Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > comp.programming > #553 > unrolled thread

Re: The Lisp Curse

Back to article view | Back to comp.programming

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.

Contents

  Re: The Lisp Curse Mentifex <mentifex@myuw.net> - 2011-06-29 15:41 -0700
    Re: The Lisp Curse "Fuschia, President-Elect of the Bright Purplish-Green Council" <fp-eotbp-gc@ibm.com> - 2011-06-29 19:16 -0400
    Re: The Lisp Curse "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-06-30 09:47 +0200
      Re: The Lisp Curse BGB <cr88192@hotmail.com> - 2011-06-30 02:27 -0700
        Re: The Lisp Curse "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-06-30 14:28 +0200
          Re: The Lisp Curse BGB <cr88192@hotmail.com> - 2011-06-30 14:06 -0700
            Re: The Lisp Curse "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-07-01 23:25 +0200
              Re: The Lisp Curse "Alf P. Steinbach /Usenet" <alf.p.steinbach+usenet@gmail.com> - 2011-07-02 00:32 +0200
                Re: The Lisp Curse "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-07-02 09:49 +0200
                  Re: The Lisp Curse "Alf P. Steinbach /Usenet" <alf.p.steinbach+usenet@gmail.com> - 2011-07-02 12:23 +0200
                    Re: The Lisp Curse "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-07-02 15:30 +0200
                      Re: The Lisp Curse "Alf P. Steinbach /Usenet" <alf.p.steinbach+usenet@gmail.com> - 2011-07-02 17:06 +0200
                        Re: The Lisp Curse "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-07-02 18:09 +0200
                          Re: The Lisp Curse James Waldby <not@valid.invalid> - 2011-07-02 17:23 +0000
                            Re: The Lisp Curse "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-07-02 20:50 +0200
                            Re: The Lisp Curse luser droog <mijoryx@yahoo.com> - 2011-07-17 22:14 -0700
                              Re: The Lisp Curse Ben Bacarisse <ben.usenet@bsb.me.uk> - 2011-07-18 14:10 +0100
                                Re: The Lisp Curse Willem <willem@toad.stack.nl> - 2011-07-18 14:13 +0000
                                  Re: The Lisp Curse Ben Bacarisse <ben.usenet@bsb.me.uk> - 2011-07-18 16:18 +0100
                          Re: The Lisp Curse "Alf P. Steinbach /Usenet" <alf.p.steinbach+usenet@gmail.com> - 2011-07-02 19:47 +0200
                            Re: The Lisp Curse "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-07-02 20:56 +0200
                              Re: The Lisp Curse "Alf P. Steinbach /Usenet" <alf.p.steinbach+usenet@gmail.com> - 2011-07-02 23:01 +0200
                                Re: The Lisp Curse "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-07-03 10:45 +0200
                        Re: The Lisp Curse BGB <cr88192@hotmail.com> - 2011-07-02 10:00 -0700
                      Re: The Lisp Curse luser droog <mijoryx@yahoo.com> - 2011-07-17 04:09 -0700
                  Re: The Lisp Curse "James" <no@spam.invalid> - 2011-07-03 12:06 -0700
              Re: The Lisp Curse BGB <cr88192@hotmail.com> - 2011-07-01 21:22 -0700
                Re: The Lisp Curse "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-07-02 09:47 +0200
                  Re: The Lisp Curse BGB <cr88192@hotmail.com> - 2011-07-02 02:10 -0700
                    Re: The Lisp Curse "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-07-02 15:29 +0200
                      Re: The Lisp Curse BGB <cr88192@hotmail.com> - 2011-07-02 10:03 -0700
                        Re: The Lisp Curse "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-07-02 20:47 +0200
                          Re: The Lisp Curse BGB <cr88192@hotmail.com> - 2011-07-02 20:29 -0700
                            Re: The Lisp Curse "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-07-03 10:53 +0200
                  Re: The Lisp Curse tm <thomas.mertes@gmx.at> - 2011-07-03 23:42 -0700
          Re: The Lisp Curse Fritz Wuehler <fritz@spamexpire-201107.rodent.frell.theremailer.net> - 2011-07-03 14:40 +0200

Page 2 of 2 — ← Prev page 1 [2]

#588

On Sat, 02 Jul 2011 19:47:31 +0200, Alf P. Steinbach /Usenet wrote:

> * Dmitry A. Kazakov, on 02.07.2011 18:09:
>> On Sat, 02 Jul 2011 17:06:29 +0200, Alf P. Steinbach /Usenet wrote:
>>
>>> * Dmitry A. Kazakov, on 02.07.2011 15:30:
>>
>>>> I meant that so long nobody knows what the intelligence algorithmically is.
>>>> In this context estimations of necessary power are meaningless. It is like
>>>> an alchemist guessing how much wood he would need to heat the cauldron in
>>>> order to produce one ounce of gold.
>>>
>>> Oh, I wouldn't have guessed that meaning, it's too meaningless to me, sorry.
>>>
>>> Probably, as seen from my perspective, the insight that you're lacking is that
>>> many aspects of intelligence are well understood, and that one such aspect is
>>> the need for a large, detailed model of the world (where "world" means whatever
>>> environment that the intelligence interacts with), a model that supports almost
>>> instant association actions and (possibly vague) set intersection.
>>
>> I don't think that a model of the world is an essential part:
>>
>> 1. Insects and other primitive species certainly do not have it, yet solve
>> problems too complex for programmed computers. They also possess very
>> limited abilities to learn, so even learning is not a key aspect in
>> achieving the level of intelligence they show.
> 
> Regarding "insects ... do not have [a model of the world]". They do relate to 
> the world, mostly with hardwired sequences of responses. You can *define* the 
> model as the system that provides the responses. Substitute some other model and 
> you get responses that won't work very well for the creature -> insect dies. 
> Thus the world model that is there is necessarily very finely adapted.

So is a plant or a beer bottle. Anything existing in the world is in some
relation to it. I think that a model of the world in its traditional
understanding is a subsystem keeping a large amount of data, which can be
identified and separated from the rest of the system. It should have some
interface and some means to adjust, refine, expand the information stored
there. I believe that insects have nothing in that sense. (An ant hill as a
whole may have something like that if we consider chemical traces left by
the ants etc.)

> Regarding "yet solve problems too complex for programmed computers", if you're 
> talking about control of the body, then that's right, but then it isn't about 
> intelligence but rather about feedback and control systems.

Yes, it is an example of a problem, which computational complexity cannot
be estimated. Intelligence is another problem of such kind, presumably even
more difficult.

My point is that because we cannot estimate complexity of these problems,
we cannot say anything about the hardware needed to solve them.

> Regarding "even learning is not a key aspect in achieving the level of 
> intelligence they show", well, what level intelligence do insects show? Not very 
> much intelligence. By any measure of intelligence an individual insect has about 
> zero, about the same as a thermostat. Give an insect an IQ test and it won't 
> even get the first question right. Zero correct answers => 0 IQ. :-)

BTW, IQ cannot be a measure of intelligence so long we have no idea how
intelligence works or a reliable measure of (to prove that IQ is correlated
with this measure).
 
> Getting an artifical system to emulate an insect brain is still worthwhile.

Sure

> For if one can't do even that, then it's rather hopeless to aim for AI, and 
> secondly, as we have discussed, as opposed to what goes on in a human brain the 
> dataflow per second inside an insect brain appears to be somewhat within the 
> capabilities of current computers (even a PC).

Yes

>> 2. A newly born child does not have any model. The model is constructed
>> later in the process of learning.
> 
> Your idea here is, I believe, one extreme position in the so called "nurture 
> versus nature" debate.
> 
> The nature versus nurture debate is so old that is has been resolved.
> 
> Summary: it's both, both innate capabilities (model) and learning.

We seemingly have different notions of model. To me innate capabilities is
not a model.

-- 
Regards,
Dmitry A. Kazakov
http://www.dmitry-kazakov.de

[toc] | [prev] | [next] | [standalone]

#590

* Dmitry A. Kazakov, on 02.07.2011 20:56:
>
> We seemingly have different notions of model. To me innate capabilities is
> not a model.

You're right that (let me paraphrase you) in order to match external reality, a 
model must somehow support updating. Models that are hardwired in insects and 
other beings, including us, may or may not support dynamic updating within the 
life of an individual. The main update for a hardwired model, however, is not 
during an individual lifetime, as a certain LaMarck-something erroneously 
thought, but through natural selection: those individuals with more fit 
hardwired models are more likely to have offspring, propagating good models.

I'm not sure what you're thinking of as a model, but I think part of the 
difficulty here is to recognize e.g. a simple reaction as a model. Say, some 
critter has a tendency to hang onto a thin branch of a tree, but upon sensing 
some trembling and increased CO2 level it releases its grip and falls, hopefully 
down onto some passing large animal. One may presume that in this creature's 
nervous system no unnecessary detail has been represented: there's nothing 
directly representing the passing animal, the goodness of its blood supply, the 
tree branch, or anything like that. Only trembling and CO2 and the force with 
which the critter is gripping the tree branch.

Another difficulty lies in the encoding. If you have ever debugged a computer 
program at the machine code level you may be aware that there's extremely little 
direct connection between memory bit values and things modeled by the program. 
Yet it's pretty simple: the model provides the relationship between inputs and 
outputs, it is to some extent influenced by the inputs, and if we send a 
philosophing engineer to study it long enough, he or she will be able to come up 
with some correspondences between things in the model (like, excitatory levels 
of a small group of neurons) and things in the world (like, CO2 level).


Cheers & hth.,

- Alf

-- 
blog at <url: http://alfps.wordpress.com>

[toc] | [prev] | [next] | [standalone]

#599

On Sat, 02 Jul 2011 23:01:50 +0200, Alf P. Steinbach /Usenet wrote:

> * Dmitry A. Kazakov, on 02.07.2011 20:56:
>>
>> We seemingly have different notions of model. To me innate capabilities is
>> not a model.
> 
> You're right that (let me paraphrase you) in order to match external reality, a 
> model must somehow support updating. Models that are hardwired in insects and 
> other beings, including us, may or may not support dynamic updating within the 
> life of an individual. The main update for a hardwired model, however, is not 
> during an individual lifetime, as a certain LaMarck-something erroneously 
> thought, but through natural selection: those individuals with more fit 
> hardwired models are more likely to have offspring, propagating good models.

Well, from the behavior of a system and the inputs you always can reverse
engineer some model assuming some class where to search for the model and
some criteria of choice. This is comparable, in fact just same, to
approximation of a function by a class of, say, linear functions using
least squares. It is possible to do, but it does not mean that the system
indeed employs this model or any model. It is you, who models the system.
You cannot equate your model to the system itself (or the model it
allegedly employs) without further [and very strong] assumptions.

> I'm not sure what you're thinking of as a model, but I think part of the 
> difficulty here is to recognize e.g. a simple reaction as a model. Say, some 
> critter has a tendency to hang onto a thin branch of a tree, but upon sensing 
> some trembling and increased CO2 level it releases its grip and falls, hopefully 
> down onto some passing large animal. One may presume that in this creature's 
> nervous system no unnecessary detail has been represented: there's nothing 
> directly representing the passing animal, the goodness of its blood supply, the 
> tree branch, or anything like that. Only trembling and CO2 and the force with 
> which the critter is gripping the tree branch.
> 
> Another difficulty lies in the encoding. If you have ever debugged a computer 
> program at the machine code level you may be aware that there's extremely little 
> direct connection between memory bit values and things modeled by the program.

Yes, it is almost impossible to reconstruct the higher-level language
semantics from machine code and plainly impossible to reconstruct the
domain-specific semantics of the program.

> Yet it's pretty simple: the model provides the relationship between inputs and 
> outputs, it is to some extent influenced by the inputs, and if we send a 
> philosophing engineer to study it long enough, he or she will be able to come up 
> with some correspondences between things in the model (like, excitatory levels 
> of a small group of neurons) and things in the world (like, CO2 level).

Even within such a narrow class there is presumably an infinite number of
designs fulfilling the requirements. But that is not the point. The point
is that x+1 is not a model of some "Linear World". There should be more
meat on it to become a model in a model-based decision making system. I
think at least required:

1. A physical separation from the other parts of the system;

2. Model output is evaluated for fitness rather than directly used (a
logical separation);

3. Modeled activity is asynchronous to the real-time circle where the whole
system is enclosed (which is why models can be so useful).

from #3 we can derive many useful properties as an ability to adjust,
change the model, have competing models within one system etc.

I think that the human brain indeed employ models in the sense #1-3. We can
easily consider ourselves a critter in some role playing game, because we
posses this powerful tool. But I doubt that critters do.

-- 
Regards,
Dmitry A. Kazakov
http://www.dmitry-kazakov.de

[toc] | [prev] | [next] | [standalone]

#582

On 7/2/2011 8:06 AM, Alf P. Steinbach /Usenet wrote:
> * Dmitry A. Kazakov, on 02.07.2011 15:30:
>> On Sat, 02 Jul 2011 12:23:35 +0200, Alf P. Steinbach /Usenet wrote:
>>
>>> * Dmitry A. Kazakov, on 02.07.2011 09:49:
>>>> On Sat, 02 Jul 2011 00:32:56 +0200, Alf P. Steinbach /Usenet wrote:
>>>>
>>>>> On the other hand, computing power is necessary for intelligence,
>>>>> because it's
>>>>> necessary for quickly performing large associative searches (like set
>>>>> intersections) in an extremely large model of the external world.
>>>>>
>>>>> Just to get an idea of what that means, put down a hypothesis of
>>>>> each neuron
>>>>> communicating 100 bytes per second on average. I dunno, perhaps
>>>>> it's just 1, it
>>>>> doesn't matter much. Assume, like, 1 trillion neurons in your
>>>>> brain. Then,
>>>>> what's going on there, is like 100 trillion bytes, or 10^11 bytes,
>>>>> being sent
>>>>> around per second. So for your PC you need a bus with that
>>>>> capacity, but as of
>>>>> 2011 your PC's bus doesn't shuffle hundreds of trillions of bytes
>>>>> per second;
>>>>> it's more like just hundreds of millions... That's the VON NEUMANN
>>>>> BOTTLENECK,
>>>>> the result of extreme centralization: currently a factor of 1
>>>>> million short.
>>>>>
>>>>> And that translates to at least some years, probably a decade or two.
>>>>
>>>> I would not be so sure.
>>>
>>> As I read this you mean "I would not be so sure that AI will be
>>> achieved in a
>>> decade or two".
>>>
>>> If that's what you mean then that's profoundly silly, arguing against
>>> an idiot's
>>> position that nobody's argued for.
>>>
>>> But if you mean that you're not so sure that it will take that long,
>>> then why?
>>
>> I meant that so long nobody knows what the intelligence
>> algorithmically is.
>> In this context estimations of necessary power are meaningless. It is
>> like
>> an alchemist guessing how much wood he would need to heat the cauldron in
>> order to produce one ounce of gold.
>
> Oh, I wouldn't have guessed that meaning, it's too meaningless to me,
> sorry.
>
> Probably, as seen from my perspective, the insight that you're lacking
> is that many aspects of intelligence are well understood, and that one
> such aspect is the need for a large, detailed model of the world (where
> "world" means whatever environment that the intelligence interacts
> with), a model that supports almost instant association actions and
> (possibly vague) set intersection.
>
> You can compare this to, not alchemist as you mentioned (which has the
> wrong kind of connotations) but -- scientific or engineering-like
> chemists. Your task as a fumbling beginner engineering chemist is to
> produce hydrogen from water. You have no idea how to do that, but even
> though you're pretty clueless about the process you're able to calculate
> that you will need *at least* as many kilograms of water as the number
> of kilograms of hydrogen to produce.
>
> This will under-estimate the amount of water that you need, severely.
>
> But it isn't an estimate that tells you how much is sufficient: it only
> tells you that that amount, at the very least, is required, and its
> valid and it's entirely correct, for you do need at least that much water.
>

yeah, probably...
one also has to do similar to estimate things like memory usage or 
internal complexity, which may then impact whether or not a given 
algorithm will be usable, ...


>
>>> The fly's nervous system reportedly has only millions of neurons, but
>>> still
>>> seems to be "smarter and more complicated than any computer built so
>>> far",
>>> according to Ann-Shynn Chiang of the National Tsing Hua University in
>>> Taiwan,
>>> who did some breaktrough work on mapping a fly brain's neural net,
>>> published
>>> this February[1].
>>
>> [...]
>>
>>> [1]<url: http://www.sciencedaily.com/releases/2010/12/101202124108.htm>
>>

read article:
it is a mystery how soon until one can write a fly-brain emulator.
say, they map the fly-brain to the GPU using either OpenGL Shaders or 
OpenCL or similar.

then one can plug this into a 3D character...


3D simulated realistic fruit fly...


>> Yes, I remember the publication of 2006 the page refers to.
>
> I'm impressed, I don't see any reference to 2006.
>
> The journal reference of the original research:
> [
> Ann-Shyn Chiang, Chih-Yung Lin, Chao-Chun Chuang, Hsiu-Ming Chang,
> Chang-Huain Hsieh, Chang-Wei Yeh, Chi-Tin Shih, Jian-Jheng Wu, Guo-Tzau
> Wang, Yung-Chang Chen et al.
> Three-Dimensional Reconstruction of Brain-wide Wiring Networks in
> Drosophila at Single-Cell Resolution.
> Current Biology, December 2, 2010 DOI: 10.1016/j.cub.2010.11.056
> ]
>
> It says "December 2, 2010", but perhaps based on earlier work that you
> remember?
>
>
>> Stimulating a
>> fly with 100% controlled input and comparing its behavior with the models
>> looks a productive approach.
>
> I agree with your idea for one productive approach, I think that would
> be good and necessary. Another approach, that might be also employed,
> might be to computationally evolve the model based on how the model
> fares in some virtual environment (that might ultimately be replaced
> with robot). This way it could possibly be evolved forward much faster,
> e.g. with many sims in parallel. And perhaps equally important, as
> opposed to the Reality comparision, it might be possible to evolve
> smaller parts of the model independently, at least some part of the way.
> Which might make the difference between publishing in one's own life or
> after. ;-)
>

yeah.

an open possibility would be to write a conventional AI (procedural 
logic) with some parts of its logic driven by genetic programs.

then its overall performance can be evaluated and used in breeding up 
more generations.


I had before developed a non-generational version of a GP evolver, which 
basically produced and evaluated programs as a stream (one generates new 
ones, then later sends them back with an evaluation as to how well they 
did, say a weighted calculation involving seconds-lived, damage dealt, 
confirmed kills, ...).

however, in my tests it seemed less able to readily adapt than the 
generational approach.

another idea is to combine this with a neural net, where the program and 
net are evolved either together or in parallel.


the downside though is that in a 3D game, one doesn't have enough cycles 
to burn on this, as it is hard enough even to keep frame-rates up using 
conventional logic (the renderer, basic physics, and basic AI tasks like 
determining if/when potential enemies are visible, all eat up a fair 
number of clock-cycles).

when one is face-to-face with poor framerates, they may realize a little 
more how limited computers may be at times...


although, I guess one could potentially try it in a special-purpose 
environment, whereby framerate is not the deciding factor (then one 
later does the magic trick of later using the "evolved" logic in the 
production engine, probably transcribed to plain C or similar...).

then one can set up large team-vs-team battles, where they have to 
basically score as many enemy kills as they can and survive as long as 
they can, with a slight bias towards minimizing running time and 
fragment length.

possibly, environmental hazards could exist as well (such as falling in 
lava, maybe turrets or crushers, ...).


or such...

[toc] | [prev] | [next] | [standalone]

#670

On Jul 2, 8:30 am, "Dmitry A. Kazakov" <mail...@dmitry-kazakov.de>
wrote:
> On Sat, 02 Jul 2011 12:23:35 +0200, Alf P. Steinbach /Usenet wrote:
>
>
>
> > * Dmitry A. Kazakov, on 02.07.2011 09:49:
> >> On Sat, 02 Jul 2011 00:32:56 +0200, Alf P. Steinbach /Usenet wrote:
>
> >>> On the other hand, computing power is necessary for intelligence, because it's
> >>> necessary for quickly performing large associative searches (like set
> >>> intersections) in an extremely large model of the external world.
>
> >>> Just to get an idea of what that means, put down a hypothesis of each neuron
> >>> communicating 100 bytes per second on average. I dunno, perhaps it's just 1, it
> >>> doesn't matter much. Assume, like, 1 trillion neurons in your brain. Then,
> >>> what's going on there, is like 100 trillion bytes, or 10^11 bytes, being sent
> >>> around per second. So for your PC you need a bus with that capacity, but as of
> >>> 2011 your PC's bus doesn't shuffle hundreds of trillions of bytes per second;
> >>> it's more like just hundreds of millions... That's the VON NEUMANN BOTTLENECK,
> >>> the result of extreme centralization: currently a factor of 1 million short.
>
> >>> And that translates to at least some years, probably a decade or two.
>
> >> I would not be so sure.
>
> > As I read this you mean "I would not be so sure that AI will be achieved in a
> > decade or two".
>
> > If that's what you mean then that's profoundly silly, arguing against an idiot's
> > position that nobody's argued for.
>
> > But if you mean that you're not so sure that it will take that long, then why?
>
> I meant that so long nobody knows what the intelligence algorithmically is.
> In this context estimations of necessary power are meaningless. It is like
> an alchemist guessing how much wood he would need to heat the cauldron in
> order to produce one ounce of gold.

But one can often put a lower bound on such a figure. (Like how much
heat the rest of the house requires.)

[toc] | [prev] | [next] | [standalone]

#605

"Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> wrote in message 
news:ydtfbfc0h12.10ud2ny33ds0q$.dlg@40tude.net...
> On Sat, 02 Jul 2011 00:32:56 +0200, Alf P. Steinbach /Usenet wrote:
>
>> On the other hand, computing power is necessary for intelligence, because 
>> it's
>> necessary for quickly performing large associative searches (like set
>> intersections) in an extremely large model of the external world.
>>
>> Just to get an idea of what that means, put down a hypothesis of each 
>> neuron
>> communicating 100 bytes per second on average. I dunno, perhaps it's just 
>> 1, it
>> doesn't matter much. Assume, like, 1 trillion neurons in your brain. 
>> Then,
>> what's going on there, is like 100 trillion bytes, or 10^11 bytes, being 
>> sent
>> around per second. So for your PC you need a bus with that capacity, but 
>> as of
>> 2011 your PC's bus doesn't shuffle hundreds of trillions of bytes per 
>> second;
>> it's more like just hundreds of millions... That's the VON NEUMANN 
>> BOTTLENECK,
>> the result of extreme centralization: currently a factor of 1 million 
>> short.
>>
>> And that translates to at least some years, probably a decade or two.
>
> I would not be so sure. I think that an average processor is far more
> powerful (in the above terms) than a fly. (How many neurons has a fly?) So
> far nobody was able to build anything even close to the "intelligence" of
> an insect.

http://www.zdnet.com/blog/emergingtech/exclusive-a-robot-with-a-biological-brain/1009?tag=nl.e550

:^o 

[toc] | [prev] | [next] | [standalone]

#572

On 7/1/2011 2:25 PM, Dmitry A. Kazakov wrote:
> On Thu, 30 Jun 2011 14:06:26 -0700, BGB wrote:
>
>> On 6/30/2011 5:28 AM, Dmitry A. Kazakov wrote:
>>> On Thu, 30 Jun 2011 02:27:42 -0700, BGB wrote:
>>>
>>>> rather, one can make any number of weak-AIs, which mimic various aspects
>>>> of intelligence, or mimic intelligent agents (for example, game AIs,
>>>> which are usually called AIs, but in effect are typically little more
>>>> than some mass of procedural logic and hard-coded behaviors).
>>>
>>> ...and some consider intelligent agents a hoax...
>>
>> well, it is only a "hoax" if one believes they are intended to represent
>> true intelligence.
>
> We don't know. If intelligence is computable there is no obvious reason why
> computation cannot be distributed among thin agents, unless the algorithm
> is violently coupled preventing parallelism. But distribution as such
> contributes nothing to the algorithmic complexity.
>

only really matters for true intelligence...

a weak-AI need neither manifest true intelligence, nor require fancy 
implementation strategies. a weak AI need only be "sufficient" to 
complete a given task requiring emulating some relevant aspect of the 
behaviors of an intelligent agent.


in the case of games, this usually means the ability to notice the 
player, run forwards, shoot, and maybe perform other basic actions 
(usually with neither any sort of reasoning or learning abilities, just 
good old hard-coded logic...).


> Some SF authors believed that intelligence arise by itself from
> computational power. That is rubbish, a not booted Intel Core i7 is no more
> or less intelligent than a not booted Intel 8086. It is the software that
> makes the play.
>

well, there is some amount of hand-waving, for example, in the "Kurzweil 
Voodoo", but I doubt that they are being *that* stupid (presumably they 
believe some sort of software is needed, just hand wave just what sort 
of software this may be...).


>> for example, game-AIs mostly serve the purpose of running around,
>> getting shot, shooting back at the player, potentially providing limited
>> interaction with the player, ...
>>
>> and, in this sense, they work fairly well...
>>
>> and, a "good" AI is one with pathfinding and which manages to do things
>> like take cover and avoid running off ledges to their death (say, the
>> player is across a gap and there is a trench with lava in between, a
>> good AI will not run off the ledge and into the lava...).
>>
>> adding speech-synthesis, and HMM-based dialogue generation (similar to a
>> chat-bot or similar), and make them seem a little more natural, ...
>>
>> some games partly ruin the effect by making the NPCs have very detailed
>> scripted events, but as soon as the event ends, the NPCs resort to
>> "stand-and-stare" statue-like behavior, which kind of ruins their
>> portrayal of intelligence.
>
> Huh, but this actually depends on whether we believe in the validness of
> the Turing's test approach. Does an inability of the intelligent system
> (us) to reveal another system's (AI) stupidity really characterize the
> second system as intelligent? If you say, no, Turing was wrong, then I
> would probably agree.
>

I don't say that the inability to disprove its intelligence makes it 
intelligent by extension, but for weak-AIs, it doesn't matter.

it is "strong-AI" which requires the ability for the machine to possess 
some form of actual intelligence, but for weak-AI, no such requirement 
exists.

from the POV of a game player, they don't really need to know or care 
whether or not the NPCs / "enemy player" / ... possess any sort of true 
intelligence, provided whatever they do possess is sufficient for the 
task at hand.


for example, passing the turing test need not demonstrate that a machine 
is intelligent, rather, it only need demonstrate that basic NLP and 
similar tasks have been mastered to an acceptable degree to pass the test.


>> they work acceptably as adaptive linear filters.
>
> But general intelligence should be from a stronger class of methods than
> that. This is why I consider NN as a hoax, i.e. an intended attempt to
> deceive.
>

I don't believe it possesses general intelligence.

I do believe though that it can be used as an adaptive linear filter, 
and that the algorithm itself works to this degree.

it would then only be a "hoax" if:
A, it can be shown that a neural net is unable to perform the task of 
being an adaptive linear filter;
B, it is presented as having abilities beyond that of being an adaptive 
linear filter, and the results of these abilities are being faked.


>>> Generic algorithms are also known since 50s (W.R. Ashby, if I correctly
>>> remember). Their problem is that the intelligence is not in the random
>>> mutations, but in the way it is decided if the mutation was good. This
>>> cannot not work for the same reason the Darwin's "natural selection"
>>> cannot. So the intelligence, if any, is hidden in the directed mutations.
>>
>> I have had some luck using them for signal-filters, but they are
>> generally unacceptably slow to adapt, and pale in comparison to just
>> using ones' head to come up with a solution (goes much faster and
>> doesn't require large numbers of generations...).
>>
>> they could be useful in some cases of adaptive algorithms though, where
>> the algorithm needs to adapt but for some reason a programmer can't just
>> write out a good solution, and it is acceptable for adaptation to be
>> really slow... personally I haven't come across any scenario like this
>> though.
>
> What about a recent hype: the XP (extreme programming) - a chaotic change
> of the source in a hope to adapt to the customer requirements, which
> haven't even been formulated.
>
> The same problem here, there is no any guaranty that the process converges
> at all. And even if it would, little could be said about the properties at
> the convergence point.
>

dunno...

I have never really been big into "development methodologies", as the 
type of strategy which is best suited to implementing something likely 
depends somewhat on the type of thing being implemented.


>>>> some of my code also contains algorithms that could be loosely termed as
>>>> AI (generally data-driven heuristics-based algorithms), but really in a
>>>> practical sense is nothing terribly special about them (in any real
>>>> sense more than invoking "Google" or an SQL query is "intelligent"...).
>>>
>>> Yes, once the problem is solved it is no more considered AI. Presently AI
>>> is a bin to put anything in, we are not sure how to deal with.
>>
>> well, it depends...
>>
>> game AIs are still called AIs despite being both a generally solved
>> problem, and being fairly obviously not-really-intelligent...
>
> OK
>
>>>> as for "magic" or "silver bullet" algorithms, very likely no such algos
>>>> exist.
>>>
>>> Yes. Example: an intelligent human is unable to calculate sine with double
>>> precision which almost any dumb processor easily can.
>>
>> yep, humans suck at calculating things, and computers suck at doing
>> human-like stuff...
>
> Actually it all was human-stuff before. Computers came later and shown
> themselves better than humans in the areas now considered computer-like.
>

fair enough...

but, people built up lots more that was never really within the realms 
of human activity.


>> very possibly, by the time an AI reached human-like intelligences, it
>> would also have absurdly huge computational and information-retrieval
>> abilities (because at its core it would probably still be a
>> computer...), and could likely not be "fooled" in any human-like manner
>> (the traditional assumption of fooling AIs through word games or logic
>> puzzles or riddles, causing them either to go berserk or freeze up...).
>
> It is also possible that an ability to become fooled is a necessary part of
> general intelligence.
>

but, what if the AI doesn't have general intelligence either?...


what if it is, instead, say:
speech recognition and NLP;
and a contextual menu-like interface.

for example, the AI may seem chatty and friendly enough, and can answer 
questions and similar, but at its core it is a menu-driven system, and 
its "mind" is simply collecting various pieces of information and 
navigating between various pre-made menu-forms?...

or, maybe its "mind" amounts to little more than: forwards, back, abort 
(return to main menu), and gathering a list of facts (IOW, a 
"wizard"-style interface).

good enough NLP and responsiveness may well partly hide from the average 
person the fact that they are just going along through a wizard or a 
menu system, as the AI talks and responds much as if it were an actual 
person (although potentially being immune to the verbal attacks of angry 
customers, or to usual sorts of trickery or social-engineering tactics).

for many tasks, this level of "intelligence" may well be sufficient.


> Anyway, due to Gödel's incompleteness, there always should be a state of
> "perplexity" in any inference system. It does not crash the human brain,
> because we have multiple subsystems monitoring each other as
> semi-independent agents. An AI system should be necessarily built
> similarly.
>

potentially, or humans don't crash because an input validation problem 
or buffer-overflow overwrite part of the control stack and/or lead to an 
invalid memory access (GPF / segfault...).


this is I guess the usual basis of the "break robots with word games" 
strategy.

one goes up to robot store clerk and says "how much wood could a wood 
chuck chuck if a wood chuck couldn't chuck wood?" full well expecting 
its head to explode as it falls backwards, allowing them to run off with 
the money from the register.

but then it may just look at them with its blank robot eyes and say: "I 
don't understand, please restate the question."

[toc] | [prev] | [next] | [standalone]

#573

On Fri, 01 Jul 2011 21:22:33 -0700, BGB wrote:

> a weak-AI need neither manifest true intelligence, nor require fancy 
> implementation strategies. a weak AI need only be "sufficient" to 
> complete a given task requiring emulating some relevant aspect of the 
> behaviors of an intelligent agent.

1. How do you know which aspects are relevant? Relevancy is derived from
the way the truly intelligent system (the player) senses the world. To know
it in a hard way could be more complex than the intelligence itself.

2. How do you know that any relevant aspect is computable?

The relation between "weak-AI" and "strong-AI" is unknown. Turing believed
them comparable, but the relation can be any. (Somebody said, that two
unintelligent persons can produce an intelligent one in just 9 months
without any mental efforts. (:-))

>> Some SF authors believed that intelligence arise by itself from
>> computational power. That is rubbish, a not booted Intel Core i7 is no more
>> or less intelligent than a not booted Intel 8086. It is the software that
>> makes the play.
> 
> well, there is some amount of hand-waving, for example, in the "Kurzweil 
> Voodoo", but I doubt that they are being *that* stupid (presumably they 
> believe some sort of software is needed, just hand wave just what sort 
> of software this may be...).

In 40s-50s they didn't write about computers in modern understanding, they
did about "homeostatic mechanisms". AI was thought as a huge analogue
control system full of feedbacks, made of coils and ebonite. Modern
computers are relay systems of then, which at that time nobody was much
interested in.

> I don't say that the inability to disprove its intelligence makes it 
> intelligent by extension, but for weak-AIs, it doesn't matter.
>
> it is "strong-AI" which requires the ability for the machine to possess 
> some form of actual intelligence, but for weak-AI, no such requirement 
> exists.
> 
> from the POV of a game player, they don't really need to know or care 
> whether or not the NPCs / "enemy player" / ... possess any sort of true 
> intelligence, provided whatever they do possess is sufficient for the 
> task at hand.

You want the NPC to behave as if it were really intelligent. Games replace
this "weak" or, I would call it "Turing", intelligence with straightforward
cheating. Cheating is unintelligent, because it looks so, this is what
makes players so angry...

> for example, passing the turing test need not demonstrate that a machine 
> is intelligent, rather, it only need demonstrate that basic NLP and 
> similar tasks have been mastered to an acceptable degree to pass the test.

Maybe, but that presumes that the intelligence is more than a sum of basic
tasks. Which is an open question so far.

>>> yep, humans suck at calculating things, and computers suck at doing
>>> human-like stuff...
>>
>> Actually it all was human-stuff before. Computers came later and shown
>> themselves better than humans in the areas now considered computer-like.
> 
> fair enough...
> 
> but, people built up lots more that was never really within the realms 
> of human activity.

A key feature of general intelligence, BTW.

> what if it is, instead, say:
> speech recognition and NLP;
> and a contextual menu-like interface.
> 
> for example, the AI may seem chatty and friendly enough, and can answer 
> questions and similar, but at its core it is a menu-driven system, and 
> its "mind" is simply collecting various pieces of information and 
> navigating between various pre-made menu-forms?...

The magic is blown away? But whatever intelligence be, if computable, there
is no magic in it, just many, many tiny cogwheels! (:-))

>> Anyway, due to Gödel's incompleteness, there always should be a state of
>> "perplexity" in any inference system. It does not crash the human brain,
>> because we have multiple subsystems monitoring each other as
>> semi-independent agents. An AI system should be necessarily built
>> similarly.
> 
> potentially, or humans don't crash because an input validation problem 
> or buffer-overflow overwrite part of the control stack and/or lead to an 
> invalid memory access (GPF / segfault...).

Interestingly that living systems process a huge amount of input
information to deal with, quite persistently to DoS attacks...

-- 
Regards,
Dmitry A. Kazakov
http://www.dmitry-kazakov.de

[toc] | [prev] | [next] | [standalone]

#575

On 7/2/2011 12:47 AM, Dmitry A. Kazakov wrote:
> On Fri, 01 Jul 2011 21:22:33 -0700, BGB wrote:
>
>> a weak-AI need neither manifest true intelligence, nor require fancy
>> implementation strategies. a weak AI need only be "sufficient" to
>> complete a given task requiring emulating some relevant aspect of the
>> behaviors of an intelligent agent.
>
> 1. How do you know which aspects are relevant? Relevancy is derived from
> the way the truly intelligent system (the player) senses the world. To know
> it in a hard way could be more complex than the intelligence itself.
>

doesn't really matter in-general though.
the goodness of the AI is like the goodness of the graphics or the 
physics. if it looks good enough, good enough, and maybe someone else 
will improve on it later...


> 2. How do you know that any relevant aspect is computable?
>

this can determined sufficiently well in context, usually by beating at 
the behaviors of something long enough that it is "sufficiently not stupid".

relevant aspects would be, say:
responds to player;
shoots at player;
doesn't run off ledges.

and, for some newer games:
goes and takes cover (vs standing in the open soaking up bullets);
maybe has path-finding, so as to avoid getting stuck in a corner 
somewhere or be defeated by being unable to run around a wall;
...


> The relation between "weak-AI" and "strong-AI" is unknown. Turing believed
> them comparable, but the relation can be any. (Somebody said, that two
> unintelligent persons can produce an intelligent one in just 9 months
> without any mental efforts. (:-))
>

well, there is a difference:
"weak AI" is "good enough for the task at hand" WRT basic behaviors and 
functionality;
"strong AI" possess some form of actual intelligence, and can build on 
this, potentially eventually exceeding its creators.


humans will not be intellectually defeated by plain procedural logic.


>>> Some SF authors believed that intelligence arise by itself from
>>> computational power. That is rubbish, a not booted Intel Core i7 is no more
>>> or less intelligent than a not booted Intel 8086. It is the software that
>>> makes the play.
>>
>> well, there is some amount of hand-waving, for example, in the "Kurzweil
>> Voodoo", but I doubt that they are being *that* stupid (presumably they
>> believe some sort of software is needed, just hand wave just what sort
>> of software this may be...).
>
> In 40s-50s they didn't write about computers in modern understanding, they
> did about "homeostatic mechanisms". AI was thought as a huge analogue
> control system full of feedbacks, made of coils and ebonite. Modern
> computers are relay systems of then, which at that time nobody was much
> interested in.
>

yet digital computers are generally more practical...


>> I don't say that the inability to disprove its intelligence makes it
>> intelligent by extension, but for weak-AIs, it doesn't matter.
>>
>> it is "strong-AI" which requires the ability for the machine to possess
>> some form of actual intelligence, but for weak-AI, no such requirement
>> exists.
>>
>> from the POV of a game player, they don't really need to know or care
>> whether or not the NPCs / "enemy player" / ... possess any sort of true
>> intelligence, provided whatever they do possess is sufficient for the
>> task at hand.
>
> You want the NPC to behave as if it were really intelligent. Games replace
> this "weak" or, I would call it "Turing", intelligence with straightforward
> cheating. Cheating is unintelligent, because it looks so, this is what
> makes players so angry...
>

usually in games, the line is ill-defined. "cheating" in the game-design 
sense is more often reserved for things like when the game pulls a 
forced-lose (where one goes through a battle, which then invokes a 
cutscene showing the player losing, even if in-battle the player 
one-shotted the enemy). or, when the game pulls out some sort of "voodoo 
magic" which makes something happen that otherwise couldn't work in the 
game (usually driven by "event triggers"). basically, the gameplay 
equivalent of the "invisible wall".

most general game AIs basically just hard-code a number of behaviors, 
but they are applied consistently, making them not "cheating" by this 
definition.


>> for example, passing the turing test need not demonstrate that a machine
>> is intelligent, rather, it only need demonstrate that basic NLP and
>> similar tasks have been mastered to an acceptable degree to pass the test.
>
> Maybe, but that presumes that the intelligence is more than a sum of basic
> tasks. Which is an open question so far.
>

well, humans demonstrate some abilities not presently demonstrated by 
computers.

the ability to be conscious/self-aware/... could very well be beyond 
their reach, FWIW.


>>>> yep, humans suck at calculating things, and computers suck at doing
>>>> human-like stuff...
>>>
>>> Actually it all was human-stuff before. Computers came later and shown
>>> themselves better than humans in the areas now considered computer-like.
>>
>> fair enough...
>>
>> but, people built up lots more that was never really within the realms
>> of human activity.
>
> A key feature of general intelligence, BTW.
>

except humans built it, so this is human intelligence, not machine 
intelligence.


>> what if it is, instead, say:
>> speech recognition and NLP;
>> and a contextual menu-like interface.
>>
>> for example, the AI may seem chatty and friendly enough, and can answer
>> questions and similar, but at its core it is a menu-driven system, and
>> its "mind" is simply collecting various pieces of information and
>> navigating between various pre-made menu-forms?...
>
> The magic is blown away? But whatever intelligence be, if computable, there
> is no magic in it, just many, many tiny cogwheels! (:-))
>

no, a menu-driven AI would be notably inferior to a real human, as it 
would not have sentience. however, such a menu-driven AI is within the 
boundaries of being likely implementable.


>>> Anyway, due to Gödel's incompleteness, there always should be a state of
>>> "perplexity" in any inference system. It does not crash the human brain,
>>> because we have multiple subsystems monitoring each other as
>>> semi-independent agents. An AI system should be necessarily built
>>> similarly.
>>
>> potentially, or humans don't crash because an input validation problem
>> or buffer-overflow overwrite part of the control stack and/or lead to an
>> invalid memory access (GPF / segfault...).
>
> Interestingly that living systems process a huge amount of input
> information to deal with, quite persistently to DoS attacks...
>

because the human mind doesn't work the same as a computer...

[toc] | [prev] | [next] | [standalone]

#578

On Sat, 02 Jul 2011 02:10:12 -0700, BGB wrote:

> On 7/2/2011 12:47 AM, Dmitry A. Kazakov wrote:
>> On Fri, 01 Jul 2011 21:22:33 -0700, BGB wrote:
>>
>>> a weak-AI need neither manifest true intelligence, nor require fancy
>>> implementation strategies. a weak AI need only be "sufficient" to
>>> complete a given task requiring emulating some relevant aspect of the
>>> behaviors of an intelligent agent.
>>
>> 1. How do you know which aspects are relevant? Relevancy is derived from
>> the way the truly intelligent system (the player) senses the world. To know
>> it in a hard way could be more complex than the intelligence itself.
> 
> doesn't really matter in-general though.
> the goodness of the AI is like the goodness of the graphics or the 
> physics. if it looks good enough, good enough, and maybe someone else 
> will improve on it later...

No, graphics can be measured (fps, pixels, color depth). AI lacks
quantitative measurement to know if it is good enough and what need to be
done to have it good. Once you had AI you could measure its quality in
"intelligent decisions" per second. But there is no any in any time.

>> 2. How do you know that any relevant aspect is computable?
> 
> this can determined sufficiently well in context, usually by beating at 
> the behaviors of something long enough that it is "sufficiently not stupid".
> 
> relevant aspects would be, say:
> responds to player;
> shoots at player;
> doesn't run off ledges.

These are just actions. Intelligence is in how actions are performed, not
in what is done. The same action can be considered intelligent or
unintelligent depending on the context. Relevant here is the choice of an
action, not the action itself.

>> The relation between "weak-AI" and "strong-AI" is unknown. Turing believed
>> them comparable, but the relation can be any. (Somebody said, that two
>> unintelligent persons can produce an intelligent one in just 9 months
>> without any mental efforts. (:-))
> 
> well, there is a difference:
> "weak AI" is "good enough for the task at hand" WRT basic behaviors and 
> functionality;
> "strong AI" possess some form of actual intelligence, and can build on 
> this, potentially eventually exceeding its creators.

I don't think this is a workable distinction. For example, sweeping floor
with a broom is a task requiring much more general intelligence than
playing chess. (Thirty years ago people believed otherwise) I doubt that an
average office-cleaner exceeds his creators.

>>> what if it is, instead, say:
>>> speech recognition and NLP;
>>> and a contextual menu-like interface.
>>>
>>> for example, the AI may seem chatty and friendly enough, and can answer
>>> questions and similar, but at its core it is a menu-driven system, and
>>> its "mind" is simply collecting various pieces of information and
>>> navigating between various pre-made menu-forms?...
>>
>> The magic is blown away? But whatever intelligence be, if computable, there
>> is no magic in it, just many, many tiny cogwheels! (:-))
> 
> no, a menu-driven AI would be notably inferior to a real human, as it 
> would not have sentience.

It is only a question of reflection, i.e. if you include the states of the
menu in the menu. Fake consciousness should not be harder than fake
decision making, the difference is only in what is considered the input,
the outer word or the inner one.

>>>> Anyway, due to Gödel's incompleteness, there always should be a state of
>>>> "perplexity" in any inference system. It does not crash the human brain,
>>>> because we have multiple subsystems monitoring each other as
>>>> semi-independent agents. An AI system should be necessarily built
>>>> similarly.
>>>
>>> potentially, or humans don't crash because an input validation problem
>>> or buffer-overflow overwrite part of the control stack and/or lead to an
>>> invalid memory access (GPF / segfault...).
>>
>> Interestingly that living systems process a huge amount of input
>> information to deal with, quite persistently to DoS attacks...
> 
> because the human mind doesn't work the same as a computer...

That is just one step away from saying that intelligence is incomputable.

-- 
Regards,
Dmitry A. Kazakov
http://www.dmitry-kazakov.de

[toc] | [prev] | [next] | [standalone]

#583

On 7/2/2011 6:29 AM, Dmitry A. Kazakov wrote:
> On Sat, 02 Jul 2011 02:10:12 -0700, BGB wrote:
>
>> On 7/2/2011 12:47 AM, Dmitry A. Kazakov wrote:
>>> On Fri, 01 Jul 2011 21:22:33 -0700, BGB wrote:
>>>
>>>> a weak-AI need neither manifest true intelligence, nor require fancy
>>>> implementation strategies. a weak AI need only be "sufficient" to
>>>> complete a given task requiring emulating some relevant aspect of the
>>>> behaviors of an intelligent agent.
>>>
>>> 1. How do you know which aspects are relevant? Relevancy is derived from
>>> the way the truly intelligent system (the player) senses the world. To know
>>> it in a hard way could be more complex than the intelligence itself.
>>
>> doesn't really matter in-general though.
>> the goodness of the AI is like the goodness of the graphics or the
>> physics. if it looks good enough, good enough, and maybe someone else
>> will improve on it later...
>
> No, graphics can be measured (fps, pixels, color depth). AI lacks
> quantitative measurement to know if it is good enough and what need to be
> done to have it good. Once you had AI you could measure its quality in
> "intelligent decisions" per second. But there is no any in any time.
>

not really, in non-cases, such as the quality of a 3D rendered scene, 
... it as well depends on subjective properties.

for many tasks, the subjective properties are sufficient, or are 
actually the important side of the thing in question.



>>> 2. How do you know that any relevant aspect is computable?
>>
>> this can determined sufficiently well in context, usually by beating at
>> the behaviors of something long enough that it is "sufficiently not stupid".
>>
>> relevant aspects would be, say:
>> responds to player;
>> shoots at player;
>> doesn't run off ledges.
>
> These are just actions. Intelligence is in how actions are performed, not
> in what is done. The same action can be considered intelligent or
> unintelligent depending on the context. Relevant here is the choice of an
> action, not the action itself.
>

for many of these AIs, the intelligence and behavior are equivalent. 
usually, the logic that drives the behaviors is good old procedural 
logic. then, the only difference between an AI and a traditional piece 
of code, is that people usually have an AI as an anthropomorphic entity, 
rather than sitting off invisibly in the background somewhere.


>>> The relation between "weak-AI" and "strong-AI" is unknown. Turing believed
>>> them comparable, but the relation can be any. (Somebody said, that two
>>> unintelligent persons can produce an intelligent one in just 9 months
>>> without any mental efforts. (:-))
>>
>> well, there is a difference:
>> "weak AI" is "good enough for the task at hand" WRT basic behaviors and
>> functionality;
>> "strong AI" possess some form of actual intelligence, and can build on
>> this, potentially eventually exceeding its creators.
>
> I don't think this is a workable distinction. For example, sweeping floor
> with a broom is a task requiring much more general intelligence than
> playing chess. (Thirty years ago people believed otherwise) I doubt that an
> average office-cleaner exceeds his creators.
>


sweeping a floor need only require a robot that can operate a broom and 
move around, and probably several layers of bitmaps, and some more 
procedural logic.

but, the point is that "weak-AI" is within the reach of non-adaptive 
procedural logic (IOW: hard-coded logic), whereas "strong-AI" is not.


>>>> what if it is, instead, say:
>>>> speech recognition and NLP;
>>>> and a contextual menu-like interface.
>>>>
>>>> for example, the AI may seem chatty and friendly enough, and can answer
>>>> questions and similar, but at its core it is a menu-driven system, and
>>>> its "mind" is simply collecting various pieces of information and
>>>> navigating between various pre-made menu-forms?...
>>>
>>> The magic is blown away? But whatever intelligence be, if computable, there
>>> is no magic in it, just many, many tiny cogwheels! (:-))
>>
>> no, a menu-driven AI would be notably inferior to a real human, as it
>> would not have sentience.
>
> It is only a question of reflection, i.e. if you include the states of the
> menu in the menu. Fake consciousness should not be harder than fake
> decision making, the difference is only in what is considered the input,
> the outer word or the inner one.
>

this may not be sufficient, and depending on the implementation could be 
prone to crashing (say, an infinite recursion takes place and leads to a 
stack overflow).


>>>>> Anyway, due to Gödel's incompleteness, there always should be a state of
>>>>> "perplexity" in any inference system. It does not crash the human brain,
>>>>> because we have multiple subsystems monitoring each other as
>>>>> semi-independent agents. An AI system should be necessarily built
>>>>> similarly.
>>>>
>>>> potentially, or humans don't crash because an input validation problem
>>>> or buffer-overflow overwrite part of the control stack and/or lead to an
>>>> invalid memory access (GPF / segfault...).
>>>
>>> Interestingly that living systems process a huge amount of input
>>> information to deal with, quite persistently to DoS attacks...
>>
>> because the human mind doesn't work the same as a computer...
>
> That is just one step away from saying that intelligence is incomputable.
>

it may well be...

the more immediate concern though is that humans lack the mechanisms by 
which to experience a GPF or stack overflow or buffer overflow.

the brain is designed to try to keep going as long as possible, even 
after critical damage, unlike software which will generally blow-up at 
the first sign of trouble ("oh crap, this pointer is null", boom...).


or such...

[toc] | [prev] | [next] | [standalone]

#586

On Sat, 02 Jul 2011 10:03:57 -0700, BGB wrote:

> On 7/2/2011 6:29 AM, Dmitry A. Kazakov wrote:

>> No, graphics can be measured (fps, pixels, color depth). AI lacks
>> quantitative measurement to know if it is good enough and what need to be
>> done to have it good. Once you had AI you could measure its quality in
>> "intelligent decisions" per second. But there is no any in any time.
> 
> not really, in non-cases, such as the quality of a 3D rendered scene, 
> ... it as well depends on subjective properties.

More fps, higher resolution, etc do it better. (I don't consider cases when
things rendered are not ones one actually wanted to show)

>>>> 2. How do you know that any relevant aspect is computable?
>>>
>>> this can determined sufficiently well in context, usually by beating at
>>> the behaviors of something long enough that it is "sufficiently not stupid".
>>>
>>> relevant aspects would be, say:
>>> responds to player;
>>> shoots at player;
>>> doesn't run off ledges.
>>
>> These are just actions. Intelligence is in how actions are performed, not
>> in what is done. The same action can be considered intelligent or
>> unintelligent depending on the context. Relevant here is the choice of an
>> action, not the action itself.
> 
> for many of these AIs, the intelligence and behavior are equivalent.

Possibly for all. Behavior is a function of intelligence and the only way
intelligence can be observed.

> usually, the logic that drives the behaviors is good old procedural 
> logic. then, the only difference between an AI and a traditional piece 
> of code, is that people usually have an AI as an anthropomorphic entity, 
> rather than sitting off invisibly in the background somewhere.

Yes, any finite number of states can be programmed by a FSM in a
brute-force way. The problem is that we don't want to go this way. A more
intelligent approach is as intelligent as it is.

>>>> The relation between "weak-AI" and "strong-AI" is unknown. Turing believed
>>>> them comparable, but the relation can be any. (Somebody said, that two
>>>> unintelligent persons can produce an intelligent one in just 9 months
>>>> without any mental efforts. (:-))
>>>
>>> well, there is a difference:
>>> "weak AI" is "good enough for the task at hand" WRT basic behaviors and
>>> functionality;
>>> "strong AI" possess some form of actual intelligence, and can build on
>>> this, potentially eventually exceeding its creators.
>>
>> I don't think this is a workable distinction. For example, sweeping floor
>> with a broom is a task requiring much more general intelligence than
>> playing chess. (Thirty years ago people believed otherwise) I doubt that an
>> average office-cleaner exceeds his creators.
> 
> sweeping a floor need only require a robot that can operate a broom and 
> move around, and probably several layers of bitmaps, and some more 
> procedural logic.

It requires classification of things into valuable stuff vs. rubbish/dirt.
No procedural logic would help here.

> but, the point is that "weak-AI" is within the reach of non-adaptive 
> procedural logic (IOW: hard-coded logic), whereas "strong-AI" is not.

Yes, if you define "weak" as "programmable by average programmer", but that
would make the classification useless.

>>>>> what if it is, instead, say:
>>>>> speech recognition and NLP;
>>>>> and a contextual menu-like interface.
>>>>>
>>>>> for example, the AI may seem chatty and friendly enough, and can answer
>>>>> questions and similar, but at its core it is a menu-driven system, and
>>>>> its "mind" is simply collecting various pieces of information and
>>>>> navigating between various pre-made menu-forms?...
>>>>
>>>> The magic is blown away? But whatever intelligence be, if computable, there
>>>> is no magic in it, just many, many tiny cogwheels! (:-))
>>>
>>> no, a menu-driven AI would be notably inferior to a real human, as it
>>> would not have sentience.
>>
>> It is only a question of reflection, i.e. if you include the states of the
>> menu in the menu. Fake consciousness should not be harder than fake
>> decision making, the difference is only in what is considered the input,
>> the outer word or the inner one.
> 
> this may not be sufficient, and depending on the implementation could be 
> prone to crashing (say, an infinite recursion takes place and leads to a 
> stack overflow).

You can prune the search tree, which many algorithms do anyway. If
exceptional states were the only problem...

> the more immediate concern though is that humans lack the mechanisms by 
> which to experience a GPF or stack overflow or buffer overflow.

Human brain is a massively parallel system, if an exception occurs
somewhere in such a system it stays local (unless you try to keep
everything synchronized).

-- 
Regards,
Dmitry A. Kazakov
http://www.dmitry-kazakov.de

[toc] | [prev] | [next] | [standalone]

#591

On 7/2/2011 11:47 AM, Dmitry A. Kazakov wrote:
> On Sat, 02 Jul 2011 10:03:57 -0700, BGB wrote:
>
>> On 7/2/2011 6:29 AM, Dmitry A. Kazakov wrote:
>
>>> No, graphics can be measured (fps, pixels, color depth). AI lacks
>>> quantitative measurement to know if it is good enough and what need to be
>>> done to have it good. Once you had AI you could measure its quality in
>>> "intelligent decisions" per second. But there is no any in any time.
>>
>> not really, in non-cases, such as the quality of a 3D rendered scene,
>> ... it as well depends on subjective properties.
>
> More fps, higher resolution, etc do it better. (I don't consider cases when
> things rendered are not ones one actually wanted to show)
>

past a certain point there is not much of a point (much past 1920x1080 
and 60fps, with the present technology, there is not much to gain).

for example, plain Quake2 in 1920x1080 and pulling 60 fps will still 
look like Quake2...


then one ends up in the land of adding detailed lighting and shadowing 
(phong lighting and stencil shadows or shadow-mapping) and using pixel 
shaders on damn near everything, which can improve aesthetics at a cost 
to the framerate.

a different direction is to have really big scenes, or to have lots of 
complex detail in the geometry. sadly though, it is "take your pick", 
since one can't have a huge scene with complex geometry and lots of 
pixel-shader effects, as then the framerate is teh-suck...

however, LOD can help, an example of this would be falling back to a 
simpler lighting strategy for geometry past a certain distance (say, 
vertex lighting and no shadows, or use of large low-resolution light-maps).



>>>>> 2. How do you know that any relevant aspect is computable?
>>>>
>>>> this can determined sufficiently well in context, usually by beating at
>>>> the behaviors of something long enough that it is "sufficiently not stupid".
>>>>
>>>> relevant aspects would be, say:
>>>> responds to player;
>>>> shoots at player;
>>>> doesn't run off ledges.
>>>
>>> These are just actions. Intelligence is in how actions are performed, not
>>> in what is done. The same action can be considered intelligent or
>>> unintelligent depending on the context. Relevant here is the choice of an
>>> action, not the action itself.
>>
>> for many of these AIs, the intelligence and behavior are equivalent.
>
> Possibly for all. Behavior is a function of intelligence and the only way
> intelligence can be observed.
>

yes, but behavior need not have much actual intelligence behind it.


>> usually, the logic that drives the behaviors is good old procedural
>> logic. then, the only difference between an AI and a traditional piece
>> of code, is that people usually have an AI as an anthropomorphic entity,
>> rather than sitting off invisibly in the background somewhere.
>
> Yes, any finite number of states can be programmed by a FSM in a
> brute-force way. The problem is that we don't want to go this way. A more
> intelligent approach is as intelligent as it is.
>

but, it is another tradeoff:
brute force works now;
if the number of possible states is small enough to be within what the 
developer can code for, then there is little problem with hard-coding 
the entire set of states.

for example, if an NPC has, say, 10 general states (stand, walk, run, 
dead, melee attack, ranged attack, ...), and a limited number of 
possible state transitions, then little stops someone from hard-coding 
it. then other more subtle variations can be controlled by flags 
(hostile/passive/player-aligned/...).



>>>>> The relation between "weak-AI" and "strong-AI" is unknown. Turing believed
>>>>> them comparable, but the relation can be any. (Somebody said, that two
>>>>> unintelligent persons can produce an intelligent one in just 9 months
>>>>> without any mental efforts. (:-))
>>>>
>>>> well, there is a difference:
>>>> "weak AI" is "good enough for the task at hand" WRT basic behaviors and
>>>> functionality;
>>>> "strong AI" possess some form of actual intelligence, and can build on
>>>> this, potentially eventually exceeding its creators.
>>>
>>> I don't think this is a workable distinction. For example, sweeping floor
>>> with a broom is a task requiring much more general intelligence than
>>> playing chess. (Thirty years ago people believed otherwise) I doubt that an
>>> average office-cleaner exceeds his creators.
>>
>> sweeping a floor need only require a robot that can operate a broom and
>> move around, and probably several layers of bitmaps, and some more
>> procedural logic.
>
> It requires classification of things into valuable stuff vs. rubbish/dirt.
> No procedural logic would help here.
>

a sweeping robot need not have valuation, rather it would just sweep up 
anything on the floor just the same. if a person worried about it, they 
could create a mechanism to separate dust from non-dust and have someone 
manually evaluate the non-dust later (say, robot has a container for any 
pieces of clothing/shoes/wallets/... it runs across)

or, it could use basic computer-vision algorithms to classify common 
object types (say, an object is encountered, and then a long list of 
photos is searched for a matching object, or similar), with rules for 
how to handle them (say, "wallet -> put in drawer", "plastic bottle -> 
send to trash", ... with an edge case, say, "unknown -> put in drawer").


>> but, the point is that "weak-AI" is within the reach of non-adaptive
>> procedural logic (IOW: hard-coded logic), whereas "strong-AI" is not.
>
> Yes, if you define "weak" as "programmable by average programmer", but that
> would make the classification useless.
>

actually, it is more defined as "programmable by people in general".
for example, an "average programmer" also lacks the skills to implement 
a compiler either, but the minority of people who can implement 
compilers make up for the rest.

however, if something can't be reasonably implemented in-general, or 
performance would not be acceptably good, then it has a problem.


>>>>>> what if it is, instead, say:
>>>>>> speech recognition and NLP;
>>>>>> and a contextual menu-like interface.
>>>>>>
>>>>>> for example, the AI may seem chatty and friendly enough, and can answer
>>>>>> questions and similar, but at its core it is a menu-driven system, and
>>>>>> its "mind" is simply collecting various pieces of information and
>>>>>> navigating between various pre-made menu-forms?...
>>>>>
>>>>> The magic is blown away? But whatever intelligence be, if computable, there
>>>>> is no magic in it, just many, many tiny cogwheels! (:-))
>>>>
>>>> no, a menu-driven AI would be notably inferior to a real human, as it
>>>> would not have sentience.
>>>
>>> It is only a question of reflection, i.e. if you include the states of the
>>> menu in the menu. Fake consciousness should not be harder than fake
>>> decision making, the difference is only in what is considered the input,
>>> the outer word or the inner one.
>>
>> this may not be sufficient, and depending on the implementation could be
>> prone to crashing (say, an infinite recursion takes place and leads to a
>> stack overflow).
>
> You can prune the search tree, which many algorithms do anyway. If
> exceptional states were the only problem...
>

fair enough, one can use recursion and infinite-loop protection strategies.


>> the more immediate concern though is that humans lack the mechanisms by
>> which to experience a GPF or stack overflow or buffer overflow.
>
> Human brain is a massively parallel system, if an exception occurs
> somewhere in such a system it stays local (unless you try to keep
> everything synchronized).
>

I think it is more likely that it has no such mechanisms in the first place.

[toc] | [prev] | [next] | [standalone]

#600

On Sat, 02 Jul 2011 20:29:49 -0700, BGB wrote:

> On 7/2/2011 11:47 AM, Dmitry A. Kazakov wrote:
>> On Sat, 02 Jul 2011 10:03:57 -0700, BGB wrote:
>>
>>> On 7/2/2011 6:29 AM, Dmitry A. Kazakov wrote:
>>
>>>> No, graphics can be measured (fps, pixels, color depth). AI lacks
>>>> quantitative measurement to know if it is good enough and what need to be
>>>> done to have it good. Once you had AI you could measure its quality in
>>>> "intelligent decisions" per second. But there is no any in any time.
>>>
>>> not really, in non-cases, such as the quality of a 3D rendered scene,
>>> ... it as well depends on subjective properties.
>>
>> More fps, higher resolution, etc do it better. (I don't consider cases when
>> things rendered are not ones one actually wanted to show)
> 
> past a certain point there is not much of a point (much past 1920x1080 
> and 60fps, with the present technology, there is not much to gain).

The point is that this is measurable and predictable, we know how human 
eyes work. "Good enough" decision making is not measurable, we don't know 
how intelligence works/is felt.

>>> for many of these AIs, the intelligence and behavior are equivalent.
>>
>> Possibly for all. Behavior is a function of intelligence and the only way
>> intelligence can be observed.
> 
> yes, but behavior need not have much actual intelligence behind it.

You cannot look behind. There is no physical difference in the observable 
effects.
 
>>>> I don't think this is a workable distinction. For example, sweeping floor
>>>> with a broom is a task requiring much more general intelligence than
>>>> playing chess. (Thirty years ago people believed otherwise) I doubt that an
>>>> average office-cleaner exceeds his creators.
>>>
>>> sweeping a floor need only require a robot that can operate a broom and
>>> move around, and probably several layers of bitmaps, and some more
>>> procedural logic.
>>
>> It requires classification of things into valuable stuff vs. rubbish/dirt.
>> No procedural logic would help here.
> 
> a sweeping robot need not have valuation, rather it would just sweep up 
> anything on the floor just the same. if a person worried about it, they 
> could create a mechanism to separate dust from non-dust and have someone 
> manually evaluate the non-dust later (say, robot has a container for any 
> pieces of clothing/shoes/wallets/... it runs across)
> 
> or, it could use basic computer-vision algorithms to classify common 
> object types (say, an object is encountered, and then a long list of 
> photos is searched for a matching object, or similar), with rules for 
> how to handle them (say, "wallet -> put in drawer", "plastic bottle -> 
> send to trash", ... with an edge case, say, "unknown -> put in drawer").

That is it! No office-cleaner has a problem of sweeping woven carpet's 
patterns...
 
>>> the more immediate concern though is that humans lack the mechanisms by
>>> which to experience a GPF or stack overflow or buffer overflow.
>>
>> Human brain is a massively parallel system, if an exception occurs
>> somewhere in such a system it stays local (unless you try to keep
>> everything synchronized).
> 
> I think it is more likely that it has no such mechanisms in the first place.

You must have such a mechanism in order to reduce overall complexity. Not 
all combinations of the inputs are possible, they are not fully 
independent. Any working system makes some presumptions about that and gets 
advantages from this knowledge. Furthermore the internal state of the 
system is correlated with the state of the inputs and their history, it is 
only a presumption that when you are in a button press callback, the button 
was pressed, and when "abc" is cached in the buffer, then 'a' was pressed 
before 'b' and so on ad infinitum.

The strength of presumptions of course vary (from the lift does not work 
today, to no more oxygen left in the Earth's atmosphere). When a constraint 
gets violated you have an exceptional state and have to transit to a state 
where the constraint is no more presumed. It is hard to do when programming 
in a poor language, using wrong methodologies etc, but possible for some 
set of constraints. The constraints which cannot be violated (recovery 
impossible) manifest errors/bugs when nevertheless violated (e.g. if 
somebody tried to dive into hydrochloric acid).

-- 
Regards,
Dmitry A. Kazakov
http://www.dmitry-kazakov.de

[toc] | [prev] | [next] | [standalone]

#612

On 2 Jul., 09:47, "Dmitry A. Kazakov" <mail...@dmitry-kazakov.de>
wrote:
> On Fri, 01 Jul 2011 21:22:33 -0700, BGB wrote:
> >> Anyway, due to Gödel's incompleteness, there always should be a state of
> >> "perplexity" in any inference system. It does not crash the human brain,
> >> because we have multiple subsystems monitoring each other as
> >> semi-independent agents. An AI system should be necessarily built
> >> similarly.
>
> > potentially, or humans don't crash because an input validation problem
> > or buffer-overflow overwrite part of the control stack and/or lead to an
> > invalid memory access (GPF / segfault...).
>
> Interestingly that living systems process a huge amount of input
> information to deal with, quite persistently to DoS attacks...

What about hypnosis (commands successfully channeled to a
communication port)?


Greetings Thomas Mertes

--
Seed7 Homepage:  http://seed7.sourceforge.net
Seed7 - The extensible programming language: User defined statements
and operators, abstract data types, templates without special
syntax, OO with interfaces and multiple dispatch, statically typed,
interpreted or compiled, portable, runs under linux/unix/windows.

[toc] | [prev] | [next] | [standalone]

#602

> > and some of us don't believe in strong-AI either...
> 
> Well, there is an old joke that if someone works on artificial
> intelligence, that is because of a lack of the natural one...

No shit!

[toc] | [prev] | [standalone]

Page 2 of 2 — ← Prev page 1 [2]

Back to top | Article view | comp.programming

csiph-web