Begin forwarded message:
From: Brad Templeton <btm@xxxxxxxxxxxxxx>
Date: July 14, 2006 7:39:08 PM EDT
To: David Farber <dave@xxxxxxxxxx>
Cc: pollack@xxxxxxxxxxxxxxx
Subject: Re: [IP] more on "Strong" AI to be here within 25 years
On Fri, Jul 14, 2006 at 06:22:12PM -0400, David Farber wrote:
Begin forwarded message:
From: Jordan Pollack <pollack@xxxxxxxxxxxxxxx>
Date: July 14, 2006 4:28:13 PM EDT
To: dave@xxxxxxxxxx
Subject: Re: [IP] "Strong" AI to be here within 25 years
I'm sorry, but baloney is still baloney, because Moore's law doesn't
increase the quality and complexity of our software. We'd see
something coming on supercomputers or grids. Alternative views about
the next 50 years are in the current issue of IEEE Intelligent
Systems,
*http://tinyurl.com/hul5g
*IEEE unfortunately charges a fee, but my paper "Mindless
Intellligence" is available free at http://ectomental.com
Oddly, Moore's law gives the illusion of increasing the quality and
complexity
of our software, particularly in the fields of "former AI." (Former
AI is stuff
that used to be called AI, but once we figured out how to do it,
people stopped
calling AI due to the fundamental theorem of AI.)
For example, in fields like speech recognition, OCR, and visual
pattern
recognition, while it would be foolish to say there have been no
algorithmic
advances in these fields, a sizeable amount of the progress has come
because
interesting algorithms designed decades ago became computationally
workable
do to Moore's law trends.
There is one path to machine intelligence (non-artificial) which does
not require
any advances in complexity of software. Namely, if we can, in the
next 25 years,
learn how to reverse engineer the nervous system at the underlying
level so that
we can build things which act just like neurons, glial cells and
their associated
systems out of something else besides natural proteins, then we can
simply copy
the "software" (patterns of interconnections, chemical flows and
rules) from
protien brains into this new substrate, without understanding it at
the higher
level.
To give an analogy, a capable hardware engineer, who knows only about
IC design
and digital logic, can build Hydra, the world's best chess player,
just
by copying the software into hardware she builds. She need not have
any understanding
of chess, or indeed anything above the level of the machine
instruction set.
However, should this transfer of a human mind into another substrate
take place,
this person would be unconstrained by many of the rules which bind us
as biological
beings. In particular, much has been written about how such a person
could
engage in "recursive self-improvement" using the ability to tinker
with their own
makeup in experiments, seeking improvements. The then improved
person could continue
this process -- doing even better at it, in an explosive increase.
While it is far from certain this will happen in 25 years, it is also
foolish to
suggest it can't happen in that period of time. While we don't yet
have a sufficiently
good model of the neuron, nothing we know suggests this is
impossible. (Rationalizations
about microtubles notwithstanding.)