From: Eliezer S. Yudkowsky (sentience@pobox.com)
Date: Tue Jun 18 2002 - 08:41:55 MDT
Anders Sandberg wrote:
> On Tue, Jun 18, 2002 at 12:08:27AM -0400, Eliezer S. Yudkowsky wrote:
>
>> If this is so, then why does evolution, a process of essentially
>> constant underlying "smartness", appear to have accelerated over time,
>> rather than slowing down as more complex organisms are created? The
>> above theory would appear to make predictions which are the diametric
>> opposite of the observed evidence from anthropology and evolutionary
>> biology.
>
>
> Does evolution really accelerate? The evidence I have seen seems somewhat
> inconclusive, and I have one paper lying around suggesting that species
> longevity is increasing roughly logarithmically - which would mean both
> that the improvement in survivability is increasing more slowly, and that
> more and more old species remain.
>
> If you look at it from the perspective of billions of years of single
> celled organisms, hundreds of millions of years of soft-bodied creatures,
> the Canbrian explosion followed by a scant few hundred million years of
> skeletal creatures going onto land, followed by the appearance of
> sophisticated nervous systems and then the appearance of humans a few
> million years ago, it sounds very much like a singularity chant. But on
> the molecular level the rate of evolution might have been nearly constant
> - the first billion years were really the heroic era when replication,
> metabolism and environmental protection were perfected. Since then there
> has just been a lot of homeobox twiddling.
I am in rough agreement with some of this - I've heard some strong arguments
recently (Terrence Deacon) that the evolution of the hominid family was much
more linear and less exponential than was commonly deemed. But I'm not
Kurzweil; it doesn't bother me whether progress is linear or exponential in
weakly self-improving processes, as long as it's not logarithmic. Either is
good enough for explosive self-improvement given strongly recursive
self-improvement a la seed AI.
> My personal view is that evolution does indeed accelerate, but in the
> form of adding new and faster layers on top of the old ones. Metabolic
> evolution is terribly slow. Body plan evolution is faster. Then nervous
> systems and adaptivity added an even faster layer,
All of which directly contradicts the assertion that more complex systems
are necessarily harder to improve. Evolution, acting entirely without
foresight, has managed to set things up so that more and more complex
systems offer more and more handles for making adaptive changes, and ways to
control more and more coordinated complexity through atomic mutations to
single genes that govern complex systems.
> which enabled an super-fast layer of behavioral and cultural change. But
> if this carries over to intelligence development, it might turn out that
> the road to superintelligence is not by looking at super-hardware or
> super-software, but by looking at super-memetics or super-culture.
Evolution's higher levels of organization are also characterized by
decreased efficiency. Proteins are held together by weak Van der
Waals forces rather than orbital sharing, to the detriment of us
protein-based lifeforms. Evolution's way of hacking evolvability without
intelligence unfortunately creates certain vulnerabilities, which are very
hard to get rid of because of evolution's inability to handle simultaneous
dependencies - evolution cannot intelligently refactor DNA to use diamondoid
instead of protein. The creation of intelligence permits us to step outside
the limited pathways of evolution and reopen the lower levels of
organization to intelligent improvement.
Or to put it another way: A human super-culture such as the Earthweb,
ultimately based on 200Hz neurons in nonagglomerative brains, could probably
beat our contemporary society in a fair fight, but would be instantly
crushed by a postbiological superintelligence that thought, oh, say, a
billion times faster (200GHz substrate).
-- Eliezer S. Yudkowsky http://singinst.org/ Research Fellow, Singularity Institute for Artificial Intelligence
This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 09:14:53 MST