Hi Matt, and others,

I have decided to forward Matt's post to the AGI list, because it's about
your critique of logic-based AI.  My list Genifer is more focused towards
technical implementation of Genifer.  Also, I have already rebuked your
argument on my list countable times.

------------- Forwarded message --------------------

>From: "YKY (Yan King Yin, 甄景贤)" <[email protected]>
>In the shortest term, intelligent virtual pet is all I can think of.  But
as I stated before, we can't be sure how long Genifer will stay in that
stage.  Probably not long if the learning algorithm is good.


OpenCog had this I think a year ago. You can also buy pet robots that
display some intelligence. This hasn't led to rapid learning. Why should it?
Most of what humans learn is through language, which these lack. The
alternative is reinforcement learning, which is slow because each
reward/penalty decision transmits only 1 bit.

Might I suggest the problem that Cyc originally intended to solve (and
failed), the software brittleness bottleneck. Lenat believed that computers
made bad decisions because they lack common sense. If computers knew what
most humans knew, then they would do the right thing. They would do what you
mean and not just what you say. This seemed like a reasonable thing to do
when the project was started in 1984, when there was no internet and
software was something you bought on a floppy disk that ran on a computer
with a text-only display and no mouse. But software today has the same
problem. To use an example relevant to that era, you still can't tell your
computer that you want to edit the document about golf that you worked on
last week, even though the computer has all of the information it needs to
satisfy your request. It knows the current date, the last update time of
every file, and the text contained in those files. But it still insists on a
folder
and file name.

Common sense is anything that most people know. Cyc still lacks a very
important aspect of common sense. Everyone knows the meanings of words and
sentences. But computers don't and Cyc doesn't. Software developers can't
use Cyc to make their software easier to use because it doesn't help their
software understand natural language.

I believe the fundamental error that Cyc made in their design was trying to
build knowledge about language on top of structured knowledge about the
world. The base layer made sense for 10 MHz computers with 1 MB of memory
and a floppy drive. Structured knowledge representation was the only thing
they could implement. Modeling language is as computationally intensive as
modeling vision. But it is the wrong design. People learn language first,
and build knowledge about the world on top of that. I suppose it is possible
in theory to specify the rules for English using first order logic, if we
knew what those rules were. But we don't, because that's not how we learn
language. Children can put nouns and verbs in the right order without
knowing the difference between a noun and a verb.

We need to start with the only known working example of human intelligence
and optimize from there.



-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393
Modify Your Subscription: 
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968
Powered by Listbox: http://www.listbox.com

Reply via email to