Structure of AI (was: Re: [extropy-chat] COMP: Distributed Computing)

[J. Andrew Rogers]

On Nov 20, 2004, at 10:12 AM, Rik van Riel wrote:
> Based on what I understand, intelligence appears to be based
> on information processing and communication, not computation.


As a technical nit, I challenge you to define "information processing", 
"communication", and "computation" such that they are truly distinct 
entities in some kind of rigorous fashion.


> Human intelligence, on the other hand, seems to be best at
> fitting together the pieces of the puzzle.  Not at complex
> calculations.  Power through correlation.  A very dynamic
> setup.


This is generally true.  The brain is essentially a giant 
context-sensitive pattern index that does everything using a few 
primitive operations on that index.  If you think of it this way, the 
limitations start to become obvious.


> Mmm, maybe I should check around, to see if AI research is
> still thinking in terms of programming, or more in terms of
> processing data ...


Most AI research is still thinking in terms of computation, or at the 
very least they view computational power as the limit on intelligence.  
If you look at the models used by most researchers, you can see why 
they might come to that conclusion.  Newer foundational mathematical 
models based in algorithmic information theory would strongly suggest 
that this view is quite incorrect.

I actually may have been the first hardcore theorist in the field of AI 
to assert that there is almost no "computation" in "intelligence", 
something which is considered less controversial and outlandish today 
than when I started publicly making such assertions five years ago.  
Still, old ideas die hard.  As one of the first people to take a 
serious stab at defining intelligent systems and AI in terms of 
algorithmic information theory, it became obvious to me that the 
pervasive view that intelligence is bound by computational power was 
not supportable in the mathematics.  As a foundational mathematical 
model of intelligence, this general area has done very well; there are 
far more reasons to think it is correct today than when it was first 
proposed, and it has generated the first really new directions in ages.

It is worth pointing out that if you take these models into 
consideration, which really are the only mathematical framework for 
generally intelligent systems we currently have,  the popular models of 
what constitutes a "human-equivalent computer" (e.g. Moravec) are 
completely and deeply broken.  The closest direct metric of 
intelligence capability for silicon would be cache line fill rate and 
total RAM; TFLOPS are almost totally irrelevant as a practical matter.

cheers,

j. andrew rogers