Re: Development models.

From: Stephen Reed (reed@cyc.com)
Date: Wed Sep 07 2005 - 12:03:32 MDT


The Cyc Project is the largest pure-AI effort, and although most of this
audience disagrees with our assumptions, approach and accomplishments, I
would like to share some of our development model in the hope of informing
skeptics.

Starting as a sponsored research project at MCC, Austin in the mid-1980's,
Cycorp has been staffed at approximately 50 persons since I joined the
company in 1999. We migrated early to open source tools with the notable
exception of using Allegro Lisp from Franz, which enabled us to move
cleanly from Symbolics Lisp machines to whitebox Linux workstations (about
80 in house, one per worker and the rest in server pools or
administration). By 1999 the programming team (about one-half of the
workers) built our own lisp runtime in portable GNU C. This investment
was substantial but allows us to deploy Cyc without the Allegro runtime.
In the last three years, our client applications have migrated from HTML
to Java. We had reached the limits of what Javascript could do for the user
interfaces.

We use Bugzilla to manage our detailed tasks, and use Cyc itself for
project management. Our tasks are oriented around the particular needs
of our various sponsored research projects, but our management tries hard
to align our tasks to enhance our core capability and make progress
towards our AI goal.

Here are some of the specialties up to this point:

1. Operating systems level programming expertise to maintain
our runtime infrastructure, and translate from lisp into efficient C for
deployment.

2. Lisp programmers for server-side development including object store,
inference modules, database integration and natural language parsing.

3. Java programmers for client-side development including GUI's and
enterprise/semantic-web integration.

4. Inference programmers who know about deductive, inductive and abductive
inference and know how to efficiently support inference in the Cyc
knowledge base.

5. Linguists (also programmers) whose specialties include natural language
lexicon, parsing and generation. We also use third party and open source
NL tools.

6. Philosophers who know how to represent commonsense objects and their
relationships in Cyc's predicate calculus language CycL. They know how to
formulate facts and rules to get commonsense questions to answer.

7. Managers who are familiar with artificial intelligence in general, who
have detailed knowledge about the needs of our sponsors, and who can herd
the rest of us like cats to get projects delivered as contracted.

8. Proposal writers. Among our managers we have some skilled at writing
winning proposals. I attribute our success to our sponsors' satisfaction
with our performance and deliverables from previous projects.

9. System administrators who can keep state-of-practice Linux boxes
operating smoothly, with a shared filesystem and secure network.

10. Human resources, finance, general administration and so forth that one
would expect from a company our size. We are quite proud that our
overhead rates are very low compared to other government contractors.

In summary, before 1999 I too had hoped that a single motivated person
could combine the best published research of others, together with open
source software components to construct an AGI. Now that I've been at
Cycorp for these six years I believe that a large integrated team will be
required because specialized knowledge is key. Any one area of
specialization could entrap the sole inventor. It's for this reason that
published AI research is so specialized.

Hope this is helpful.

Cheers.
-Steve

My statements are my own opinions and do not represent Cycorp in any way.

 --
===========================================================
Stephen L. Reed phone: 512.342.4036
Cycorp, Suite 100 fax: 512.342.4040
3721 Executive Center Drive email: reed@cyc.com
Austin, TX 78731 web: http://www.cyc.com
         download OpenCyc at http://www.opencyc.org
===========================================================



This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:52 MDT