On 5/25/20 2:36 PM, Alan Grimes via AGI wrote:
Ok, it seems we have lost a concept here, that of metaprogramming. In
classic interpreted languages such as Lisp, etc, it was possible to
write programs that write programs.
When you think of AGI that needs to optimize itself for its platform and
continually adapt to its environment, it should quickly become apparent
that we are not looking for an AGI program per se, but a program that
generates and continually tunes and updates the AGI program.
Yes, and that gets closer to a discussion of architecture. When you
consider the "program" of the AGI is will be more like a collection of
routines that is navigated by the existence of circumstances (data.)
Like the habits you mention, the collection of routines will be acquired
over a period of the AGI's development. The AGI can quickly alter it's
overt behavior by "switching" on and off the habits or micro-habits.
I've proposed that the habits be considered as "opportunity" that
deliver a benefit, and that the opportunity be triggered by changes in
data. One need not continually reprogram the whole thing (retrain the
neural net.) Instead add a new opportunity that is of greater
"effectiveness" to alter the behavior. This "architecture" requires
that there be a mechanism of arbitration going on continuously. Several
possible opportunity may exist for a given moment, and arbitration
determines which is "best." The opportunity that shows the greater
promise (greater effectiveness) is chosen to execute for the next cycle.
I get that people will call this method absurd - someone has to write up
new opportunity and add it to the collection - "a real AGI learns on
it's own..." The quick answer to that objection is that the "unit" can
be taught to recognize opportunity, package it, and bring it into it's
collection. It will assess and adopt based on the routines it contains
for that purpose. Even those "assess and adopt" routines can be
replaced as better methods come along - a better opportunity for that
purpose.
Consider also your habbits, not just your external habbits but you also
have hundreds of micro-habbits for dealing with routine social
interractions and cognitive tasks. These can all be thought of as
generated programs that were generated by the metaprogram that is your
underlying intelligence.
Anyway, don't let your mind get trapped in the limitations of commonly
used programming techniques.
------------------------------------------
Artificial General Intelligence List: AGI
Permalink:
https://agi.topicbox.com/groups/agi/T5364f6a30c9294ea-M24c1a23546c6ddb9f812ab00
Delivery options: https://agi.topicbox.com/groups/agi/subscription
