Letter to Roger Schank

On Sun, May 29, 2016 at 10:32 PM, Douglas Miles <logicmoo@gmail.com> wrote:
> Roger Schank wrote:
> I started a company called Cognitive Systems in 1981. The things I was talking about then clearly have not been read by IBM (although they seem to like the words I used.) Watson is not reasoning. You can only reason if you have goals, plans, ways of attaining them, a comprehension of the beliefs that others may have, and a knowledge of past experiences to reason from. A point of view helps too.

I agree with nearly everything that you said  in your well-written ( scathing )  about Watson .  I appreciate your points of view.    Why do we not have voices like yours within the realm of public understanding?   Sadly (I'll confess) over the years most money I've gotten for AI research is becasue I've held my tongue and went along with flow knowing  it'd eventually lead to getting a chance to do things the way I knew they needed to be done.   Well I was wrong.. going with the flow just leads to more going with the flow.

Any how this email is about you (not me):
I tend to create my own story of the history of AI (we all do) and you are an important part of that story.  Forgive me for exaggerating or misrepresenting your role in some parts  of the following narrative:    The  first of many books of yours that I read (when i was 13 years old in 1983)  was the one published with Abelson (SPGU).    towards the end of the book either I came to (or you came to the conclusion) that the  software that you outlined (involving SAM/PAM) that was capable of understanding  complex scenarios  was held back only by the lack of scripts.  I know that these "prewritten" scripts were intended to be very generalizable and we needed only  a  trivial number (though  a large  enough  number  to offend the anti-"a-priori" crowd  at that time *smile*) ..    I don't remember if it was specifically said in the book ( just due to the time period)  but undoubtedly  we could use hierarchies of generalizations to make just a moderate number of scripts applicable to several different types of scenarios ( later on  people call this ontological engineering  (where we craft methodologies for dealing with the  various   idiosyncrasies  of how generalization could be used to create equality)).    You were part of a a handful of scientists whom had reasonable theories of machine understanding (1970-1983) .  But most theories seemed to have a some things in common.   Need to deal with search sometimes on a  PSPACE complexity.  Needed to planning  switch between between representations.  Needed  blackboard  that could hold an undefined amount  of knowledge structures.  All needed to be easy to reconfigure ( to experiment with ) and the ability to combine various modules.

  I felt  was the energy  behind the viability of the CYC project.   Meaning I believe that CYC was funded to test *your* theories out.    ( I doubt  that Doug Lenat's version of history  coincides with mine )..    For example PAM  and SAM  have strong overlaps in types of knowledge structures they consume .   and more importantly produce.     Meaning  we needed an architecture that was capable of supplying your modules  with  the type of  generalizable /subsumptive  knowledge structures  that could be worked with over a long period of time  and from the many modalities, contexts and points of view that might be needed.     Not only  using your  current theories (at the time)  but future theories.  And of course from

other

 scientists to!   Like even Lenat!  Though before he can test even his own some PSPACE computational problems have be solved.

(Combinatoric explosions)


Also there is a learning curve in what does it mean to produce the CycL language.  Not sure how familiar you are with CYC but some examples:   20 years in a mass scale change of separating WFF rules (VocabularyMts) from GTheoryMts.   30 years in a mass merging of #$hasAttribute with #$isa (Attributes into Collections) (Simular to merging KL-ONE  Roles and Concepts into one thing!)  a massive

addition of a parallel knowledge like the


concept of "*-Underspecified" versions of the most vital parts being cloned

into their own knowledge bases


.   My point is even Lenat can predict at least a couple more massive rethinks to the system...

BTW... I'm not referring to CYC as people like to describe it (sadly it is as conflated as Watson)  My prescribed use of is less about modeling human understanding

From Lenat's point of view


and more  about using it  as a general  blackboard.

( Like KnowledgeWorks )


CYC is merely an ideal "content control system" for the types of research you've written about.  For example we can create a section of the data that holds goals  of  each agent in various scenarios are represented  homoiconicly  (I just mean readable by the agents themselves) and able to be access by various modules.     The main focus of my work is the provide infrastructure in which we can logically (explicitly) describe the information (that is forward chain data to and make calls into/between modules that involve planning and current state recognition (as in how the state relates to the goals of something). In other words to put some of  your theories to the test :)


 What really bothers me (  fists shakingly! )  is that we have not yet  disproved or even tried ( to my knowledge)  to continue the experiments that you have outlined  on a scale that we seem to  have available with Watson.     Or have we?


Thanks in advance,
Douglas