00100	Dear Dr. L:
00200	
00300		At a recent meeting of leaders  of  projects  concerned  with
00400	artificial  intelligence  (Minsky and Papert from M.I.T., Raphael and
00500	Hart from  S.R.I.,  Newell  from  Carnegie-Mellon,  and  Feldman  and
00600	McCarthy  from  Stanford University), Ed Feigenbaum and Steve Crocker
00700	tried to explain a new ARPA policy concerning the support of research
00800	in  artificial  intelligence  and  related  areas.   We say "tried to
00900	explain", because we did not all leave with the same impression,  and
01000	we  agreed  that we were still not clear on what the policy is to be.
01100	However, it was made clear that there is to be a  large  change,  and
01200	most of us understand it to be along the following lines:
01300	
01400		The  general  support  of  laboratories  will  stop  and,  in
01500	particular, proposals like the Stanford proposal that ARPA  has  just
01600	funded will no longer be accepted.
01700	
01800		Instead,  ARPA  support  will  be concentrated on a few tasks
01900	with goals that can be realized in a specific time, and  laboratories
02000	will be asked to undertake to realize subtasks of these tasks.
02100	
02200		Management   of  these  projects  at  all  levels  is  to  be
02300	emphasized.  Specifically, a larger ARPA staff will monitor the work,
02400	advisory  committee structures to help them will meet frequently, and
02500	laboratories will employ lower level management  personnel  for  each
02600	subtask.
02700	
02800		All work performed in the laboratories with ARPA support will
02900	be directly related to the subtasks assigned to the laboratory.
03000	
03100		The models cited for the new policy were the  speech  project
03200	and dendral.
03300	
03400		In  our  opinion,  the  new  policy  will  have the following
03500	consequences:
03600	
03700		1. Universities will have to look elsewhere for  support  for
03800	basic  research  in  artificial  intelligence.  Such support can most
03900	likely be obtained from NSF but on a much smaller scale than ARPA has
04000	heretofore afforded and after some delay.
04100	
04200		2. Universities will eventually have to give up competing for
04300	the  subtasks  of  the  tasks  undertaken  by  ARPA.   We  base  this
04400	conclusion   on  the  following  facts.  Organizations  supported  by
04500	contract research initiated by a contracting agency  rather  than  by
04600	their own unsolicited proposals or as laboratories with substantially
04700	self generated research programs require a different  structure  than
04800	university laboratories have.  University laboratories are managed by
04900	professors acting as principal investigators with a small  amount  of
05000	professional management help.  Research organizations like BBN or SRI
05100	or RAND or Lincoln Laboratory are run at  the  top  and  for  several
05200	levels  down  by professional managers who have the responsibility of
05300	getting enough contracts to keep the organization busy.  Overheads in
05400	these  organizations  are  usually  about 150% of direct costs rather
05500	than the 50% typical of universities.  This reflects  the  management
05600	costs  and  also  the  need  to  have  overhead accounts that support
05700	personnel between projects and while  soliciting  work.  Since  these
05800	organizations typically prepare elaborate proposals for projects with
05900	a low probability of acceptance, their costs in this area  are  quite
06000	high.
06100	
06200		Occasionally,  universities  have  tried  to  operate in this
06300	mode, but almost all have  abandoned  the  attempt.   In  particular,
06400	Stanford  University  will not allow it, and M.I.T. discourages it at
06500	least on campus.  I don't think Carnegie-Mellon does it either.
06600	
06700		Certainly,  this  mode  of  operation  is  well  suited   for
06800	government  initiated  projects  where  the  goals  can be formulated
06900	precisely and are better understood by the government personnel  than
07000	by  the  contractors.  I can also see why ARPA feels pressure to move
07100	in this direction, and I see that the support of  basic  research  in
07200	information  processing by ARPA has been somewhat of an anomaly among
07300	its programs.
07400	
07500		However, the government initiated contract  has  not  been  a
07600	successful  mechanism  for  basic  research,  because  the government
07700	cannot get enough of  the  talent  required  to  do  the  initiation.
07800	Modifying  the  mechanism  by  giving  the  power  of  initiation  to
07900	committees may mitigate the problem, but  it  won't  do  a  good  job
08000	either,  because  new  things  start small and unheard of, and policy
08100	committees concentrate  their  attention  on  what  is  familiar  and
08200	salient for their chairmen and the majority of their members.
08300	
08400		The  problem  of  supporting basic research in an agency like
08500	ARPA is  not  relevance.   Committees  or  government  personnel  can
08600	evaluate  the  case  for  relevance  to the agency's mission that the
08700	laboratory makes.  The problem is initiation.
08800	
08900		Well, it was probably inevitable that ARPA support  of  basic
09000	research in computer and information science has to end, but a number
09100	of questions arise concerning the  transition  to  other  sources  of
09200	support.
09300	
09400		1.  Can  we continue to use the computer facilities for basic
09500	research?  Perhaps  the  older  facilities  could  be  given  to  the
09600	universities.
09700	
09800		2. What is the planned timing for the phaseout?
09900	
10000		3. How rigorously will the criteria for inclusion of all work
10100	in the subtasks be interpreted?
10200	
10300		4. Is there a possibility of some basic research support even
10400	if on a greatly reduced scale?
10500	
10600		5. We think that some university participation in the applied
10700	programs being initiated will  be  able  to  continue  for  a  while,
10800	because   these   programs   are  based  on  work  initiated  in  the
10900	universities, and therefore a number of laboratories are well set  up
11000	for  it.   Is  there some possibility of relaxation of the management
11100	structures proposed and the intensity of  the  committee  work  while
11200	this is still going on?
11300	
11400		Clarification  of  these  points  will  assist  us during the
11500	transition  period.   The  ARPA  program  in  information  processing
11600	techniques  has  been a big help in getting computer science research
11700	started in this country.
11800	
11900						Sincerely yours,