From: Jef Allbright (jef@jefallbright.net)
Date: Fri Jun 09 2006 - 14:11:40 MDT
On 6/9/06, Martin Striz <mstriz@gmail.com> wrote:
> On 6/9/06, Eliezer S. Yudkowsky <sentience@pobox.com> wrote:
> > Martin Striz wrote:
>
> > An AI with random-access memory *is* a complete internal model of
> > itself. Why should an AI bother to quine itself into its scarce RAM,
> > when two copies contain exactly the same information as one? What good
> > does it do to model yourself perfectly? What more does it tell you than
> > just being yourself?
>
> Wouldn't it be smart to test designs in a model before you dedicate
> them to your source code, rather than willy-nilly rewriting the stuff
> without being sure empirically what the changes would do? That seems
> even more dangerous.
>
> Either way, my point stands: you can't guarantee that AIs won't make mistakes.
>
> Martin
Consider that your're not so interested in modeling the agent, but
rather its interactions with its environment.
- Jef
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:56 MDT