Re: Seed AI (was: How hard a Singularity?)

From: Eliezer S. Yudkowsky (sentience@pobox.com)
Date: Sun Jun 23 2002 - 12:47:50 MDT


James Higgins wrote:
> At 06:37 AM 6/23/2002 -0400, Eliezer S. Yudkowsky wrote:
>
>> You think my design is too complex. Okay. Nonetheless, the more
>> complex a design is, the more mind arises from the stuff that
>> implements that design, and the more opportunities there are to improve
>> mind by improving the implementation (never mind actually *improving
>> the design*!) I think that
>
> Complexity = capability? I don't think so. Now, I'm not talking AI here
> but, I've found that a simple well architected system usually outperforms
> a large complex system. Excessive complexity usually arises from a lack
> of true understanding of the problem domain. Maybe this doesn't apply to
> AI, but it applies to most (everything?) else I'm aware of.

In my experience, your instincts are correct; this doesn't apply to AI, but
applies to everything else. Excessive complexity usually arises from a lack
of true understanding of the problem domain, and simple systems with good
architectures are usually far better than complex ones. Except in AI, where
the problem is far more complex than your brain "wants" to think - in part
because of all the complexity of thinking is invisible to you, and in part
because usually humans just don't deal with things that complex.

> Plus, very complex systems are much more difficult to produce and almost
> always contain many more bugs. Considering the tenets of FAI the
> potential for bugs that could cause the system to stray (and thus
> possibly not remain "Friendly") worries me considerably, especially in a
> very complex system.

Any Friendliness error that does not actually kill you or permamently warp
the goal system is at worst a temporary inconvenience, and permanently
warping a system with causal validity semantics isn't easy.

-- 
Eliezer S. Yudkowsky                          http://intelligence.org/
Research Fellow, Singularity Institute for Artificial Intelligence


This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:39 MDT