em will experience a century of subjective time in about a month
of objective time. A mega-em experiences a century of life about
every hour. If the age of em is largely driven by a population which
is kilo-em or faster, it will evolve with a speed so breathtaking as
to be incomprehensible to those who operate on a human time scale. In
objective time, the age of em may only last a couple of years, but to
the ems within it, its history will be as long as the Roman Empire.
What comes next? That's up to the ems; we cannot imagine what they
will accomplish or choose to do in those subjective millennia or millions
of years.
What about humans? The economics of the emergence of an em society
will be interesting. Initially, humans will own everything, but as
the em society takes off and begins to run at least a thousand times
faster than humans, with a population in the trillions, it can be
expected to create wealth at a rate never before experienced. The
economic doubling time of industrial civilisation is about 15 years.
In an em society, the doubling time will be just 18 months and
potentially much faster. In such a situation, the vast majority of
wealth will be within the em world, and humans will be unable to
compete. Humans will essentially be retirees, with their needs and
wants easily funded from the proceeds of their investments in
initially creating the world the ems inhabit. One might worry about
the ems turning upon the humans and choosing to dispense with them
but, as the author notes, industrial societies have not done this with
their own retirees, despite the financial burden of supporting them,
which is far greater than will be the case for ems supporting human
retirees.
The economics of the age of em will be unusual. The fact that an em,
in the prime of life, can be copied at almost no cost will mean that
the supply of labour, even the most skilled and specialised, will be
essentially unlimited. This will drive the compensation for labour
down to near the subsistence level, where subsistence is defined as
the resources needed to run the em. Since it costs no more to create
a copy of a CEO or computer technology research scientist than a
janitor, there will be a great flattening of pay scales, all settling
near subsistence. But since most ems will live mostly in virtual
reality, subsistence need not mean penury: most of their needs and
wants will not be physical, and will cost little or nothing to
provide. Wouldn't it be ironic if the much-feared “robot
revolution” ended up solving the problem of “income
inequality”? Ems may have a limited useful
lifetime to the extent they inherit the human characteristic of the
brain having greatest plasticity in youth and becoming
increasingly fixed in its ways with age, and consequently less able to
innovate and be creative. The author explores how ems may view death
(which for an em means being archived and never re-instantiated) when
there are myriad other copies in existence and new ones being spawned
all the time, and how ems may choose to retire at very low speed and
resource requirements and watch the future play out a thousand times
or faster than a human can.
This is a challenging and often disturbing look at a possible future
which, strange as it may seem, violates no known law of science and
toward which several areas of research are converging today. The book
is simultaneously breathtaking and tedious. The author tries to work
out every aspect of em society: the structure of cities,
economics, law, social structure, love, trust, governance, religion,
customs, and more. Much of this strikes me as highly speculative,
especially since we don't know anything about the actual experience of
living as an em or how we will make the transition from our present
society to one dominated by ems. The author is inordinately fond of
enumerations. Consider this one from chapter 27.
These include beliefs, memories, plans, names, property,
cooperation, coalitions, reciprocity, revenge, gifts,
socialization, roles, relations, self-control, dominance,
submission, norms, morals, status, shame, division of labor,
trade, law, governance, war, language, lies, gossip, showing off,
signaling loyalty, self-deception, in-group bias, and meta-reasoning.
But for all its strangeness, the book amply rewards the effort you'll
invest in reading it. It limns a world as different from our own as
any portrayed in science fiction, yet one which is a plausible future
that may come to pass in the next century, and is entirely consistent
with what we know of science. It raises deep questions of philosophy,
what it means to be human, and what kind of future we wish for our
species and its successors. No technical knowledge of computer
science, neurobiology, nor the origins of intelligence and
consciousness is assumed; just a willingness to accept the premise
that whatever these things may be, they are independent of the
physical substrate upon which they are implemented.
September 2016 