I spent some time with Lenat's crew in Austin circa 2000 as part of my deal
with HP's "Internet Chapter 2" project ala "eSpeak".  In tha tdeal I rather
insisted that I be permitted my own resources to pursue some fallback plans
in the event that their primary "vision" failed.  What I was interested in
pursuing with Cyc was an application development platform based on first
order logic, utilizing their extensive "ontology" as a standard library.
It was actually a really great potential that could have headed off all
kinds of brain noise in the industry since then.  I even got one of Paul
Allen's quantum computer language designers from Interval Research involved
to revive Russell's Relation Arithmetic toward that end.  The root problem
with FOL as a programming language is similar to the problem with
functional programming.  Time is treated as a second class citizen.  The
quantum core is time symmetric but at least it is a solid foundation formal
foundation for introducing time as a first class citizen.

Lenat went back to HP HQ with me to do a presentation on their "ontology"
but the folks running the project weren't really serious.  They were more
interested in importing H-1bs thatn delivering on the $500M got dumped on
things like 4 page center spreads in the WSJ, etc.  I had to threaten to
resign to get the Interval guy when the management said I could only hire
H-1bs.

The big practical problem with the Cyc engine was the lack of incremental
tabling -- a problem that didn't even get addressed in Prolog until some
years later with XSB.  We had a Prolog engine company from UK on staff
along with Guido, but I couldn't peel off enough resources to get the
incremental tabling done.

There's still some potential for Cyc I think, but only as a programming
platform -- at least until a more principled approach to logic-program
synthesis work along the lines pursued by Richard Evans
<https://arxiv.org/abs/1910.02227> advances to the point that it can bring
the Cyc ontology to heel within a principled machine learning platform.

On Thu, Feb 20, 2020 at 9:36 PM Matt Mahoney <[email protected]>
wrote:

> On Thu, Feb 20, 2020 at 7:40 PM <[email protected]> wrote:
> >
> > On Thursday, February 20, 2020, at 7:02 PM, Matt Mahoney wrote:
> >
> > Cyc was a failure. But it is interesting that in 34 years we haven't
> found something better.
> >
> >
> > How so? Is Cyc that good? Show me, I've never seen Cyc in action.
> 
> I have, and it was pretty bad. But coding millions of common sense
> rules by hand in first order logic made sense when computers had 64 KB
> of memory and there was no internet. Cyc might have been useful with a
> natural language interface, but that would have been millions of more
> rules, and we know how those attempts have gone.
> 
> Now that we have terabytes of text and the hardware to process it,
> statistical language models make sense. But that still isn't enough.
> Half of what you know (about 1 Gb) is knowledge you were born knowing,
> encoded in your DNA. A lot of that knowledge is still likely to
> require hand coding.
> 
> --
> -- Matt Mahoney, [email protected]

------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T887757e45bfd1342-Mdc3b313e3dfa803d5b1bd584
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to