COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	1. Here is an approximate subdivision of AI.
C00010 ENDMK
C⊗;
1. Here is an approximate subdivision of AI.

	a. Heuristics. Search through spaces of alternatives.

	b.  Representation.  What  kinds  of  facts  are  there  that
programs  should  know and  how  should they  be  represented  in the
memory of the computer.

	c. Learning.  This  comprises generalization from  experience
and  also   compiling  rapid  running  programs   from  independently
received  pieces of information.   This last  probably corresponds to
what is now called automatic programming.

	d.  Sense  and   motor  activities.    Vision,   speech,  and
manipulation.

2. Successes in AI

	a.  Understanding of how  far tree search  on straightforward
representations can go in playing chess and proving theorems.

	b. Showing that  it is  possible and necessary  to go  beyond
the  classification  problem  and methods  in  perceiving  the  world
through the senses.

	c.  Identification of the representation  problems as the key
to further progress.


3. Areas of research springing from AI.

	a. Present approaches to speech.

	b. Present approaches to natural language.

	c. Sophisticated manipulation for applied purposes.


4.  AI  has  disposed  of  the  idea  that  the  main   problems  are
probabilistic, it  has disposed  of Chomskian  linguistics, classical
pattern recognition, and the automaton model of the world.

5.  Tools produced by  AI include list processing  languages which is
being used  in  non-AI type  symbolic  calculation, it  is  partially
responsible  for  the development  of  time-sharing  in  that my  own
contribution to  time-sharing  was motivated  by  needing it  for  AI
research.  The  same is true of  my and some others  contributions to
mathematical theory of computation.

6. Major intellectual advances contributed by AI are

	a.  My approach to epistemology,  although I cheerfully admit
that this approach is not widely accepted yet.

	b. The  use  of information  processing models  in  cognitive
psychology.

	c.  The procedure  approach of  the  M.I.T. AI  group to  the
representation of information.

	d.  Showing the  way to  go beyond Chomskian  linguistics and
really put the semantics in.

	e. The  development of  mathematical  logical systems  within
which  one can  work  rather than  systems  which are  good only  for
proving  metatheorems  about.    This  advance  is  also  not  widely
recognized yet.


7. I don't think that  AI as a field has 3 to 5  year goals, although
ARPA  might have such  goals in AI  and particular  projects may have
such  goals.   My   main  personal   goal   is  to   understand   the
representation problem better, because I  still feel it has long been
the   major   bottleneck   to   large   advances   toward  human-like
intelligence.

	Our project has  goals as described  in our latest  proposal.
I believe we  can do vision well enough to  assemble and to navigate.
I believe that in  five to ten  years we shall  be able to  represent
the information  in a simple  narrative like  a newspaper story  well
enough  so  that   a  program  will  be  able  to  draw  the  obvious
consequences from such a narrative.   I have been interested in  this
problem since 1958, and my present  optimism stems from the fact that
many people  such as Schenk and Winograd and Sandewall are converging
on this problem.


8. Medium term goals (10 to 20 years) include:

	a.  The  ability   to  automatically  generate  programs   of
reasonable efficiency compiled  from common sense information about a
subject matter and a simply stated goal.

	b. The ability for a  walking or driving machine to  navigate
safely in  a  human environment.   (Flying  is much  easier).   Human
level  ability   to  build  objects  from  materials  given  suitable
instructions.

	These are ambitious goals for  this time frame and might  not
be  met. I  believe they  would  be met  if my  understanding of  the
required intellectual mechanisms is correct and becomes accepted.

	The   long  term   goal  is   to  make  programs   which  are
intellectually as capable as people and therefore  much more capable.
I do not have a  set of ideas for intellectual mechanisms adequate to
achieve  this  level  of  performance.  Therefore,  I  cannot  put  a
time-scale on this, because  I cannot predict when someone  will have
the necessary ideas.   However, there are smarter people than me, and
I am  reasonably confident  that intelligence  is understandable  and
will be understood.


	As you know,  Steve, I will be debating  Lighthill next week,
and  the state  of my  preparation will  be represented  by documents
light.*[2,jmc].   These documents  will be  in their  final state  by
Thursday night,  but there is a  lot of redundancy now.   Your set of
questions has been  helpful in my  preparation. I don't  mind if  you
show this memo to people in its present state,  but I imagine I could
think of more important points if motivated and given time.