Re: [agi] A difficulty with AI reflectivity

From: Eliezer Yudkowsky (
Date: Fri Oct 22 2004 - 08:21:17 MDT

Christian Szegedy wrote:
> I liked Schmidthubers paper. It contains several important ideas that
> could be important for construction of a practical AI.
> Schmidthubers paper is a greatly modified version of Levins (and
> Hutters) universal search algorithm. If one understands the original
> idea, it becomes immediately clear that although the universal search is
> a very simple and provably almost optimal algorithm (at least
> asymptotically), its practical value is simply zero: the unimaginably
> huge additive constant makes it practically untractable.

*nod nod*

> For me, it appears to be some kind of cheating. Assume that some people
> develop by hard work some really good program-/proof-generation
> techniques, then we simply puts their plugins into his framework and
> gets provable optimality and practical performance at the same time and
> can reap the credits for his "important contribution"

Oh, come now, that's not fair; we haven't seen him try anything like that.
And it *would* be an important contribution, to correctly say the X in
"what kind of really good <X> techniques you need to make an AI". My own
fear is that the basic AIXI/OOPS/GM formalism lacks basic, foundational
elements needed for practicality - AIXI might work given enough brute
force, though I'm not sure GM would work at all as it seems underspecified;
but to make an AGI run in our universe, you might need additional basic
theoretical elements not discussed in the Solomonoff-Levin-Schmidhuber
-Hutter tradition.

The value of Schmidhuber's and Hutter's work is precisely that we *can*
talk about it and try to figure out what's missing. All other proposals
for AGI are so underspecified we can't even do that.

Eliezer S. Yudkowsky                
Research Fellow, Singularity Institute for Artificial Intelligence

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:49 MDT