Re: The Future


"Charles Hottel" <chottel@xxxxxxxxxxxxx> wrote in message
news:IkVUh.4970$3P3.2819@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
"Oliver Wong" <owong@xxxxxxxxxxxxxx> wrote in message
news:bEOUh.10432$iq5.280939@xxxxxxxxxxxxxxxxxxxxxxx

You might be interested in reading Roger Penrose's "The Emperor's
New Mind". In it, Penrose convincingly argues that there's "something
going on" in the human brain which current physical theories cannot
explain. He suspects we'll need to develop a unifying theory for
quantum-gravity before we can make further progress in understanding
how the mind works.
[...]

Kurzweil says that there is as yet no proof of that and even if such
computation goes on in the brain there is nothing that prevents a
machine from functioning as a quantum computer and duplicating it.

Well, yes and no... Before we can build a machine which simulates
whatever quantum magic is going on in the brain, we'd have to first find
out what quantum magic exactly is going on in the brain. I mean, the
source code we provide to the quantum computer can't simply be:

PERFORM sameComputationThatTheBrainDoes UNTIL problemSolved
.

I thought his explaination of quantum computers, how they work and what
their limitations were the clearest and easiest to understand that I
have read to date. Having to figure out and duplicate these quantum
processes will delay the time table, but we already see computers
increasing their intellegence rapidly without quantum computing. We can
probably have very intelligent computers without it although it would be
required for uploading minds.

I'm a bit skeptical about this part. With a lot of hand waving,
computer science has categorized problems according to their "difficulty"
(how long it would take a computer to solve the given problem as a
function of the size of the input), and two classes of difficulty they've
come up with are P (which stands for "within Polynomial time") and NP
(stands for "Non-deterministic Polynomial time").

Polynomial time means it's possible to write a polynomial where the
variable is the size of the input, and the value of the whole polynomial
is an upperbound on the number of computational "steps" (and I'm
intentionally being vague about what constitutes a "step" here, but you
may approximate it as being a single CPU instruction) that a "traditional"
computer would have to perform to solve the problem. For example, sorting
a list of numbers is within polynomial time, because given a list of N
numbers, we can devise an algorithm that will sort the list within N^2
steps. If we're sorting a list of a million numbers (and thus the "size"
of the problem is said to be 10^6), it'll take at most 10^12 steps. If you
have a computer which can perform a billion steps per second (e.g. a 1Ghz
computer), you could solve the proble in 10^3 seconds, under 20 minutes.

Contrast this with Exponential-time problems (or EXP problems), where
the time it would take to solve the problem can expressed in the form of
2^N, for example. Solving a problem of size 10^6 would take 2^1000000
steps. Even if you had a billion Ghz computer, you probably wouldn't be
able to solve this problem before the universe collapsed upon itself.

NP problems lie somewhere between P and EXP. If we had some sort of
"magical" non-deterministic computer, we could solve NP problems in
polynomial time. But without these "magical" computers, it would take our
"normal" computers exponential time. "Cracking encryption" is an example
of an NP-problem: When some device claims it uses 1024 bits of encryption,
cracking its encryption is a problem of size 1024, which would take 2^1024
steps, which again with our computers would likely take longer than the
age of the universe: Google calculator won't let me compute 2^1024, but
2^999 =~ 5 * 10^300; a 100Ghz computer which can perform 100 billion
(10^11) steps per second would solve hte problem in 5 * 10^289 seconds, or
about 10^282 years.

It turns out that quantum computers have the properties of the above
mentioned "magical" non-deterministic computers. That is, quantum
computers can crack most encryption schemes that exist today within
polynomial time.

So while it's true that a traditional computer can simulate a quantum
computer, it would do so extremely slowly, and only feasibly for extremely
small problems. For example, for an NP problem of size 5, the traditional
computer would take 2^5 steps, or 32 steps, while a quantum computer would
take perhaps 5^2 = 25 steps. Not a big difference. Essentially (and again,
with a lot of handwaving), this means we could simulate a 5-neuron brain
at real time speed (and in fact, I think AI researchers are currently
simulating 300 neuron brain at real time speed; recall that a computer is
much faster than a brain, if we ignore the advantages of parallelism,
which doesn't show up very much in a mere 5-neuron, or even 300 neuron
brain). It becomes a big problem when you want to simulate a 100 billion
neuron brain.

But as you mentioned above, it's not clear whether you really need
"quantum magic" to simulate a human brain or not. If you don't, then
merely having a billion-core CPU ought to be able to do it. If you do,
then you'll probably need a quantum computer, and right now, quantum
computer technology is still in its infancy, so it's not clear what kind
of progress rate we can expect from it (e.g. will they too follow Moore's
Law?).

And as you mention below, we might not need to actually fully simulate
the brain in order to achieve Turing-test passing AI.

[...]

Kurzweil talks about simulating the brain at two levels and gives order
of magnitude estimates for how much computer power will be needed.

On level is to simulate the brain at level function units. This
requires somewhat less computations per second (cps) than simulating the
neurons and all of their non-linearities. The later level of simulation
would be need if you want to scan your barin an upload a copy to a
computer.

Right now we have functional level simulation of some regions of the
brain. There are probably several hundred of these functional regions.

Interesting, I hadn't heard anything about this area of research.

Right now the unused cycles of computers attached to the internet is
probably more than enough for functional simulation.

Yes, but good luck trying to get all the computers attached to the
internet to cooperative towards a single goal.

Around 2020 supercomputers will be powerful enough to do it and around
2025 a PC will have that much power. He predicts computers will start
passing the Turing test around 2030 and by around 2037 full simulation
of the neurons and their non-linearities will be possible.

Well, I'm predicting around 2047, so what you say sounds a tad
optimisitic, but not entirely unreasonable.

[...]

The AI guys seem to want to do it so they can say "see we were right all
along". Plus the idea of immortality and super intelligence with
god-like powers is very seductive to guys who like to learn things and
increase their knowledge. What worries me is that humans are a mixture
of both good and evil potential and the million times faster processing
will amplify both. Despite the rosey projections and optimistic
senariors I think nobody knows what the consequences will be. For every
positive picture presented you can imagine an equally powerful negative
picture. I think implementing it without some idea of what will happen
is irresponsible. Techology always seems to result is some unintended
consequences. As it gets more powerful those consequences could result
in the end of us all.

The whole idea of "The Singularity"
(http://en.wikipedia.org/wiki/Technological_singularity) is that you
*CANNOT* have some idea of what will happen once the technology is
implemented. And obviously, a lot of Sci-fi fiction play with this idea as
well.


He says we evolved from bacteria and we did not exterminate them so
super intelligent computers will not wipe us out.

It's more like we kill billions and billions of bacteria every second
without even noticing it (via our immune system), because they're so low
our radar, we don't even pay attention to them.

Hmm... suddenly, super-intelligent AI isn't sounding like such a good
idea...

In fact they will be human and will value and respect their human
heritage. He is optimistic about human values.

I'm not. It also depends on whether we program them to value and
respect human heritage, whether or not we allow them to reprogram
themselves, and whether or not we program them to not want to reprogram
the part of themselves which values and respects human heritage.

If that last part confused you, consider this: Let's say you have the
ability to reprogram yourself. Would you decide to reprogram yourself so
that you could be a cold blooded killer, feeling no empathy, regret or
remorse for taking other people's life? Probably not. Why not? Because
society has brainwashed you into thinking that reprogramming yourself to
be a cold blooded killer is a "bad" thing.

But even if we program the AIs to like humans, there's tons of sci-fi
fiction depicting bugs (either in the code, or even in the original
design/requirement specifications themselves) that let the AIs kill humans
anyway.

Take Asimov's three laws of robotics, for example. The first one is
says that a robot may not allow a human to come to harm, even through
inaction. Obviously, this an impossible requirement, as humans are
continuously "coming to harm" (in the form of aging). So the AI might then
decide to reformulate the rule as "minimize the amount of harm that any
human will come to". From there, it may conclude that the best way to do
that is to kill all humans. You'll have a sudden spike in harm in the
immediate future, but then for the rest of eternity, you've guaranteed no
harm will ever come to a human ever again.


Well we have not wiped out bacteria but we don't mind killing them by
the millions or billions when they get in our way. We experiment on
monkeys that share our genome to 95 to 99 percent.

Right; the way we handle animals with presumably inferior intelligence
is mixed. Some we tend to kill with little remorse (so called "pests",
such as rats, mosquitos, etc.) Others we try to protect, without
disturbing (animals considered to be endangered). Some we keep as pets.
Some we hunt (for food or for sport). How would we feel if AIs
"domesticated" us, so that we were no longer able to survive in "the
wild", but became entirely dependent on the AIs to take care of us? We may
have a higher quality of life than what "wild" humans might have
experienced, but we lose our independence, and become "second-class"
citizens in a world dominated by superior life forms.

Human values have produced world wars, mass exterminations, genocide and
many other horrible events. We often have good conscious intentions
that are thwarted by subconscious programs. Often these subconscious
programs are developed in childhood and they have a good purpose for us,
but their methods are those that make sense for a child but are not
appropriate behavior for an adult.What happens if these god-like super
intelligences get in a power struggle to see who will be the head god?.
What will the collateral damage be? What happens if the one million plus
increase in speed results in mental illness or frustration and
impatience with the slow responses of unenhanced humans? When they
sneeze will we all catch cold? If they can modify their design and
improve upon it at an ever increasing rate how long will they retain
their "humanity". He projects that late in this century a single
computer will have trillions of times the processing power of 10 billion
people. It does not take super intelligence to realize that mere humans
with nuclear weapons and nanoweapons could be preceived by these AIs as
a threat to their existence.

On the other hand, something with a trillion times the intelligence of
humanity as a collective might not consider mere humans with nuclear and
nano weapons to be a threat at all. Or at least, only as much as a threat
as the common cold virus is to us. Yes, it's *conceivable* that someone
might die from the common cold, but it's so low on our priority list of
things to worry about... The AIs might have more important things to worry
about than humans with bombs that can barely destroy a single planet,
nevermind a galaxy... they might be more concerned with what the *other*
AIs are up to.


He does not address this very deeply.

Understandbly, as we're way out into wild-speculation-land here.

Most of his arguments are with those who say that what he is proposing
cannot be done. He says the super intelligence will find a way to
overcome any obstacle.

It really depends on what he means by this. On the one hand, we still
face the same basic problems that animals face (how can I raise enough
money/find food to survive? how can I claim territory/buy a house? how can
I find a good mate? how can I raise my family successfully? etc.) only at
a higher abstraction level. It seems that mere intelligence is not enough
to solve these problems, because these problems arise from competition
with other agents of comparable intelligence. On the other hand,
presumably AIs don't need to worry about food or raising families (or do
they?).

From another interpretation, we as humans have essentially "solved",
at a superficial level at least, all the problems that animals were facing
(how to satisfy basic survival needs), and we've moved on to problems that
are probably inconceivable to the animals, like what to do about global
warming, pollution, corrupt government, etc.

He does not mention much about trying to control it, although some of
the refrences are to material that discusses steps we should take to
make sure we have friendly relations with them. I suppose if we try to
control it and it circumvents us then relations might not be very good.

That might not be the case. Think of park rangers to find an injured
animal such as a wolf. They might try to help the wolf, and the wolf, not
understanding what is going on, is aggressive and tries (futily) to attack
the rangers. The rangers are not upset with the wolf, or anything like
that. Rather, they understand that the wolf does not have the intellectual
capability to assess the situation, and is behaving from baser instincts.
The AIs might have the same outlook towards us.

It might be amusing to picture an AI training other AIs into
capturing, tagging, and releasing entire human societies back into the
wild for further study. "Careful there, if you aggravate them too much,
they may start launching missiles at you. It's not deadly, but it *is*
bothersome to deal with. You want to distract them here, and while they
watch that point, quickly place the nano-RFID tags on them before they
realise what happened, and then back off. We don't want to disturb their
natural habitat."

- Oliver


.

Relevant Pages

  • Re: The Future
    ... on" in the human brain which current physical theories cannot explain. ... from functioning as a quantum computer and duplicating it. ... But even if we program the AIs to like humans, ...
    (comp.lang.cobol)
  • Re: Can operant conditioning account for all learning?
    ... considering the genetic complexity of the human brain. ... not as much that humans "created" culturally inheritance as evolution ... goals" is decision theory, based on an economic analysis, known as "expected utility maximisation". ... Rational economic agents consider their possible actions, ...
    (comp.ai.philosophy)
  • Re: Qualia Question
    ... >> working on the reverse engineering problem. ... In the "everything is physical" model, the mind and the brain are the same ... And that's a simple PC designed by humans. ...
    (comp.ai.philosophy)
  • Re: Could anyone explain this please?
    ... their blood - the brain is an oxygen glutton so.... ... HUMANS have A1 and A2. ... in Negroes, rarely in whites, never in yellows. ...
    (sci.anthropology)
  • Re: What are we going to wish for?
    ... customers happy, and in the end, this makes all the individual humans ... But these giant and very smart and very powerful AIs only work for us as ... long as we keep control of them - as long as we keep control of their ... true world government. ...
    (comp.ai.philosophy)