quote:Original post by Cedric
Then what example can you give of something that really represents artificial intellgence?

Since we don''t currently *have* artificial intelligence, no.

But I suppose this really depends upon what we mean by ''artificial intelligence''. If we use a Turing-based definition, then a program is an artificial intelligence if its behaviour cannot be distinguished from that of a human being of average intelligence.

This is very domain-specific. From a question-and-answer perspective, the nearest I''ve seen are sophisticated expert systems. But there are issues of language recognition, visual and audio perception and motor control that many people would expect of an artificial intelligence. For perception and motor control, ''neural'' networks appear to be a partial solution.

quote:
And what do you expect from AI scientists? That they build every bit and every instruction of a program that is intelligent?

Until we understand how humans process reasoning, language, and perception, we don''t know how to tell a computer how to be intelligent. So no.

quote:
Learning is a process that is inherently done autonomously (otherwise, it''s not learning). And genetic programming is just another way of learning, by trial and error.

To say that a GA ''learns'' to sort lists is rather like saying that fish evolved into land-animals by learning to breathe air. I consider that to be an overgeneralisation of the word ''learn''.

One may associate learning with intelligence, but I think it''s wrong to associate the evolution of innate abilities with intelligence.

The key note of intelligent learning is that intelligent beings can apply what they learn outside of the context in which they learn it. When a human learns to sort lists of numbers, he is able to abstract that knowledge and apply it to any list of comparable things. A GA cannot do that, simply because a GA cannot generalise what it learns -- it has no reasoning ability.

The nature of GAs is that they only learn the correct responses to the fitness tests. Train a GA to sort a list, and it will sort that list and no other. Train a GA to sort randomly selected lists of length N, and it will learn to sort arbitrary lists of length N. Train a GA to sort randomly selected lists of random length, and you *might* get what you want, but it may be difficult to prove it.

Anyway, is anyone going to make with the language ideas, or shall we debate more off-topic issues?

Is there a way to get this discussion output into some niceer format? I would like to save it to hard disk.

" ''No one has control -- control is just a fantasy. And being human is difficult.'' "

quote:
Anyway, is anyone going to make with the language ideas, or shall we debate more off-topic issues?

Mayrel: I agree, but as I''ve begun work on some of the ideas presented in this discussion (in the past two years) I believe that AI based "patterns" and even OO patterns are pointing in the right direction.

Some of the complication is envisioning AI, but isolating part of a neural net to code generation, unit testing, and compilation-loggging is a dirrection worth going I believe ( I just have to prove it -- I can feel my thesis paper getting really, really long.)

I have to say I appreciate the discussion -- very good insights that I truly was bumping into in my design but didn''t have a good vocabulary to define them. For instance problem domains on a program by programmer level -- interesting thought there.

Especially when compounded with the idea of a program generating code specific to a set of constraints -- which makes me think of prolog only becuase it deals greatly with constraints and unions -- but at the same time Lisp is powerful enough to implement prolog in about 200 lines of code (with comments) which is to say yes, that lisp is very flexible, and powerful, but the idea here is rather simple a program that writes itself right? Well how about a programmer that simply constrains, optimizes and make efficient the areas of self writing program.

Matched with OOP and plug ability you are instead dealing with a set of semantics that are sort of "data/level/event/hardware-abstraction"--driven.

And like a net, wire frame, the program would be dealtwith at a landscape sort of level, defining places of height, and places of shallowness, places of quick descent/ascent. Whatever -- its a form linguistic geometry that is instead treated to mold itself to hardware abstraction and diverge where data-driven models require it too. (Linguistic Geometry is fairly new, but does exist.)

Anyway --
later --
L-