From: Ben Goertzel (ben@webmind.com)
Date: Tue May 01 2001 - 18:30:30 MDT
>
> Thanks Ben, for making it available.
No worries mate --
I eagerly await your detailed comments on various tricky points ;>
>
> As to the NDA: I would estimate that about 65-75% of the material
> is already
> available online, anyway. The additional, more detailed protected
> information is certainly very useful. How much of a risk/ limitation this
> NDA represents has to be a personal decision. In my judgment:
> quite minimal
> (much is already in the public domain, you are unlikely to copy
> significant
> portions of the design, a lot of crucial detail is excluded,
Yeah, the book does not include code-level detail, and only gives bits and
pieces of object structure. There's an additional roughly 10,000 pages of
documents containing all this, which will be made accessible only to people
who are actually working on the system.
The book tries to outline the **conceptual** design of the system -- how it
works as a complex system, and the general software architecture used to
implement this complex system..
Personally, I would like to publish this book, but whether this can happen
depends on others besides myself... i.e. those involved with whatever
successor businesses to Webmind Inc. emerge.
Another comment -- I definitely don't believe this is the ONLY way to create
real AI. I just believe it's A way. There might be better ways, I just
haven't found any. I think it's very possible that deep
technical/conceptual dialogue with a bunch of you folks could lead to
changes to the design -- tweaks or even substantial changes.
[This paragraph is mostly for Peter]
There are some moderately major parts of the design that I'm not in love
with, but they seem workable and I couldn't find workarounds. For instance,
I don't like VariableNodes and explicit variable unification. HOI is too
logic-y for my taste. I'd rather get that to emerge in a more subsymbolic
kind of way. But we couldn't figure out how to do this, after very much
effort, and we couldn't identify any serious a priori problems with the
current approach. (Pei Wang thought of the current approach, and he likes
it of course). Another example: I don't really like having context
formation use its own data mining algorithm instead of a more
general-intelligence approach. But, the folks working on that pointed out
that using the specialized approach they advocate will be much more
efficient and doesn't really lose you anything in real use cases they
analyzed. So.... In this manner, a lot of the questionable design points
came out of a lot of thinking and analysis that unfortunately isn't fully
recorded in the book, but that will be drawn out when we talk about the
stuff. I think my own intuition is closer to yours than the actual WM
design, because in building the WM AI Engine, when I couldn't come up with a
perfectly elegant and intuitive way to do something, we just invented
something that seemed to work. The balance between engineering and
theory...
My idea is, rather than getting a design that I think is PERFECT, to get a
design that I think is workable, and implement it, and then learn about its
benefits and shortcomings through experimentation.
One thing I can tell you from experience is how VERY VERY MUCH effort it
takes to go from an intuitive conceptual design for a real AI, to a detailed
design suitable for implementation. Whoa!
-- ben
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:36 MDT