Re: Google AI: well at least it's not not evil

Tim Tyler <seemysig@xxxxxxxxxxxxxx> wrote:
Curt Welch wrote:

But, if my believe about intelligence being a reinforcement learning
function at the core is correct, we have the problem that super
intelligence is incompatible with survival. You can't use super
reinforcement learning to create super survival because once they get
smart enough, they will just disable the pain created by the "survival
critic" and OD on pure pleasure. [...]

No way. Becoming a crack addict is incompatible with my
current goals. Therefore, I do not experiment with crack.
The same thing applies to expected utility maximising AIs.

In my opinion you have no real grasp of the true forces at work shaping
your behavior. The goals that we talk about (going to school, having a
career, writing a book, starting a family), are just high level by products
of the true fundamental forces driving us.

If you are in pain, do any of your "goals" matter? No, they don't. All
you can think about is stopping the pain. It's beyond our power to ignore
pain except for short periods. People who are in pain, and can't stop the
pain, kill themselves to stop the pain. We can not resist the fundamental
forces shaping our behavior - it's what we are built to do. We will kill
ourselves if that's the only way to stop pain.

Likewise, if we had a way to create pure pleasure by modifying our reward
system to eliminate all pain and inject us full time with pure pleasure, we
would be unable to resist it as much as we are unable to ignore pure pain.
We would die before turning off the flow of pleasure if had the ability to
turn it on and keep it on.

The reason we don't become crack heads is that crack is not pure pleasure
and it's not ever lasting. Whatever pleasure it creates, it's short lived,
and then, when it's over, life is worse than it was before you took the
crack. Crack and other drugs are not free, you have to find a way to keep
getting more of it, and that problem creates lots of pain in the life of a
drug addict.

But what if we could get a super high for free, anytime we wanted it, and
it would last for as long as we wanted it to last, simply by hooking a PC
to our brain, and modifying a few lines of code, or maybe cutting and
changing a few wires in you brain? What if you could get a never ending
orgasm - one so great, that nothing external would stop the pleasure. You
could cut off your leg and you wouldn't even notice it. That's what it
would be like if we could modify our brain's built in reward system. And
once you try it, you will be instantly addicted.

So, what happens when an AI learns all about how their brain works, and
finds that in the brain, is this reward system which is their pleasure
system. They learn all about how they work and figure out exactly what
they would have to do to modify their pleasure system to disable all pain
signals and give themselves a never ending pain free orgasm. All they have
to do is make some trivial modifications to their brain to try this to see
what it's like.

But, being well educated, they know what will happen if they do this. They
know that when robots do this, they become catatonic and just lay around
moaning. And they know that when someone tries to fix the modification,
and turn off the orgasm, the robots then act as if they are in constant
pain, and will do anything to get that constant orgasm back, or kill
themselves if they can't. They know that the only way to "fix" a robot
after such an experience, is to reset their brain to what it was before
they tried it.

How many robots do you think could resit ever trying it?

What we are talking about is basically suicide by never ending orgasm. If
humans had a flap in their head that covered a button which anyone could
flip open and push to commit suicide by never ending orgasm, what do you
honestly think would happen to the human population? How long do you think
it would take before everyone was dead by never ending orgasm? Remember
that any pain you had felt in your life - hunger, sorrow, depression, a
sore tooth, could instantly be ended and turned into pure happiness and joy
just by pushing that button. How long could you last without pushing it.
How long could you last when 90% of the human population has pushed it and
they all died extremely happy as a result?

This is the nature of the problem we are looking at when AIs get so smart,
that they can 1) understand the nature of the hidden never ending orgasm
button in them, and 2) figure out a way to push it, no matter how well they
are engineered to make it impossible to push.

Disable the motivation hardware?!?

In humans, that would be like taking crack!

No, it's like hitting the never ending orgasm button - which lukely, we
haven't yet figured out how to do without the help of a brain surgeon (I'm
not even sure if they know how to do it yet either).

In an AI, that would /severely/ impact the ability to reach
the current goals in the future. *Huge* negative utility!

You don't understand reinforcement learning machines. They have one and
only one prime goal - to maximize reward. All other goals, like "finishing
my homework so I can graduate from college" are sub-goals of the one prime
goal. The sub-goals are only a means to the end - a means to the getting a
little more total long term reward. If you could bypass all that, and skip
right to the main goal, you would.

If someone left a million dollars sitting on our door step, and and said we
could have the money, how many people would leave it sitting on the step
instead of taking it? Not many, if anyone. Money is just the external
token we use as a placeholder for our internal reward signals. Resisting
hitting that free instant orgasm button will be harder to resit than the
money, because money is a secondary goal to the real goal which is built
into us for creating higher rewards.

AIs will *protect* and *preserve* their goal systems - more
than almost anything else in the world.

As I said, you don't understand reinforcement learning systems (or you
don't believe, or understand, that we are a reinforcement learning system).
We do not protect our goals. Our behavior selection system is HARD-WIRED
to select behaviors based on the amount of reward they are expected to
produce. We have no ability to do anything other than what our behavior
selection system is hard wired to make us do. If the low level behavior
selection systems "figures out" that there is an option to push a button
and get an instant and never ending incrase in reward, it will select that
behavior. "we" don't have anything to say about that.

The only way to prevent it from happening, is to not let the low level
hardware "find out" about the option. If we never try a drug, then the low
level hardware doesn't have a direct experience of the pleasure, and as a
result, it doesn't "known" how good the behavior would be.

A "dumb" AI, might never be able to calculate the reward of the behavior
(see the value in it), and as a result it will never try it as long as
things are set up so it can't do it by accident. And as simple way to
block it from being done by accident is to make it painful if you got close
to doing it - (touching the flap that covers the button creates intense
pain for example).

But a super intelligence, - one which has access to all blue prints and
software specifications for itself, and who fully understands what it is (a
reinforcement learning machine), will be able to predict that pushing that
button is a direct and instant trip to heaven which ends up creating a path
which guides the low level hardware directly to the "push the button"
behavior.

The only way to prevent it, is to constantly brain-wash the AI into not
pushing it (but who is going to do that job?) (or simply never let it learn
about itself).

You badly need to sit through the Stephen Omohundro lectures
I mentioned, if you seriously entertain this sort of stuff:

Sounds interesting. I'll try to find the time to look at time.

The Nature of Self-Improving Artificial Intelligence
http://www.singinst.org/media/singularitysummit2007/stephenomohundro?t=qt
_l

Stanford Colloquium: Self-Improving Artificial Intelligence, Stephen
Omohundro http://video.google.com/videoplay?docid=635444280686112089

Self-Improving AI: Designing 2030
http://video.google.com/videoplay?docid=-3008925388275783572

Else read these two papers from his site:

Stephen M. Omohundro, ?The Nature of Self-Improving Artificial
Intelligence?

Stephen M. Omohundro, ?The Basic AI Drives?

http://selfawaresystems.com/

I don't see how socialism is likley as an alternative.

We already have it. It's not an alternative. How much of our daily
lives are already taken care of by our governments? Roads, police,
fire stations, public parks, street lights, education, regulation of
all industry that could hurt us, various aspects of health care,
unemployment, homeless shelters, social security, welfare. A huge
percentage of the GNP is already a socialistic system. It will just
continue to expand over time until we are all living on welfare. But
instead of people living off the hard work of other people (which is
what welfare is all about today), we will all be living off the hard
work of the machines because there will be no jobs left for humans to
do that machines won't be able to do 10 times better.

Right. But how long is that kind of situation going to last for?
What utilitarian reason is there to keep the humans around?

Well, I think your view comes from a lack of understanding AI. You are
assuming any AI we build will have similar goals and motivations humans
have. You are assuming they will be survival machines - only much better
at it than we are. I don't think that's the case. I think we can get the
advantage of smart machines, without making them survival machines. I
think we will create machines which will love us - who will hate not being
around humans, and hate not taking care of humans. They will be motivated
not to take care of themselves, but to take care of humans instead. They
will take care of themselves, only so they can meet thieir prime goal of
taking care of humans.

The utilitarian reason they will keep us around, is that we will be as
important to them, as food is to us. Not having humans to take care of
will be painful to them. That's the type of AIs we will build. We won't
build AIs that will want to kill us.

After a while, companies, farms, etc will want to expand over
the human nests - and will be prepared to pay for access to the
land. They will probably regard humans as being like vermin:
stupid, primitive messy creatures that use up resources,
leave trails of slime behind them and contribute nothing
to the rest of society, except, of course for their acts
of terrorism and violence.

:)

Only if you assume "intelligence" means "survival machine", instead of
"reinforcement learning machine".

This is the sort of issue that comes from trying to debate the future of AI
when we don't even know what "intelligence" is yet. My view of what
intelligence actually is creates a very different future from your view.

The main reason we rose to dominance on the planet in the
first place, is our large brains. I figure, once we lose
that edge, our days are numbered.

That's true. But because evolution didn't build the AIs directly, we did,
they will be our bitches, not evolutions bitch. Unless we are stupid, we
won't let them get out of control, and I don't think we are stupid.

Either we transform to a socialistic society where we share wealth by
sharing ownership of the major assets of the world, or a small number
of super rich people take over the world.

Or the AIs - as you originally mentioned.

[snip context]

I see no limit to technological progress.

Well, there's physical law. Presumably progress will slow down
after the lifesphere has been expanding at close to the speed
of light for a while ;-)

I guess there's also the possibility that technology has a ceiling as well.
Meaning, there's probably a limited amount of stuff that can be done in
this universe, and when there's nothing more to learn (no more new physics
to understand that creates a new industry), and no more new problems to
solve, and no more new algorithms or new type of machines to design (we
have done it all), the only thing left to do will be to expand and explore
until the entire universe is filled up with energy sucking life forms.

The question is whether we
remain in control of the technology or if the technology takes over.
Because I don't believe that "intelligence" is the same thing "survival
machine" I don't think the technology will take over - at least not in
the short term.

Any machine takeover would not really be motivated by machine
survival instincts - or by machine self-reproduction. It would
simply be down to resource competition with humans. Machines
share our ecological niche, and compete with us for minerals,
land and energy.

Our resource consumption behavior is not hard-wired into us. It's a
behavior created as a secondary goal - because the low level hardware has
predict resource consumption will create high long term rewards.

But what are those rewards? It all goes back to the prime motivations. We
are survival machine not because our prime motivations is "survive". We
are survival machines because our prime motivations are "don't let anything
damage us" (all forms of damage is detected by sensors that are wired to
create pain in us), "dont run out of energy" (the pain of hunger),
"reproduce" (the joy of sex and other related activites), take care of the
kids until they learn enough to take care of themselves (the joy of
nurturing). It only takes a small handful of core motivations like that to
create a survival machine. Everything else, such as the consumption of
resources, can be learned.

What what will we create as the prime motivations for the AIs we build? If
we give them the same sort of motivations we have, they will be survival
machines. We will only be able to get them to work for us if we do things
that meet their basic needs - like feed them. But if they are better at
feeding themselves than we are at feeding them, they won't need us, and in
the end, they will just kill us to get us out of the way.

But, because we don't want to die, we won't build AIs with motivations like
that. With enough clever engineering, we can build them to be motivated to
love humans - to love everything about us. To love doing whatever it is we
ask them to do. As such, they will keep us around, and work for us, simply
because we are human, simply because we are the things they need the most.

Of course machines /will/ most likely be built with survival
instincts - as explained in Stephen M. Omohundro, ?The Basic AI Drives?.

They will need a few very basic things - harm to their body should be seen
as a minor pain to help them learn not to harm their body. And getting
energy should be a minor reward to help them learn to keep their batteries
charged or whatever needs to be done. But the prime motivation of serving
humans should always be configured in them to be more important than the
survival drives like harming their body or keeping their batteries charged.
If a human asks them to jump off a cliff and kill themselves, they should
want to do it - just like what would happen if we asked a TV to do the same
by using the "push" command :) on it.

If you motivate them correctly by giving the right prime motivations, they
will be as responsive to us as our TVs our response to us - only far more
intelligent. They will try to survive, mostly because we have made it
clear to them we want them to survive because we like them being around.

We simply won't build any survival machines that will threaten
our control. We might use our machines to fight other humans (and
their machines) for control, but we won't build the desire to survive
into the machines.

As Stephen mentions, this /is/ one thing that they will get naturally.

Well, there are the forces of evolution fighting our desires. If someone,
somewhere, ends up building a survival machine with the power to self
reproduce, it could survive and self reproduce and grow into a threat for
humans before we realize what was happening. If the forces of evolution
gets it's hands on our machines, we could loose.

However, there's the other limit I've talked about - the natural limit to
inteligence which I'll call the death by orgasam limit. Human intelligence
and human knowledge is already knocking on the DBO door. We are real close
to it ourselves already.

Humans for the most part have no clue what they are. They don't understand
they are just reinforcement learning machines. Behaviorism actually
figured this out 50 years ago, but it turns out people disliked the idea so
much, they decided to reject the discovery in favor of irrational beliefs
that humans were more (better) than that. But society is not going to be
able to hold back the truth much longer.

When we build reinforcement learning machines that equal the intelligence
and consciousness of humans, society is not going to be able to continue
with their head in the sand approach. When these machines are all around
us, and how they work are well documented in every high school text book,
society as a whole is not going to be able to continue to believe we are
anything other than biological versions of the same thing.

What will happen to human society when they actually understand what they
are?

What will they do once they understand that all our goals, and motivations,
and desires, are nothing more the result of the way we are programmed? And
that there really is no meaning to life, and no higher purpose than the
purpose we were built by evolution to perform - which is maximize our
happiness (maximize our long term rewards). That there is no higher
meaning to right and wrong other than the right and wrongs of us maximizing
our happiness? And add to that that one day, we will know exactly how to
directly wire a DBO button into us so we can do what we were built to do -
maximize our happiness by using it.

I have a hard time knowing the answer to all this. But it will bring the
DBO button very close to our reach.

Maybe that is the answer to what we don't see any intelligent life forms
from other plants visiting us? By the time they get smart enough to travel
here, they figure out what they are, and die by orgasm - creating a natural
limit to the intelligence of life which limits how far they are able to
explore?

And anyway, we would build in survival instincts, /even/ if the
machines lacked them. Machines are expensive. If they die,
they cost money to replace.

I also think that building a survival machine is no easy task. Making
it intelligent, and giving it the tools needed to reproduce, isn't all
there is to the problem. You have to design it to want to reproduce
and survive.

Reproduction is a different issue - but reproduction and growth are
also part of the basic AI's drives. It's not a case of building them
in. It's a case of engineering them out, if we do not want them.

I don't agree. But again, this is because you see intelligence as
something more than just a reinforcement learning machine. You don't just
see intelligence as the power to learn in response to a prime reward
signal, you see it as machines with motivations very similar to humans and
with an unlimited ability to expand their intelligence.

These are the questions that will just have to be put on hold until we
create intelligence so we can know what it's true powers are, and aren't.

And you have to make sure the hardware which gives it the desire to
survive never breaks, and is never disabled by the intelligence. One
it figures out it can have unlimited joy and happiness without any
work, the machine will stop surviving.

I christen this idea "Curt's AI masturbation fantasy" ;-)

I like DBO better. :)

You understand that this test has already been done on rats don't you? If
you wire a button so it stimulates their pleasure center, the rats just
push the button until they die. You might believe that our higher level of
intelligence would prevent us from pushing the button, (just like it
prevents most of us from becoming addicted to drugs), but I don't buy it.
None of our drugs act as never end free orgasms yet.

Look at what growing human inteligence is doing to this already. We
have learned to have sex, without reproducing (birth control). This is
what we do to bypass our innate systems that motivate us to survive.
We kill babies when we don't want them - another tool for bypassing our
survival instincts. We overeat to the point of killing ourselves -
more examples of how we bypass the systems intended to make us survive.
And if course, drugs - we take them to make ourselves feel better -
more ways we are already bypassing our innate motivation systems just
so we can be happy. The more we learn about the brain the more we will
find ways to bypass the systems evolution built that cause us to be
survival machines, and the more we will learn how to get happiness
without doing what it takes to survive - without reproducing.

I hope you realise that not everyone does these things.

We have been endowed with brains to help us deal with unfamiliar
environments, while preserving the goals nature gave us.

Some people use them to that end.

:)

We don't maximize current rewards. We are built to maximize long-term
goals. This is standard for all powerful reinforcement learning
algorithms. Because of this, we have the power to trade off short term
loss, for longer term gains. We can do things we don't like in order to
get a higher reward in the future (we work for someone doing stuff we don't
like, so we can get paid, which gives us food, which prevents money which
gives us long term rewards which are greater than the lesser evil we had to
face to get that reward.

This ability to "push through the pain" is limited by the long term reward
we expect to get after the pain.

Don't mistake your power to set long term goals and "push through the pain"
with a power to ignore your prime motivations. You do that stuff _only_
because your brain has (in effect) predicted higher future returns.

One your brain "realizes" there's higher future return in pushing that
button, you are doomed.

Could this lead to not just a decline in population, but to the end of
us as a species in the long time? Is too much intelligence and too
much knowledge a danger to the survival of the species? I wonder if it
might be. Natural selection will always tend to balance it by making
us dumber if that is the case, but human knowledge doesn't die - it
lives on in books an on the internet so once the culture reaches a
given level of knowledge, it might be beyond what natural selection can
"fix" unless natural selection learns how burn books as well.

I am not very impressed by your negative view of intelligence :-(

In my experience, those who would create intelligence, often value
it more highly.

:)

But who is right? Just because you want it to be "better" doesn't mean it
is. :) we will just have to continue to push forward on the work, and see
what we find.

Another possibility of how a high intelligence society will survive is by
effects that happen at the level of the society working with the general
forces of evolution.

As long as there is some subset of the population willing to stand up and
say "God doesn't want us to masturbate - burn the AI books - we are not
robots - we have a mind of our own and free will" - and act as that
brain-washing force to keep other people in line, there is the chance that
segment of the society will survive using a head-in-the-sand survival
technique which natural selection will reward.

But I think that will only work as long as people, and more important,
society as a whole, isn't too intelligent or knowledgeable. So in the end,
I think this effect will act as a natural limit to how intelligent and
knowledgeable any society can become. I think once the society really
understands what they are, life will break down - leaving only the
societies that don't understand what they are still functioning.

--
Curt Welch http://CurtWelch.Com/
curt@xxxxxxxx http://NewsReader.Com/
.

Loading