Just recently I have just finished reading The Singularity is Near by Ray Kurzweil. The basic premise of the book is that
information processing capability has been increasing exponentially since the beginning of
life on Earth. Furthermore, this exponentially increasing rate of processing is going to
increase, if not forever, for at least another two hundred years. By that point the human
society, certainly not flesh and blood humans, will be masters of all matter reachable at
half the speed of light.
Specifically, Ray Kurzweil posits that this will occur because of the Law of Accelerating
Returns. The premise behind this law is that as progress marches on an ever greater
proportion of every industry will be equivalent to information processing and thus able to
take advantage of the exponential increases in capability.
Ever increasing mastery of genetics, nanotechnology, brain mapping and simulation, and
Artificial Intelligence are noted as the keys to not only continuing the exponential path of
information processing, but also to solving all the problems of humankind. Of these problems
are specifically mentioned disease, poverty and death. The vast majority of the book is
spent showing examples of how various important industries have historically taken advantage
of the increases in information processing capabilities to progress further and then showing
more examples of how technologies which are extremely primitive at the time of publishing,
2005, and are only ten or twenty years from production will ensure the continued exponential
increase in information processing capabilities. A small portion of the book is devoted to
debating, rather weakly in my opinion, about how it is a moral imperative to maximally
develop these major technologies despite the serious innate risks.
The two primary arguments for the continued exponential increase in information processing
capacity are a history of exponential increases and strong AI. The first argument is that
because for all of history, recorded or otherwise, the information processing capability has
been increasing exponentially it is reasonable to expect it to continue otherwise until some
limit is hit suddenly. The strong AI argument predicts that before we hit the limit of human
understanding we will have AI strong enough to design even smarter AI and computing
machines, leading to continued exponential increases in processing capability per gram up to
the theoretical limits of physics.
The historical data provided is convincing of the increasing power of hardware in terms of
raw MIPS, however I believe that there is a flaw in the argument of Ray Kurzweil in that it
is not raw MIPS which determines capability, but effective MIPS. Effective MIPS is the
measurement of usable MIPS available after various efficiency losses are taken into account.
These losses are significant and include computer architecture losses such as cache misses,
communication losses such as synchronization, complexity management losses such as
abstraction and innate problem limitations such as necessary serialization. Though the raw
MIPS of hardware has been increasing exponentially the effect of these limitations have also
been increasing exponentially. While I do not have any hard data on hand I would still agree
that there is a net increase in the capability of hardware, but if this increase is
exponential then at the very least the exponent is closer to a linear function than doubling
every eighteen months as predicted by Moore's Law.
Though a decrease in the exponent does not entirely invalidate the argument it does
drastically change the time scales involved according to the Law of Accelerating Returns.
There is some evidence that the effective processing capability is not increasing as fast as
Ray Kurzweil believed provided by the five years of hindsight since the book was published.
Specifically the predictions of the available processing power for 2010 are off, already, by
a factor of two or three.
The second argument, that we will have strong AI before we reach some limit is sound in the
theoretical world. The limit which Ray Kurzweil eludes to is that of human creativity, that
is that we will create strong AI before humans are unable to contain enough of the design of
the necessary systems in our head to make forward progress. Whether the human mind is
sufficiently capable to create strong AI before reaching our assisted limits or not is
unclear. However, there are other limits which are not discussed in any depth which threaten
much more than humans just not being smart enough. In chemistry there is the concept of an
activation energy level. For many types of reactions if you plotted the energy of the
reaction, positive for energy put in, negative for energy put out, you will see a bump just
before the reaction starts to output energy. Unless sufficient energy is put in to crest the
hill the desired reaction will not occur. There is a similar requirement with the
development of any technology. Certainly a nuclear power station can generate trillions of
Watt-hours of energy throughout its lifetime, but if you do not have the energy to build the
plant in the first place you can never tap that energy, even if you already have all the
necessary knowledge.
Similarly it is with information processing capability. Though we can create systems of ever
greater processing capability, they will require ever greater energy to perform. In the book
it is argued based on the theoretical minimums that this will not be a problem, but to
achieve these minimum energy levels we require a minimum level of processing capability
which we do not currently meet. Thus ever more energy will be required until we have reached
the activation energy to enable low energy computing. It is much in doubt whether human
industry will choose to support this level of energy in competition with the other demands
on the finite energy generation capability. There is further the considerable possibility
that the cheap energy provided by oil will run out before the processing capacity necessary
for efficient computing is reached. Oil currently accounts for 37% of the world's energy
production. The loss of this proportion of the energy supply will greatly exacerbate the
competition information processing research faces and it seems likely that maintaining the
current industries will take precedence over new research.
With the threat of insufficient energy supplies in the near future ultra-efficient computing
may not come to fruition at all. Pushing back the timeline for sufficient processing
capability due to a reduced effective rate of increase makes it more likely that the energy
will run out before strong AI becomes a possibility.
Now that I have expressed my concerns relating to why I believe that the singularity may not
come about at all it is important to express my reasoning behind why it should be avoided.
My arguments are essentially that on the path to the Singularity lies the inevitable
extinction of the human race. I will demonstrate this by referring to the destructive power
of the major technologies Ray Kurzweil believes are necessary to power humanity to the
Singularity. In the book Ray Kurzweil covers each of these threats and concludes that they
are insufficient reasons to stop progress through two lines of reasoning.
The first line of reasoning is that these technologies hold the ability to reduce human
suffering and it is thus morally required that these technologies be developed to reduce
human suffering. This argument misses the point that there already exists the technology and
capability to drastically reduce the aggregate human suffering in the world and that if the
funds used to power technological progress were instead directed to making existing
technologies cheaper, more reliable and in distributing these tools to those in need the
majority of the human suffering in the world could be solved forever. It merely requires
sacrifice. Though it is not obvious, I also believe that this argument, when used in the
context of a specific technology, may lead to ignoring the unintended side effects of the
new technology and thus cause further suffering. One major example of this is manufacturing
automation. While the advantages of automated manufacturing are deemed quite valuable,
cheaper goods, the reduction in the number of unskilled jobs and the resulting unemployment
(It is not always possible or economical to retrain for the limited number of more skilled
positions which are soon to be automated) are often ignored.
The second line of reasoning, that as long as we are sufficiently security conscious these
technologies contain the necessary defencive tools, is based on two themes of invalid
reasoning. The first is that we are currently dealing quite satisfactorily with the
artificially created threats of computer viruses and their familiars. As counter evidence I
present the Internet. Even with the best security software a person can buy a careless user
will quickly have their computer infected with several virii and trojans and spam bots.
Currently the only real payoff for the creators of this malicious software is money or
personal information. A radical madman cannot effectively gain control of a significant
number of critical military systems to be able to launch missiles. However a single crackes
can easily amass a botnet of millions of nodes on the public Internet for the purposes of
DDOSing or spamming. If we are unable to protect ourselves against quite limited malicious
software should we attempt to allow malicious 'software' to take the much more potent form
of a custom virus or nanobot swarms? With a greater effect on the real world the ideological
payoffs increase greatly. Why limit yourself to getting the message out when you can devote
a couple of years of your life to wipe out the heathens yourself?
The second invalid theme Ray Kurzweil invokes is quite surprising for a book which is all
about the exponentially increasing power of technology. It is simply this, he makes the
assumption that the new threats of custom virii and malicious nanobot swarms will always be
of the same magnitude of existing diseases and threats. That is, that things which can wipe
a nation off the map instantly are going to be restricted to large governments, as are
nuclear weapons due to the cost of their creation, and that small threats will act pretty
much as diseases do now and take sufficient time to spread that they can be detected and
fought. Both these assumptions do not hold when it comes to powerful custom virii and
malicious nanobot swarms.
The most critical flaw is in assuming that these tools do not provide the power to instantly
destroy an entire population. The human immune system has evolved over millions of years to
handle threats of the sort that exist in nature. The threats which exist in nature, on the
other hand, have evolved to spread using the tools at their disposal. This means that
diseases which kill too quickly are limited in their spread by the size of the village.
Disease which don't spread or kill quickly enough give humans sufficient time to either
defend themselves directly, or evolve at least a partial defence. Now imagine a custom virus
which works like HIV, that is lays in wait for years before attacking and destroying the
immune system, but instead of stopping at the immune system proceeds to destroy any tissue
it has infected. Now imagine that it is transmitted like the common cold. Such a virus could
kill the vast majority of the human race in a decade.
Consider further the nanobot case. Since the human immune system has never seen a nanobot it
is likely ineffectual in defending against a nanobot infection. Let us further assume the
best case of Ray Kurzweil's future by assuming that we have a nanobot immune system covering
the Earth to prevent a grey goo scenario
which is ten times more effective against novel nanobots than the human immune system is
against novel viruses. Under such an assumption it is safe to further assume that at least
sometimes such an immune system will fail. This is likely because the determined madman can
just isolate a sample of the immune system and test thousands of nanobot/virus variants
against this sample to determine a set which either strain the system's limits or sneak by
entirely. If it only takes one nanobot swarm to convert a nation into goo or one virus to
destroy a population then any failure is not an option.
It is further not a valid assumption that these destructive technologies will be restricted
to large governments. The entire point of the singularity argument is that as progress moves
on more and more of the processes of creation will be information processing based and the
tools for that will become ever cheaper and wide spread. You may not be able to procure the
necessary technology from your corner store, but you could certainly steal it from a
University laboratory.
I believe it is clear that, given the range of mental reactions and states of all the people
in the world, it is unavoidable that there will be numerous disasters resulting from
maliciously designed viruses, bacteria (fast plastic eating bacteria anybody?) and nanobot
swarm which will kill millions on a regular basis. Further I believe that the ideological
bias which Kurzweil places in his book, that Neo Luddite beliefs are indefensible, is not
near as clear cut as his flippant responses to valid concerns may have you believe.
Overall Ray Kurzweil in The Singularity is Near does a good job of playing the
starry-eyed futurist, but fails to convince not only that the Singularity is likely to
happen, but even that it is desirable to cause it to happen.
Add a Comment