Re: [sl4] An attempt at empathic AI

From: Matt Mahoney (
Date: Mon Feb 23 2009 - 06:56:40 MST

--- On Sun, 2/22/09, Johnicholas Hines <> wrote:

> <> wrote:
> > "Analyzable" and "modular" imply
> low algorithmic complexity. Do you want a system that is
> predictable, or one that is smarter than you? You can't
> have it both ways.
> I'm not sure we're talking about the same thing. In this situation, it
> is probably appropriate to make more detailed, mathy arguments. Is
> algorithmic complexity the same as Kolmogorov complexity?
> The length of the shortest program that outputs the given string?


> Assuming that it is, consider this pseudocode:
> function main():
> if complicated_subroutine():
> print "True"
> else:
> print "False"
> Further assume that complicated_subroutine doesn't do any input or
> output, a property which can be verified structurally. Then we can
> prove a non-trivial property of the program. It will print "True" or
> "False" if it prints anything. For example, it will not print "Hello
> World!".

It is possible for an algorithm with high complexity (i.e. there is no simple description of it) to do something simple, meaning there is another small program that does the same thing. My claim is that an intelligent algorithm can't be reduced in such a way. The simplest equivalent program is still complex.

> Theorem provers might be structured to pass a discovered proof through
> a filter before outputting it. If you look at Schmidhuber's and
> Hutter's various constructions, I think most of them provide
> guarantees on the behavior of the whole despite untrusted, arbitrarily
> complex components. So far, the guarantees tend to lean toward "keeps
> making progress" rather than "Friendly", but I think structural
> arguments for Friendliness are the way to go.

Unfortunately there is no known simple description of "friendly". Descriptions like CEV get around the complexity issue by pointing at the human brain without describing how the human brain works.

-- Matt Mahoney,

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:04 MDT