From: Eugen Leitl (firstname.lastname@example.org)
Date: Sun Jun 16 2002 - 03:53:51 MDT
On Sat, 15 Jun 2002, Eliezer S. Yudkowsky wrote:
> Kurzweil will always argue for the creation of AI based on
> neuroanatomical modeling of all cortical areas, and will never admit
> that a general understanding of intelligence is necessary or even that
> it could speed up the process [...] defend the proposition that an
> understanding of intelligence is possible.
> [very nice dissing continued]
I am not sure there are general principles for intelligence that allow you
to shortcut the process. Sure, computational physics describes cognition
better than vanilla physics, and is a source of constraints to guide your
design. There might be even better models.
However, an effective intelligence lives in the physical world. A world
made rather complicated in patches by emergence of intelligent life (so
far the only instance of the intelligence emergence process and its
properties to study). Something mindlessly munching on congealed star dirt
is not something we're interested in.
Here a short vector of integers in the right context codes for a richly
prepatterned agent, a human baby. A human baby is rife with encoded
expectations about the world, allowing it learn in the first place. If the
baby wasn't prepatterned with such hardcoded expectations about the shape
of the world it is dumped into it couldn't make any progress at all.
All I'm saying this is a lot more complicated and dirtier than, say, is
necessary for the correct design and operation of a nuke plant. What
you're trying to do is to build a baby from scratch, not a fertilized egg.
And without guidance from prior art, because this either means replaying a
few 100 megayears of evolution, or modelling life in machina. With the
capabilities of a few 100 people (limited to whatever complexity the state
of the art allows for) which rely on ad hoc and introspection to derive
what they think happens inside a baby.
Does sound rather hard, doesn't it.
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:39 MDT