To those doing AGI development: If, at the end of the development stage of
your project  -- say, after approximately five years -- you find that it has
failed technically to the point that it is not salvageable, what do you
think is most likely to have caused it? Let's exclude financial and
management considerations from this discussion; and let's take for granted
that a failure is just a learning opportunity for the next step.

Three possibilities seem viable:

1) The design is totally workable but just requires much more hardware
than is currently available.  (Our current estimates of hardware
requirements for powerful Novamente AGI are back-of-the-envelope
rather than rigorous calculations.)

2) The design proves unworkable.  More specifically, there is one key
aspect of the design that might fail, though I doubt it will.  A key
premise of the NM design is that the different learning/reasoning
components can work together in such a way that each one of them helps
the other one to adequately the combinatorial explosions intrinsic in
its design.  There are fairly strong theoretical reasons to believe
this will work, but the hypothesis has so far only been tested in
simple cases...

3) Teaching the system proves too time-consuming, because of the lack
of the "inductive bias" that baby humans are born with.  (I think we
can work around this by using the logic-based nature of NOvamente's
knowledge representation which will allow us to feed knowledge into
Novamente's knowledge base directly, to serve as a highly rough
analogue to a baby's inductive bias... but this is an untested
speculation...)

Much of my work the last couple years has been focused on minimizing
risks 2 and 3, by simplifying and better understanding the design; and
by working out the elements of a theory of AGI Developmental
Psychology.  Regarding Risk 1, I'm basically counting on the engineers
of the world to keep accelerating hardware development and minimizing
hardware cost....  The NM implementation is already pretty efficient
due to the use of fairly hairy C++ code...

-- Ben

-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/[EMAIL PROTECTED]

Reply via email to