RE: Whatever happened to Cyc?

From: Ben Goertzel (ben@webmind.com)
Date: Sun Apr 29 2001 - 07:11:48 MDT


Hi,

I talked to Dwight Lodge (their CEO) and also briefly Doug Lenat recently...

According to Dwight, the firm's goals at the present time are NOT to create
a program that passes the Turing test or carries out any kind of
conversation or dramatic AI activity on its own. Rather, its goals are to
create knowledge databases and reasoning tools that can be used inside other
firms' products, thus making the universe of software gradually more and
more intelligent. (I'm paraphrasing him, not quoting, but you get the
idea.)

They have a huge commonsense knowledge base now, and also a pretty
sophisticated reasoning engine that has specialized modules for dealing with
temporal logic, number logic, and so forth. I have not used their reasoning
engine, but my guess is that it's very "brittle" in the classic AI sense.

I think that their knowledge database could well be useful to a "real AI"
system. However, it could never be the primary knowledge source of such a
system -- that primary knowledge source has ultimately got to be the
system's own experience. The "definition as necessary and sufficient
condition" framework that underlies most of their knowledge database is not
really adequate for intelligence. In order for a mind to have a working
knowledge of a concept, it needs a whole complex network of facts and
intuitions regarding that concept, of various levels of abstraction and
confidence, not just necc.-and-suff. condition definitions of the concept
and related concepts. But once it has experience with a concept, being able
to load in the necc.-and-suff.-condition definition of that concept and
other related ones from CYC will be a useful ability for a digital mind to
have.

In terms of Lenat's original AI goals, I very much doubt that he personally
has given up on them. However, CYCorp is a business and at this point, a
surviving one. My sense is that Lenat feels "true AI" will be achievable
eventually once enough knowledge has been accumulated and once his reasoning
engine is hooked up to the right other software components (NLP, perception,
etc.). But the mission of Cycorp as a company is not to do this "hooking
up" but rather to build the database and the reasoning engine. Of course,
if they become fabulously financially successful then they can take on the
larger project.

I tend to think Lenat's approach to AI is wrong, in that, in my view, you
can't just develop one aspect of mind in isolation and expect it to work the
right way for interaction with the other parts.

In terms of the "failure" of Cyc, you have to understand that the hypothesis
behind it was that once a reasoning engine had enough common-sense data, it
could make intelligent common-sense inferences about the world. Well, this
is a tough hypothesis to falsify. If the approach has failed up to date,
maybe it's just because not enough data has been loaded in!!!

ben

> Does anyone have some kind of inside information about what happened to
> Cyc? The project clearly failed massively to achieve its original
> goals, but they won't admit it and say what went wrong. I've been to
> the website of the corporation that is theoretically selling it
> (cyc.com) and it's full of upbeat propaganda about how wonderful Cyc is,
> and how it'll be real AI any year now, and we can sell it to you now,
> but we won't tell you what it does. There are also technical papers
> there, none more recent than 1998, and none giving any hint that
> anything was less than hunkydory.



This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:36 MDT