RE: Minimum complexity of AI?

From: Ben Goertzel (ben@goertzel.org)
Date: Sun Feb 20 2005 - 22:15:18 MST


Well, I think that a completed Novamente AGI would have less than 100K lines
of C++, and that's without much effort for compactness. I guess this could
be done in about 20K lines of LISP, again without obsessive effort for
compactness. If you wanted to be really annoying and aggressive about
code-length optimization, call it 5K lines of LISP.

And the Novamente design hasn't been created with minimal size in mind,
quite apart from coding-style issues. I imagine that if one wanted to
create a modified design that took up as few lines as possible one could get
it down to 2000 or so lines of highly cryptic LISP code. But this amount of
compression would have an impact on the speed of self-modifiability because
the system would have a harder time understanding and modifying its own
cryptic code in the early stages. I think the 10K-20K lines version would
be better.

If one wanted to use LISP, which we have decided against at this point,
though creating a LISP Novamente in the future isn't totally out of the
question.

-- Ben

> -----Original Message-----
> From: owner-sl4@sl4.org [mailto:owner-sl4@sl4.org]On Behalf Of sam
> kayley
> Sent: Sunday, February 20, 2005 11:18 PM
> To: sl4@sl4.org
> Subject: Minimum complexity of AI?
>
>
> Another question unlikely to currently be usefully answerable .. how large
> do you expect the smallest AGI implementation to be (e.g, in lines of
> obfuscated lisp)?
>
> To make the question somewhat well defined, assume:
>
> Memory available around number of synapses in human brain * a small number
> of bits.
>
> Learning from sensory-motor experiences in a robot body, with
> teaching done
> by people with minimal special knowledge of the AI's workings (any special
> knowledge required must be included in the Line Count).
>
> Total number of CPU cycles used in learning + proof of being a functioning
> intelligence no greater than a reasonable estimate of equivalent that a
> human brain uses in 20 years.
>
> Goal system not entirely arbitrary (FAI Line Count is a separate
> question).
>
> Defining roughly human level or greater general intelligence exactly and
> concisely isn't easy, so I will give some examples as cold water for any
> cheating philosophers on the list: Learning to play Go to a reasonable
> standard given an english definition of the rules and/or a partner to play
> against, arguing whether it is sentient, designing a bicycle given an
> english description of requirements, driving a car in realistically
> uncontrolled situations. Challenge task: creating a kit that replaces a
> persons appendages with blue tentacles with an awkward tendency
> to curl into
> the shape of paperclips when not in use, and deciding whether
> this should be
> applied to Michael Jackson.
>
> Perhaps the compactness and subtlety of the program that generates the
> mandelbrot set, or some of the IOCCC entries is an appropriate
> comparison ..
> length of program + sufficient documentation/comments to be comprehensible
> is a different question, as is the length of the shortest AGI program that
> will be written before the science of mind is mature.
>
>
>
>



This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:50 MDT