From: William Pearson (wil.pearson@gmail.com)
Date: Tue Feb 21 2006 - 19:25:11 MST
Cribbing from Chaitin mercilessly....
The number of friendliness preserving programs a program can safely
self transform itself into is bounded by the length of the program.
Let us posit a number Phi which is equivalent to Omega but for
friendliness. That is bit 1 is 1 iff the shortest program expressible
in a programming languages (breaking ties in a defined fashion) is
friendly and has friendliness maintaining transformations. And so on
with the second bit and the second shortest program. And omega is bit
1 is 1 iff program 1 halts (I don't know whether this is the classical
definition)
Knowing a bit of Phi is 1 is equivalent to knowing a bit of Omega is
1, unless not halting is friendly :)
The easiest way to know whether n programs halt (without empirically
checking them) is to have them encoded in an n bit number. Each bit
representing one bit of omega, so similarly for Phi. Quite often this
is expressed as knowing the first n bits of omega, but as we are not
interested in the ordering of Phi we shall treat them as arbitrary
bits of Phi and hence arbitrary length* programs.
So the maximum number of bits of Phi and hence number of known or
provably friendly programs is bounded by the length of your starting
program. It would be less than that because you would have to take
into consideration the bit requirement of the machinery to transform
your bit of Phi into a program and other things you want your program
to do. But tightening the bound is not of much interest I don't think.
Anything wrong with this semi-formal proof?
I hope I haven't covered ground already well trodden, I couldn't find
anything explicitly positing an equivalent of Omega for the
friendliness quality and not much on Omega and Chaitin together apart
from some stuff by Geddes and a very early one by Mitchell Porter.
Will Pearson
* But not arbitrary kolmogorov complexity programs, by definition.
Note
This is a rejoin. I am going to try and not argue with people about
their approaches to AI or hard take off and so not flog the poor
deceased equines. I will limit myself strictly to self-improvement
theory and intend to automatically ignore a large percentage of list
traffic.
My own work isn't AGI it has evolved into defining a parallel hardware
system that can be used as a metaphor for the system like the brain
including equivalents to neurovascular coupling and the like. As such
I don't see my hardware likely capable of hard take off (quite apart
from my theoretical dislike for it) as I expect if or when human level
software is implemented on it it, decades from now or possibly never,
it would only be able to modify its own software and goal system in
the same way humans can i.e locally and not totally.
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:55 MDT