Re: Seed AI milestones (was: Microsoft aflare)

From: Christian Szegedy (
Date: Wed Feb 27 2002 - 10:56:57 MST

Ben Goertzel wrote:

>There is, however, a qualitative difference between (here I go again with
>crude categorizations):
>1) self-modification like that of the human brain, in which the low-level
>learning algorithms and knowledge representation mechanisms are fixed, but
>high-level learning algorithms and knowledge representation techniques are
>learned and modified
>2) self-modification like that of an uber-brain in which new
>neurotransmitters could be inserted, modifications to the conductive
>properties of neurons could be made, and so forth.
>3) self-modification like that of a brain that could rebuild itself at the
>molecular level, replacing neurons, synapses and glia with entirely
>different sorts of structures
To me, this distinction is far from clear, but this must be my fault.
My main problem is that you mix the conditions imposed on the substrate
with that on the
abstract structure. I believe that no principial distinction is needed
between different
(but reasonably rich) substrates.

>A mathematical formalization of the different varieties of self-modification
>would be possible, but I'm not going to undertake that project today. Maybe
>next month!
It sounds very exciting, provided it is true...

Perhaps you found my scepsis on the mathematical formulation arrogant.
Let me explain it
a bit more thuroughly: Let me assume that yuo have two computational
system. In order
to compare them, you will have to investigate the differences between
them. There are
two types of differences:

A) Is something achiaveable in one system which is not in the other?

B) Is something more effectively achievable in one system than in the other.

Normally you can rule out differences of type A), since almost all
reasonably complex
systems are universal Turing machines. That is, almost all substrates
are equivalent,
therefore the only question is the efficiency overhead compared one to
the other.

In order to discuss the efficiency, you have to select some (class of)
tasks and analyze the
efficiency of performing that particular (class of) task using both of them.

The first reason of my scepsis is that giving lower bound on the running
time of some
task within some system is a very hard, let us say almost impossible
task for most systems
and most tasks.

In order to tell the difference between the type 1) , type 2) , type 3)
modifications, you have to
1) Select (and accurately describe!) some relevant tasks.
2) Create some measures for the efficiency of their implementation.
3) Show some differency between their efficiency in the different

I am not a real expert in complexity theory, but I can tell that it is
quite unprobable that you have
significant results in the above three points without generating so much
hype in the complexity
theorist community that I would not know about it.

Best regards, Christian

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:37 MDT