RE: Comparative AI disasters

Billy Brown (bbrown@conemsco.com)
Fri, 26 Feb 1999 09:42:57 -0600

Eliezer S. Yudkowsky wrote:

> Now comes the interesting point. This very day, I realized that goal
> systems are not a necessary part of the Elisson architecture.
> In fact, they represent a rather messy, rigid, and inefficient method of
> "caching" decisions. The goals I had designed bear the same relation to
> the reality as first-order-logic propositions bear to a human sentence.
> If you regard a "goal" as a way to cache a piece of problem-solving
> logic so that it doesn't have to be recalculated for each possible
> situation, then you can apply all kinds of interesting design
> patterns, particularly the "reductionistic" and "horizon" patterns.
> A "goal" caches (and *considerably* speeds up) a piece of logic that is
> repeatedly used to link a class of facts to a class of choices. This
> may not hold true for BAI's #initial# goals, but it does hold true for
> practical reasoning about a subproblem, and BAI will still notice..

For an SI, or even a distributed consciousness, I can see real merit in this approach. I'm not so sure that it makes sense for a small, highly concentrated mind like Elisson (or humans, for that matter). But perhaps I've misunderstood you.

I've been assuming that Elisson would have a single goal representation system, a reasoning engine (presumably implemented as a collection of specialized domdules), a central world-model, a sensory processing system (which provides input for the reasoning system), an effector system (which carries out instructions issued by the reasoning system), and various memory subsystems. The reasoning system has an implicit goal of "figure out what to do next", which is what drives it to populate the initially empty goal system.

With this kind of architecture, the centralized goal representation is what keeps the whole system moving in the same direction. If one part of the reasoning system adopts a new sub-goal, everything else can immediately see and act on that goal. If a goal is deleted, every module that is working on it will quickly notice the change.

What would change if you replaced the IGS with and FCCS, and how would this be beneficial?

Billy Brown, MCSE+I
bbrown@conemsco.com