Re: "true" AI Hardware Development

"JGCASEY" <jgkjcasey@xxxxxxxxxxxx> wrote:

We don't really know how much complexity is required to produce
what level of intelligence.

Of course we don't. But we can open a machine and see that it's only got
100 moving parts and get a good rough feel for how many moving parts we
will need to duplicate it's function. And likewise, we can open a brain
and see how many parts it's made of and get a good rough idea of how many
parts we will likely need to duplicate it's function.

What it tells us is more about what won't work than what might work. What
we know is that the computers we had 50 years ago had no hope of
duplicating human intelligence because they had far too few "moving parts"
(memory and speed). What we know now is that our largest machines are now
in the same ball park and in the case of something like the entire
internet, we know it's a many orders of magnitude out of the park - by all
known estimates.

Essentially the brain is a regulator that helps to maintain
the essential variables that define the dynamic system we call
an animal. The brain blocks any environmental disturbances
that may put the animal into a fatal unstable state.

Sure. But we are not trying to build a human brain. We are only trying to
produce it's high level learning ability. Most likely, this is only the
neocortex and a few associated structures needed to make it work.

The gene patterns most likely to survive would have placed a
defense between them and any environmental factors that may
destroy them. This begins at the cellular level but at the
multicellular level the genes coded for passive defenses like
a shell or active defenses like a brain.

Sure, but to create AI, I see no reason to believe we need cellular level
defense systems. The Wright Brothers didn't need to build a modern jet
fighter in order to solve the problem of flight and we sure the hell don't
need to duplicate 90% of the functions of a human body to solve AI. We
just need to build a simple learning machine to act as a generic
controller.

How well the animal's actions protect the animal depends on
having a action to counteract every environmental action that
may bring the animal undone. So the complexity of the
behaviors would have to match the complexity of the forces
acting against the animal.

Getting food, avoiding being food and reproducing are the
name of the game and the foundation of all our behaviors.

Yeah, well I see it as maximizing a reward signal. Evolution spent a lot
of time and effort creating a set of rewards and a learning system that was
optimized for our survival needs. AI has nothing to do with that. AI can
be solved simply by building a machine that's got strong general purpose
learning powers. What we choose for it's motivations is a completely
different problem.

My comments about complexity are just meant to stress the fact
that we have already build machines far more advanced and
complicated than the human brain. We just haven't built the
correct complex machine yet to duplicate human intelligent
behavior.

But they are not "more advanced" in doing the things the
brain does well

True.

and we haven't decided on how to measure how
"complicated" something has to be to duplicate the human
brain.

Yeah, but we know more than enough to make some good guesses. And all our
guesses says that the machines we have already built are already many
orders of magnitude more complicated than the brain. The odds that our
guesses are that much wrong is minuscule.

I suggested it depends on the complexity of the
environment

Learning machines create all their complexity from the environment. This
isn't a suggestion. It's a well known fact. A learning machine with the
power to duplicate human intelligence will look very dumb if put in an
environment where all it can do is push, and hold down, a button to get a
reward. It will stand there holding the button and do almost nothing else.

To make it create the behavior of a typical human, you have to give it a
very rich and complex environment to interact with along with a rich and
complex set of rewards that have no easy answers.

and I would also suggest it is more a matter
of quality than quantity when it comes to intelligence.

Well, what I suspect is that we will find some fairly simple reinforcement
algorithms that of the rough complexity of my networks that have all the
power they need to learn human levels of complex behavior. The only thing
they will need, to outperform humans, is quantity - more speed and memory.

So sure, there is a level of quality we are still missing - or else we
would have solved AI - but I suspect that when we find it, it will be no
more impressive or advanced than some compression programs - probably even
less impressive. What makes it produce human level behavior vs mouse level
behavior will simply be quantity (size and speed) combined with a complex
enough environment (problem) to solve.

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

Quantcast