00100	.PAGE FRAME 45 HIGH 69 WIDE
00200	.AREA TEXT LINES 4 TO 42
00300	.TITLE AREA FOOTING LINE 45
00400	.DOUBLE SPACE
00500	.FONT 1 "NGR40"
00600	.FONT 2 "NGR30"
00700	.TURN ON "%"
00800	.EVERY FOOTING(,%1{PAGE}%*,)
00900	.NEXT PAGE
01000	.SELECT 1
01100	.CENTER
01200	ARTIFICIAL INTELLIGENCE LABORATORY
01300	STANFORD UNIVERSITY
01400	CALIFORNIA,94305.%2
01500	.FILL
01600	
01700	
01800	
01900	.VERBATIM
02000	
02100	
02200	November 23rd.
02300	
02400	
02500	
02600	Dear Sirs,
02700	
02800	.FILL
02900		Bernard Williams' review of Dreyfus WHAT COMPUTERS  CAN'T  DO
02902	was  very  welcome,  because  workers  in  the  field  of  Artificial
02904	Intelligence [AI] seem  unable  or  unwilling  to  defend  themselves
02906	against  Dreyfus'  attacks on their work, and so ought to be grateful
02908	for a philosopher prepared to champion them.
02910	
02912		However, there is a  slight  confusion  near  the  center  of
02914	Williams'  discussion  that  needs comment, since it is precisely the
02916	elimination of endemuc miuddle that is the refreshment a  philosopher
02918	might  have brought to the laborers in the mechanical vineyard of AI.
02920	Dreyfus'  thesis  is  that  distinctive  human  abilities  cannot  be
02922	simulated  by  a  digital  computer,  and Williams meets much of this
02924	argument head on.  But throughout  his  review  he  makes  use  of  a
02926	distinction  between "simulating a human activity ", on the one hand,
02928	, in the sense of  producing  the  relevant  behavior  in  a  machine
02930	(playing  chess  , say) and "producing the behavior in the way humans
02932	do it" on the other.
02934	
02936		This distinction, though widely accepted  in  AI,  is  almost
02938	wholly  without  content, and moreover was not needed by Williams for
02940	his excellent attack on Dreyfus' much publicised arguments.
02942	
02944		Williams  introduced  the  distinction  (p.   36)   when   he
02946	contrasted  AI  and Cognitive Simulation [CS] :   where CS is said to
02948	take "tips from the  ways  humans  actually  solve  problems".   That
02950	should  of  course have been:    "takes tips from the ways humans SAY
02952	they solve problems" since no one has the least idea how brains solve
02954	problems,  nor  even any clear idea of what it would be like to know.
02956	Given that fact, the AI-CS distinction begins to look rather  weak  .
02958	Next  Williams  says  ".   .  .  for AI researchers the aim is not to
02960	solve problems "the way we do", but just to get a  machine  to  solve
02962	problems"(p.  38).
02964	
02966		By  his  suggestion  that  one  MIGHT try to get a machine to
02968	solve problems the way  we  do,  Williams  endorses  the  distinction
02970	again,  and  is also utterly unfair to the motivation of many, if not
02972	most,  AI  workers.   For  their  aim  is  to  simulate   intelligent
02974	problem-solving  activity as a way of understanding it, not just as a
02976	way of solving the problems.  Moreover , at present that is the  only
02978	way  in  which we CAN understand mysterious processes like thinking :
02980	by simulating them.  It may be second best but it is  all  there  is.
02982	And  when  Williams  writes  near  the  end (p.  40) ".  .  .  and if
02984	construct [an artificial organism] is what we do.  .  .  .   .   then
02986	we understand it" he seems to appreciate the last point.
02988	
02990		But  immediately  after  the  last  quotation  Williams falls
02992	straight back into the muddle in his peroration ".  .  .  This is the
02994	direction  in which in which content can be found for the notion that
02996	a machine.  .  .  .  .  .  .  .  might solve problems .  .  .   .   .
02998	as  we  do".    But  this  is  merely rhetorical, for he has shown no
03000	direction at all.  Indeed, if he had , he would have given conceptual
03002	thinking  a  gigantic  push forward, since at the present time no one
03004	has any serious suggestion as to what it would be  like  to  discover
03006	the  "language  in which the brain processes information" , and hence
03008	how we "really" solve problems"?    To  see  this  one  only  has  to
03010	realise  the  problem of discovering the language in which a computer
03012	was really processing a problem, if one had no more to go on than the
03014	contents  of  the  computer's  registers.   It  would  be  an  almost
03016	impossible code cracking job, even though we know all the high  level
03018	computer  languages  to  which the changes of state could conceivably
03020	correspond.  In the case of the brain we have no such information.
03022	
03024		Williams seems in his  review  to  give  his  blessing  to  a
03026	fundamental  muddle  in AI:   he could have ignored it , since he did
03028	not need it to deal with Dreyfus, or have cleared it up.  Instead  he
03030	endorsed it.
03032	
03034		One  additional point is germane:   Williams refers more than
03036	once to the "result" that any analog process can be  simulated  on  a
03038	digital  computer.   This  is  very  important  to  any discussion of
03040	Dreyfus, who holds(1) that the human activities in question cannot be
03042	simulated by digital means , though (2) they can perhaps be simulated
03044	by analog means.   Clearly  this  "result"  closes  the  gap  between
03046	Dreyfus'  (1) and (2) to vanishing point , and so suggests Dreyfus is
03048	contradicting himself.  .
03050	
03052		But there may be a loophole in  this  anti-Dreyfus  argument:
03054	for like all such formal proofs, the result inquestion holds only for
03056	the class of objects  defined  in  the  proof.   There  may  well  be
03058	entities  or  systems  that  are  analog  devices in the common sense
03060	meaning of the term, though  are  not  covered  by  the  result.   In
03062	particular,  it  is  still  an  open  question  whether  or  not  the
03064	three-body problem can even be recursively  approximated  by  digital
03066	means.  If it should turn out not to be so, then any three bodies, in
03068	virtue of their mutual gravitational attraction , would form a device
03070	not simulable by a digital computer.   It may be a little hard to see
03072	how such a device could be used to simulate  human  behavior  but  it
03074	would   constitute   an   important  theoretical  looophole  in  that
03076	particular argumant against Dreyfus' general claims.
11700	
11800	
11900	
12000	.VERBATIM
12100	
12200	Sincerely,
12300	
12400	
12500	
12600	Yorick Wilks.
12650	
12652	
12654	
12656	
12658	To: The editors
12660	New York Review of Books.
12662	
12700	.FILL