From: Eliezer S. Yudkowsky (sentience@pobox.com)
Date: Sat Mar 27 1999 - 19:07:57 MST
Alexander 'Sasha' Chislenko wrote:
>
> Maybe, somebody could write a review?
> (A positive one, please - Ben is my boss! ;-) )
Well, the Webmind architecture looks an awful lot like my own best idea
of an AI architecture as of about two, three years ago. I don't get
much more complimentary than that!
At the time, I was still thinking in terms of nodes, nodes that acted on
nodes and nodes that referenced nodes, where the programming effort was
in creating the basic heuristics and the heuristics that could tell how
well heuristics worked; the latter would act upon the former to let
pattern arise from the nodes and their interactions. The nodes and the
links between nodes would be far more powerful and complex than those of
a mere semantic net, and the nodes could examine other nodes and apply
heuristics to them. In this sense, I suppose my architecture actually
was more general than Webmind; the agents themselves were nodes.
But it was still fundamentally flawed. A mind isn't a semantic net, not
even a super-net. That's not to say your *project* is fundamentally
flawed, in the sense of not being able to achieve the design goals; the
supernet would have been more powerful than any classical AI or neural
net on the market, being a cross of both, and having the capability to
incorporate both into the nodes. Perhaps the supernet could even go all
the way and Transcend, but I doubt it, because I don't trust the ability
of pattern to arise in the supernet sufficiently more powerful than
what's put in; and where nodes must be simple enough to understand
common data formats and each other, they do not have the complexity to
create true understanding. Understanding would have to be built on top
of the nodes, as a pattern of pattern of patterns. This is the
Hofstadterian paradigm applied to semantic nets instead of neural nets,
which unfortunately happens to be wrong in the second case and probably
unworkable in the first; it is certainly inefficient and hard to document.
Again, to remain complementary, I speak of true intelligence, not
achieving the relatively modest design goal of predicting financial data.
>From my current perspective, I would say that a Webmind implements a
specialized (more polite than "crippleware") case of the general version
of the Elisson architecture, in which domdules are incarnated as agents,
and the function of symbolization/domdule interoperation is incarnated
as a limited set of common shared data formats called "nodes". Or to
use the RNUI design principle, the agents incarnate a Notice level and
the nodes incarnate a Represent level; because domdules are broken up
into agents and a set of common data formats used, the problem of
interfacing between full domdules does not arrive.
(RNUI - Represent, Notice, Understand, Invent. You have to represent
something before you can notice it and "notice something before you can
understand it". Drew McDermott invented the NU part of this requirement.)
Unfortunately, this limits nodes to the Represent level and prevents
agents from Noticing any fact that cannot be expressed at the Represent
level. No node formats for higher-level Notice or low-level Understand
data is present. For example, Copycat has Notice-level information of
bonds and correspondences that it uses to create analogies, and codelets
analogous to agents. But how can Webmind represent bonds if the
programmers didn't think of it and provide it as a basic data format?
And won't the explosion of data formats cripple node interoperability by
forcing the programmers to write O(N^2) pieces of interface code?
And how does the Elisson architecture handle the problem? I'd answer
that, but I have to go make dinner. More on this later, and I ought to
read the whole book before I go on. Nice book, though.
--
sentience@pobox.com Eliezer S. Yudkowsky
http://pobox.com/~sentience/AI_design.temp.html
http://pobox.com/~sentience/singul_arity.html
Disclaimer: Unless otherwise specified, I'm not telling you
everything I think I know.
This archive was generated by hypermail 2.1.5 : Fri Nov 01 2002 - 15:03:24 MST