|
|
 | | From: | wtkiii at hotmail.com | | Subject: | Re: AI will never work in 100 years !!!! | | Date: | 12 Nov 2004 15:29:20 -0800 |
|
|
 | In my view, thoughts are mechanical and emotions are difficult to
explain. I guess, thoughts can be seen as complicated and emotions as
simple. Possibly, no matter how you adjust it, an inexplicable
element remains. An obvious limiting factor is our understanding of
psychology. Thats why I advocate starting from scratch and modeling
single/multi-cell organisms.
Another alternative is the adoration of the pattern matcher. That
means developing a fast/fundamental way of storing related info and
using it to make predictions. Eventually, this system might be equal
to an artificial organism.
|
|
| | From: | BlackWater | | Subject: | Re: AI will never work in 100 years !!!! | | Date: | Wed, 17 Nov 2004 21:54:58 GMT |
|
|
| On 12 Nov 2004 15:29:20 -0800, wtkiii@hotmail.com wrote:
>In my view, thoughts are mechanical and emotions are difficult to
>explain. I guess, thoughts can be seen as complicated and emotions as
>simple. Possibly, no matter how you adjust it, an inexplicable
>element remains. An obvious limiting factor is our understanding of
>psychology. Thats why I advocate starting from scratch and modeling
>single/multi-cell organisms.
I could program-up an "emotional ameoba" crawling
on a cybersurface with others pretty quickly. Indeed,
the rules in many cellular automations - variants of
the "Game of Life" - could be re-phrased as "emotional
responses" to nearby cells ... affecting their behavior.
For an e-organism however you'd want finer, fuzzier,
gradiations which interact in an "emotion space" to
yeild a final synthesis that will increase/decrease
a collection of 'canned' behaviors to varying degrees.
>Another alternative is the adoration of the pattern matcher. That
>means developing a fast/fundamental way of storing related info and
>using it to make predictions. Eventually, this system might be equal
>to an artificial organism.
Organic life absolutely EXCELS at pattern-matching. If
you place a vaguely cow-shaped object in a field, mosquitoes
will fly to it. They can determine "cow" even from a variety
of angles, under a variety of lighting conditions, against
a variety of backgrounds ... even though there's way less
than a cubic millimeter of brain in there (some of which must
be dedicated to other things). So far, attempts to get
e-life to do as well as o-life in this department haven't
been very successful. We kind-of know how - with neural
networks - but the best approach and the computer hardware
stand in our way. Nature had a LONG time to perfect this.
Ordinary recognition tasks - object, sounds, words - are
accomplished very quickly by o-life. Of course, just
turning on a light that says "cow" when you see, hear
or smell it isn't "intelligence". A lot has to be done
with that information afterwards in order to form
thoughts and plans. Pattern-matching is a vital PART
of intelligence, but not intelligence in and of itself.
HOW humans and such recognize patterns/things so quickly
isn't entirely known. My guess is a quick sequence of
fuzzy matches until the subject and the tentative match
share enough points of similarity - then the "Good enough,
yes that's it" flag goes up. Of course o-life doesn't
always get it RIGHT - something which may aid adaptability
if it isn't fatal.
Roboteers have projects in mind where it's very important
to get it exactly right EVERY time, so I wonder how much
of the o-life model they can really make use of. Those who
want to create e-life however may be able to tolerate,
perhaps even covet, such little mistakes. E-Life wouldn't
be flying your airliner or driving your car ... at least
not until it was of near human complexity and capability.
Robots can be stupid, so long as they quickly and cheaply
get their little task exactly right every time.
Oh well, it's a broad, deep subject. So many little parts,
each so complex ... and then we will have to fit them all
together correctly and properly balance their contributions
to the overall e-mind. Get it just a little wrong and we
get what happens when human brain chemistry is just a
little wrong - insanity. It's SO complicated that I doubt
humans will directly accomplish the task. Instead I expect
we'll have to program the basics and then create situations
and/or e-environments where "evolution" can take place. We
may not understand how the final results work any better
than we understand how our own minds work.
|
|
| | From: | wtkiii at hotmail.com | | Subject: | Re: AI will never work in 100 years !!!! | | Date: | 1 Dec 2004 14:53:52 -0800 |
|
|
| After reviewing pattern matching and even human mental behavior, I
think a bottom up approach is the only way.
I'd say about the best project to promote progress in this area is a
set of programs to assist with programming a collection of interacting
cells using C++.
|
|
| | From: | EarlCox | | Subject: | Re: AI will never work in 100 years !!!! | | Date: | Wed, 01 Dec 2004 23:27:10 GMT |
|
|
|
First sentence -- OK, I think many emergent behavior, swarm intelligence,
and cognitive scientists in the AL field would agree with you.
Second sentence -- meaningless technobabble. Not picking a fight, just that
it doesn't say anything (and what it does say is a distillation of the broad
cellular automata approach that has been tried for 30+ years).
wrote in message
news:af26eb2.0412011453.23318864@posting.google.com...
> After reviewing pattern matching and even human mental behavior, I
> think a bottom up approach is the only way.
> I'd say about the best project to promote progress in this area is a
> set of programs to assist with programming a collection of interacting
> cells using C++.
|
|
| | From: | BlackWater | | Subject: | Re: AI will never work in 100 years !!!! | | Date: | Thu, 02 Dec 2004 13:41:54 GMT |
|
|
| Maybe the 'cells' are too simple ?
I can understand the joy in trying to set up an
environment where some kind of e-life can evolve
literally from the 'neuron' level on up ... pure
research ... but that may not be expedient. We
know a little about how brains are set up. Why
not do a little 'seeding' - pre-writing a variety
of fairly simple 'canned' abilities and behaviors
that can actually DO something and then let the
e-nvironment work on evolving useful connections
and alliances between those elements ?
We could wait forever to evolve functional
pattern-matching functions of even the simplest
kind if we start totally from scratch. Even
when they take shape, there's no guarentee
they'll be amenable to whatever evolutionary
approach we're using ... ie they may not have
enough 'handles' to link to other routines or
be too stable (or unstable) to be evolved
further.
Just because 'nature' got it right ONCE doesn't
mean we'd have the same luck. An interesting
experiment - true - but then there exist things
known as 'budgets' and 'deadlines' and the
dreaded "Gee, I'm wasting my time with this"
factor to contend with.
Even a crude simulation of e-life in the form
of e-pets and household robots has proven
commercially profitible. If NOT starting from
scratch yeilds sellable results more quickly
they I'd say it's worth doing. The cash derived
will fund the 'pure' research and attract the
people needed for such research.
A couple of years ago, I picked-up a copy of a
book by Dr. V.I.Ramachandran - a neurologist -
entitled something like "Zombies in the brain".
While it wasn't super-technical, it DID impart
some potentially useful ideas about how minds
operate at the middle level - ideas that might
be translated to e-life.
There are layers upon layers of "little brains",
each more simple and task-specific, inside our
skulls. The sum-total comprises "us". Many of
the upper-level minds are fully intelligent
and muse over different subjects of interest
until needed, whereupon they manifest fully
as the mainstream consciousness. You can easily
experience this effect as you listen to someone
in conversation ... some subject that interests
you will come up, but you can't respond at that
instant, but even as you're listening to the
rest of the persons speech you're also formulating
the point you're going to bring up. It's a
"background task" which can then be brought to
the foreground. A similar experience of "little
brains" can be had if you go to a quiet place
and try to shut-down EVERY thought in your head.
It's not easy at all ... you'll become aware of
"mini-me's" in there practicing bits of conversation,
obsessively going-over a bit of music, analyzing
pictures, making-up stories ... all at the same time,
although just not the mainstream conscious process
at the moment.
Now well below the level of 'conscious' mini-brains
there are well-developed pattern-matchers linked to
the memory store, a database of well-refined
behaviors and such. These ultimately feed-forward
to the more 'conscious' mind(s). Each bit however
is already quite sophisticated - something you
could build/evolve some pretty good e-life out of
without having to worry about 'consciousness' proper.
Such 'functional modules' are the brains middleware
and I think basic ones ARE within our ability to
create whole in software. If we start THERE, with
bits we KNOW work, we KNOW serve useful purposes
and we KNOW can be integrated with others - and then
let e-volution take its course ... we may get a lot
further a lot faster.
I don't have the time or resources to persue this, but
SOME people do ...
On Wed, 01 Dec 2004 23:27:10 GMT, "EarlCox"
wrote:
>
>First sentence -- OK, I think many emergent behavior, swarm intelligence,
>and cognitive scientists in the AL field would agree with you.
>
>Second sentence -- meaningless technobabble. Not picking a fight, just that
>it doesn't say anything (and what it does say is a distillation of the broad
>cellular automata approach that has been tried for 30+ years).
>
>
>
> wrote in message
>news:af26eb2.0412011453.23318864@posting.google.com...
>> After reviewing pattern matching and even human mental behavior, I
>> think a bottom up approach is the only way.
>> I'd say about the best project to promote progress in this area is a
>> set of programs to assist with programming a collection of interacting
>> cells using C++.
>
>
|
|
| | From: | wtkiii at hotmail.com | | Subject: | Re: AI will never work in 100 years !!!! | | Date: | 3 Dec 2004 17:09:59 -0800 |
|
|
| C++ is more complex than it has to be for this job, but an "object"
with some data, some functions, and some pointers to connect it to
other cells may be an adequate model for a cell. Maybe that reduces
the techno-babble. I think that the Alife and automata people have
been spending most of their time with grids of cells that are supposed
to do something interesting if they run awhile. I don't think a group
of cells will evolve itself very successfully. I think the programmer
has to copy nature by learning to program with simple cells that don't
do much by themselves. It is tempting to try to speed things up by
using more sophisticated units, but then you loose the parallel to the
nervous system. It shouldn't take too long to figure out how cells
get a tadpole tail to make swimming motions, etc. Then the layres can
be built up from those simple units. The observation that this would
be a tedious, time consuming project seems correct, but people have
been trying the brilliant, breakthrough method for 53? years. If you
ignore the "impossibility" of the method, it looks like it should
work. I'd like to know who the SOME people are who can afford this
research. I'm still bankrupt from trying to figure out the mind and
program it. Some of the high level activity isn't too bad, but the
support layres have a way of disappearing into unfathomable circuitry.
|
|
| | From: | BlackWater | | Subject: | Re: AI will never work in 100 years !!!! | | Date: | Mon, 06 Dec 2004 23:01:37 GMT |
|
|
| On 3 Dec 2004 17:09:59 -0800, wtkiii@hotmail.com wrote:
>C++ is more complex than it has to be for this job, but an "object"
>with some data, some functions, and some pointers to connect it to
>other cells may be an adequate model for a cell.
C++ and other 'object' languages seem excessive for
a lot of purposes. All that inheritance tends to
add a burden unless the compilers are VERY clever.
Hey, remember when you could just draw a line without
having to do "Aaaa.Bbbb.Cccc.Dddd.Eeee.Ffff.DrawLine()" ?
>Maybe that reduces
>the techno-babble. I think that the Alife and automata people have
>been spending most of their time with grids of cells that are supposed
>to do something interesting if they run awhile. I don't think a group
>of cells will evolve itself very successfully.
They're hoping for 'evolution'. That's fine for 'pure'
science ... but you could diddle and diddle with those
cells and NEVER get anything useful to evolve. I think
they need to start at a somewhat higher level - with
'cells' that actually can DO interesting things - and
then try some more evolution.
>I think the programmer
>has to copy nature by learning to program with simple cells that don't
>do much by themselves. It is tempting to try to speed things up by
>using more sophisticated units, but then you loose the parallel to the
>nervous system.
Hmmm ... not necessarily. Animal nervous systems (and brains)
aren't entirely ad-hoc. They use functional "modules" -
clusters of nerves that are pre-programmed to perform some
specific task. In 'lower' animals we're talking reflexive
behaviors - some of which can be quite sophisticated. In
"smarter" animals we're talking about units within the
brain dedicated to a certain class of problem - and capable
of interacting with similar units to form 'emergent machines'
of even greater complexity.
Going too simple won't get you far - and likewise being
TOO ambitious might mean you spend decaded trying to
perfect one clever 'module'. Fortunately, there's a broad
middle ground to be explored.
>It shouldn't take too long to figure out how cells
>get a tadpole tail to make swimming motions, etc.
We've pretty much got that now.
>Then the layres can
>be built up from those simple units. The observation that this would
>be a tedious, time consuming project seems correct, but people have
>been trying the brilliant, breakthrough method for 53? years.
Fortunately, both the simple units and a selective
environment can be simulated with considerable speed
nowdays. It seems that some of these problems are
amenable to parallel processing methods - opening
the door to using cluster computers or even distributed
computers over the net to run parts of the sim.
>If you
>ignore the "impossibility" of the method, it looks like it should
>work. I'd like to know who the SOME people are who can afford this
>research. I'm still bankrupt from trying to figure out the mind and
>program it. Some of the high level activity isn't too bad, but the
>support layres have a way of disappearing into unfathomable circuitry.
I would love to see a collective EI project similar to
the approach used to refine the Linux operating system.
Hundreds, even thousands, of people and institutions
could assist. So long as common methods and templates
were adhered to it could work. Most people don't have
a LOT of time to spend on EI ... but a lot of people
have a LITTLE free time to persue the subject. Get
them all on the same page.
|
|
|
