The problem of intelligence.

Lester Zick wrote:
You keep mentioning the "paradox of intelligence". What exactly are you talking about? Where is the paradox? 

Nothing special about the term. We might say "puzzle" instead of
"paradox". I think the reason I used "paradox" is that no one seems
able to say exactly what "intelligence" means except metaphorically.

A paradox is not only a puzzle but a seemingly contradictory one. I suppose we can brainstorm a few ways intelligence fits in there but lets just call it a problem and move on. :)

What do you mean by concrete here? Are you asking whether or not we can measure intelligence? 

No. I'm asking what you and others mean by intelligence in mechanical
terms? You used "concrete" to specify a gauge for progress in solving
the problem of intelligence. So what do you mean by "concrete"? If you
mean something like an arbitrary gauge of whatever you take to measure
intelligence then I'd have to ask whether that "concrete" gauge really
measures intelligence in mechanical terms?

I don't think you would disagree that a "concrete" definition of intelligence (more on that later), capable of being used as a progress meter, would be very important. But we need to address the philosophical side of things first.

Well I consider your original argument against "we'll know it when
we see it" techniques for assessing artificial intelligence perfectly
valid. And I am definitely not categorically opposed to defining
intelligence. But I'm definitely categorically opposed to defining
intelligence anyway we want just so we can get on with mechanizing
the definition without actually defining the mechanics of intelligence
just so we can say we've made progress on some arbitrary definition.

Granted the approach of the field of AI has been something to this effect with no relevant results.

Do you actually know what a paradox is? 

Well I'm not stuck on the term. If you want to call it a puzzle
instead go ahead. It's still an unkown in mechanical terms and it
seems to produce certain paradoxical effects such as the mind and
mental effects not generally evident in material interactions.

We can't sense these effects, but then again with our technology its like trying to feel an ant with a hammer. I wouldn't say that they are "unsensable" though.

Sure but the question here is why exactly you expect intelligence to
be tangible if it produces intangible effects such as the mind and
mental effects? So either you're stuck trying to substitute tangible
effects for intelligence so you can measure them or you have to
explain how it is that intelligence is mechanical yet intangible. I
thought your original objection was that simply saying "we'll know
it when we see it" was unsatisfactory because we don't know what
we're looking at in mechanical terms. But adding further tangible
qualifications to a tangible "we'll know it when we seet it" doesn't
make the effect itself or it's explanation tangible.

The original objection says we can't "know it when we see it" unless we have some mechanical means of judging. I think our problem needs more definition.

I suppose I read your original objection as opposed to tangibility for
intelligence. So I have to ask whether you would expect to be able to
"see" intelligence at all and if so in what terms?It's one thing to be
able "see" various behavioral, neural, and skilled tangible effects of
intelligence without being able to "see" the thing itself in tangible
terms. So if you don't expect to be able to "see" intelligence in such
tangible terms then all you can really expect to be able to do is list
a whole slew of intelligent effects and mechanize them artificially.


I never meant to say intelligence is intangible.

Well if your point is that intelligence in itself is a tangible thing
then I can see that the more the better as far as measurement is
concerned. What I don't see is how or whether tangible measures
can have any bearing if intelligence is mechanically intangible. I
doubt very much that intelligence itself is a tangible thing even
though it produces tangible effects. But simply substituting tangible
things for intelligence simply because they're measurable doesn't
have any bearing on whether or not intelligence is tangible. Nor
would it have any bearing on figuring out what intelligence is if it
is intangible. And that is the fundamental issue requiring a solution
before intelligence can be mechanized artificially or otherwise.

I think we need to take a stab at putting more definition on our problem. I mean, what are we looking for really? What sort of product would satisfy our goals?

* First and foremost, we are looking for an algorithm. We want this algorithm to run on a computer and exhibit certain properties.

* There may not be a single answer. There are likely many ways in which our constraints can be satisfied. There is no one "intelligence", what we are seeking may one of many different ways of doing the same thing. It is common in Computer Science to see completely different algorithms do the same thing optimally.

* Our answer may not approximate human intelligence. The perennial example here is the bird and the airplane. Both fly, but which one makes more sense for us to use for human flight? It is likely that a top-down approach, working backwards from criteria, is the best way to approach the problem. Reverse-engineering the human brain is not practical today.

* Our answer is not the silver bullet. Humans are capable of a great variety of skills and tasks, but we should not be overly optimistic with our goals. What we create may only be capable of intelligence within a limited domain. We are looking for generality, but we should not expect to find it right away.

* To define what properties we want to have, we need to pick the smallest set of traits with which we will be satisfied. There are many superfluous requirements that get placed on the problem. Ones such as "it must be reproducible in neurons", "it must be scalable", or "it must be built out of elements x, y, and z". If we heap on too many demands we will never find a solution, but then again, if we require too few we will not find an acceptable answer.

We can make all of these points without even touching on a definition for intelligence. But if we want to go further, we need to look at the defining traits of intelligence. What would we require? What qualities are we looking for?

* Agency. Intelligence is a quality that pertains to an agent. Agency requires that there exist input signals from and output signals to an environment. These signals need to be sufficient for whatever task we want to accomplish.

*  Cognition. The agent should be capable of...
   ...accepting an arbitrary goal within the environment.
   ...accepting relevant information about its environment.
   ...planning actions that would lead to a more favorable position.
   ...creating sub-goals and plans for investigating them.
   ...abstracting, generalizing, and classifying information.
   ...generating relevant information from basic environmental inputs.
   ...reasoning about the environment from known information.

Most of these requirements are about mental effects. If we look at each item individually it is not impossible to think of algorithms that could do this.

Another point is the importance of the environment. We can't define intelligence without referring repeatedly to the environment. What is intelligent in one set of circumstances may not be in another. Is it possible that a human could be placed in a set of conditions under which we would cease to be intelligent? If not, can we remove environmental restrictions from our definition?

This is a mechanical definition of the problem. Although it is still "we'll know it when ..." it can be satisfied, and it is also possible to look at the items individually and partially qualify a solution.

Is this an adequate definition of the problem?

-- Risujin
.