Re: Slow-fast singularity

From: Matt Mahoney (matmahoney@yahoo.com)
Date: Thu May 08 2008 - 14:14:52 MDT


--- Stuart Armstrong <dragondreaming@googlemail.com> wrote:

> > > Fast as in a single AI achieves dominance before any rivals can
> emerge
> > > or be created.
> >
> > If competing AI reproduce at different rates, then that is the most
> likely
> > scenario.
>
> What makes you claim that? We have little understanding of
> intelligence; we don't know how easy or hard increases in intelligence
> will turn out to be; we're not even certain how high the advantages of
> increased intelligence will turn out to be.
>
> It could be a series of increasing returns, and the advantages could
> be huge - but we really don't know that. "Most likely scenario" is
> much to strong a thing to say.

I believe that recursive self improvement is an evolutionary algorithm. I
know not everyone agrees with me on this. My argument is that a parent
cannot deterministically design a more intelligent child or it would
already be that smart. Nor can a parent evaluate the intelligence of an
experimental child once the child is more intelligent. I believe this is
true both in an information theoretic sense (using Legg's universal
intelligence, which is bounded by algorithmic complexity) and in the
practical sense that we have IQ tests for child geniuses but not adult
geniuses.

Evolution favors rapid reproduction and acquisition of resources. In a
competition between slow improving and fast improving agents, the fast
agents will win. This does not necessarily favor intelligence. The human
race could be wiped out by fast growing bacteria or nanobots.

Evolution is not a steady state process. It is chaotic. It is punctuated
by population explosions, plagues, mass extinctions and other
catastrophes. The only stable attractor is a dead planet. It is only by
the anthropic principle that all life is not yet extinct. Right now we
are experiencing an event possibly bigger than the transition from RNA to
DNA based life or the Permian mass extinction, possibly just a few decades
away.

-- Matt Mahoney, matmahoney@yahoo.com



This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:02 MDT