coMPUT'll'iTELLv o L .3 . t 9 t 7 I 14 cccdings of rhc ?th IntcrnationalJoint Confercnccon Anificial t n t c l l i g l n c c , V a n c o u v e r8. . C . , p p . 3 3 1 - 3 3 7 ' 1981. Narural languageProccssint:crucible for compuu' rional theories of cognirfn.-Procecdings of rlc 8th Intcrnational Karlsruhc'WcstGer' on AnificialIntclli-ecncc, JointConfercncc p P . l l 8 0 1 1 8 6 . many. S r - r c u i ,i . n . t S Z t . A n i f i c i a li n t e l l i g c n c teh: e h e u r i s t iPc r o g r a m McGnu''Hill. Ncw York' NY' mingapproach, Two lessonsof logic Bmrx Cexrrvett Stqlrn PaloAIto'CA 94304'1J'S'A' Center, XerorPalo AItoResearch ond CA 94305' tl'S'A' SranJord' Centerforr,c srud.rof LanguagcandInJomatiotr,SrartfurdUniyarrig" C m p n . l n r c l l .3 . ? l { - 2 l t (19t7) In particular bc eleanedfrom is rulesand reprcsenlalions' l. Background results (incompleteness ."ri, thit canbe madequiteconcrete daredto 1879'rThis means for arithrnetic,for example),but the lessonitself is gencra'l' Modernlogicis mostreasonably toform' hadmorcthana cenluryto studya Panicu-- For discussion, of content tharlogicians-have I'tl call itthe irreducibili4' m.uclt surprisingly' Not systemsindependent formal so-called somewhat of two f"triifv to distinguish iui To getat it, we need of anysymbolsystem' of the enormousamounttheyhavelearnedis peculiarto their uroi.tt or dimensions about rhc foundationsof mathematics' rpr"in" assumptions WhatI'll call the,frsr/actor of a symbolsysteminvolvesthe however,areunivenal.holdingfor any shapesof rhesymbols,thewaysthey canbe put togetherand Stme of theirinsights, or sentences. symbols, signs, of system mkin apart, and tlre behaviouror operationsdehnedover ' logic's of many that to see come Particular letlers,and modusponensin has predicate McDermott them. Exprcssions, By humanreasoning' in comto general operations aren'tapplicable corresponding and ryP€s dau assumPtions Iogic;absiract - thatsonof thing.You canthink of this*'hole and laige I think he's right, andI aerecthattheconsequences pulerscience a combina' system'smechanics: "t a tiitte appalling.Bui McDermoitgoeson to imply f"t y.' tackaeeasa rcprcsentational AI' for basis a as least whole. al ovcr defined as a logic ntions rcjeir ope dynamic and mustthercfore iion oiitt staticsrntctures and AI bctween affair in a love realized end the dircctjl' what bc suggests,-to must he time, is also factor It's rhem.Ttrefint work' any do going to is if the system logic. substrate, -No* "logic" 'physical t:q' mean' l I don't carc too much about the term s1'mbols what the the s)'mDols with what to do do with has'to second Tlre Tlresecond facror has factor wherherwe shoulduseit, broadlyconstrued'for the full range what they'rc aLout- their content.Interpretation(in thc.logi- | of rationalbelief revision(roughly,\'hat you shouldor arc cian's sense!)is a secondfactor phenomenon,as arc truth ano I wc likely to belicvenext, if you bclievel- no.*), or whetherrefercnce. the latter in the senseof the rclation bdwe-enthc t to it confining logic in "McDermon" and a personwho works at Yale' Shadcs rto"ia foUowtraditionalmathematical name i thc enrailmentrelation(roughty,what follows, if P is true)' of the old declarativc/proceduraldistinction (bur only shades: J andadopta moregeneralretm iot thehumancasc'(As I say'I 198?). I'll usc sec Smith ultimatesense,but in whatfollows don't carc, in som-e logic dretwo storiesarecalledprooftheory mathematical In "thought" ''togic" for thenarrow!entailment for serse,and and modeltheorj, rispcctivcly;semanticsis talicnto bc thc - *re storyabout what we do.) lanir. Funhermorc,semantics we can srudyof-rhe What I do careaboutis this:thatwc learncvel,ything s1'stemgin this these about maners wha( rcally is .on,Lnt "more equal" thanthefirst factoraccount'There'sa history.Morespecifically' from those100yean of intellectual it's sense I worry thatMcbermon,in rejectinglogic'sparticularassumP- reasonfor thisasymmetry: Witiout somesecondfactoraspecu' tions, is also discardingsomeof its universallessons'Tu'o you couldn'tbc sureyou hauea symbolsystemat all' Everylessons,in panicular' lhing t'"t a mechaniialnature'to put rhis anotherwa1';it's prior (that'swhy ! called-itfint)' Havingconto meriphysically I.l. Issson one: theirreducibiliryof conrent form to a and thereforebeing amenable The firsr lessonI take to be the deepcsttnrth that logicians tent,'on rlre other hand' rEPreor symbolic is whatdistinguishes havc uncovered:thattherc'smorcto a symbolsyslemtlrancan secondfacrorana-lysis, (what philosophencall intentional)systemslikc sentational and compulersfrom ordinaryphysicalobjeculike ''Thc ycar 1879wasrhepublicationdateof Frcge'sBegrifsschrilt' lansuages -pucks foundation logical a hoc-ky develop to 'ntst and oii refineries. thc firsisrep in his life-longancmPt lessonof logic.then,canbe suted in termsof the ffre for marhcmatics.AhhoughI will arguethat wc shouldreuin thc "second wo* Frcgc's that rigour and precisionof ihe enormoJstradition factor": the cJntentof a symbol,at lcastin general' it' butarises proximate inipircd. I bclievcthar the inruitionsand insighuof variouscarlicr isn'tan inrrinsic or causally ProPen)'of Peirce.art rt For domain' espcciallyCharlcsSanders other wrirerson symbolicsyslems. and some iyirem thc between asa relation in Icast as rcic.'"nr to AI's .oni.-s. From our point of vicw it's ANE'?' Pt of thesymbol or reference .irrpt.. rhemeaning durdcvelopmcnt. unfomrnatctharthe vastmajoriryof mathematical er LaGuardia' or rnffic air of ihis morning's purcly"mathe- "n axiomatization ing rhis cenruly.hasbcendevotcdto tbc narrower, ol 20@ft up' No amount aiqplanc. actual some would involve maticel." case. PnmJ o C:nrb , lmnnc ru Crnr! COITTTIEYIARIES i n v c s t i g a t i n gh o w t h e s y m b o l P L A N E , ' i s u s e dw i t h i n t h e a i r ' traffic iontrol syslemcould ever tell you what planc it refen r o . T o 8 e l t o t h e p l a n ei t s e l f y o u ' d h a v et o l o o k o u t s i d ct h es y s t c m . t o s e e h o w i t w a s c o n n e c t e du p t o , a n d u s e di n , i t s e n v i ' ronment. Or imaeine tr;'ine to determinethe tnrth of a repon claiming that 70% of all doctorsrecommendCrest toothpastc' You woutdn't bring out your microscopeto studythe paperthe regort was written on. or submit it for rypognphic analysis; around and acruallytalk to doctors. 'vou would drive Cont.n,, in other words, doesn'thang arounda symbolsystcm, like a nervousleenaserafraid to leave thc house;it's out thcrc, in the world. Nor can the secondfactor be deducedfrom thc first. Funhermore (this is perhapsthe most surprisingthing of all) this externaliryof content doesn't arise only for rcalrime iystems, like air-traffic control systems.It holdseven for as absiract, circumstantially independent,causallyinen, and completely disembodied a system as formal arithmetic. that contentrela' Ir's a major corroltary to this first lesson "Povungnituk" to tions arcn't computed. lf I use the name rcfcr lo a small town on Hudson's Bay, for example,a content relation holds between my uneranceand a major sourceof Inuit stone carvings. But. jusr like the Propertyof being the avemge age in a collection of pcople, this relationjust is; no to hold- Adminedly' in *ork-neeJs to be done in order for it "comPutc" sometling, you may unefiInce. my interprerine but the purPose of your comPutation will only bc to arrange yourself lo stand in something like the same kjnd of (nonlompured) content retation to the town that I did, when I said it. Tirc town itself. which is a part of the contentrelation,is not "proximate" cause a part - not what the philosophen call a . - of any comPutationalactiviry of saying or undentanding' In conirasr with the first factor, to Put this anotberway, the second facror of a symbol syslem doesn't needdirect physical realization. Therc is no way the Inuits could deploya sensorin Povungniruk to detect whether their town was being rcferrcdto by an arbitnry speaker in an arbitrarily remote location' It should beadmined that how this dl works - how symbols "'reach out and touch someone" - remains an almost total mystery. Some people (Winograd. for example).argue that thly only do through human use; others (I'm in this category) belicvc that human interpretation is sufficient' but no( neccssary. But whatever one's view, the facts that theseviews have to dcal with are impressivc. To sun with, refercnceoutstriPs causaliry, at lcast locally; with four simple leners I can rcfer lo e composer who hasn't existed for more than 200 years not to a set-theoreticmodel of him, to his name, or lo your comprehcnsion of him. but to his very hean and soul (eventhough' untite Povungniruk. he doesn't exist any more)' Referencc relations arcntl even constrainedby'the light-cbne;as Church "sPeed of logic"' In fact oncc pur it,r semanticstravels at lhc refcrence almost outstriPs comprehension; if we didn't know that tanguage works, we would sPurn rumoun of its long. distance capabilities. Whar's more. the linle we do know isn't reassuring,at least I for AI. To take just one example, it seemsthat the axioms of ' arithmetic musr be connecled to thc numbers-(ratherthan to I other nonstandard inlerpreutions) not by anyrhing intrinsic to I them. but by us humans. It seems,that is. that agentsarc what I mancr. for semantical connection. But lhat only shifu thc i mystery - squarely onto AJ's subject maner: cognitive agen1. i interprcrers (igain. in the philosophical sensel)of the symbols :CSLI scminar.l"lay 3. t98{. ! E _q 1 l \ a n ds i g n s . But if we don't know low rcferenceand content g.ork. ar leastwe know tAal they work, and that therc is more to it all rhan proof theory, Funhermorc, I take it to bc the job of y n d e r s t o o dt o. e x p l a i n .i n a s s e m a n r i c sa.t l e a s l a s c l a s s i c a l l u way a as. possible, the interplar' rigorous and systematic "formal" (i.e., fint factorr) propenics.on thc onc betwecn rela' hand,and theserclatively more mysterioussecond-factor tionsof meaningand content,on the other.Admittcdll', in the face of considerableignorancc,we can't yct fill in all the details(though we do what we can - which, to date, mostlv meansenumeratingtlre relata, but somedaymore should be possible).What matlers- what logic's ftnt lessonrc:Iy tclls us - is that the second factor conlentstory must bc told. ].2. I-essonneo: a single theoreticalsrance Logic's second lesson is a theoreticalone, in the scnseof being about theory - about how symbolic systemsshould be expliined. Related to the ftnt, it arisesfrom the recognition theory and model theory, in the that the two factors (p*f related, in spiteof bein-econceprua.lly be must case) traditional distinct.This is the role, in logic's case,playedby completenessproofs,notionsofsoundnessandvalidity,etc' ln conLrast, you could aJsoargue that therc are two storiesto bc told about money: one about is physical embodiment,one about is socialandeconomic import. But, at leaston the surfacc,there is no obviousreasonwhy the storiesshouldrelate;engineenat the FranUin Mint dcsigning new dollar bills probablydon't need to k-now Gresham's law (that bad money drives out good).In logic, however, the connectionis morc direct' You Iouldn't really claim to have a (lirst-factor)inferenceregimen if you couldn'r relate it Pretty directly to (second'factor) semanticinterPrctation. Becauseof r-hisglobal but crucialconnection,logicianshave developeda single theoreticalstanccfrom which to tell both stories. The stories, furthermore, overlap in vocabulary: the samerlieoretician'sgrammar that spcllsout the linguistic regularities of logical formulae is usedby proof tieoretician and model theorelician alikc. And thc overlap is necessarl"That logic's two facton arc rclatively independent(more on this in a m6ment), and yet must ultimately be related,can onJybc said from a vanhge point from which t}ey can borh bc scen' In laying thcse things out it is importantnot to confuscthe .on..pru"l-distinctnesi of facton, or the singlenessof tbcorttical sranapoint,with the questionof how relatedthc rwo factors arc. By inalogy, Seometrydistinguishestengrh and area. but thengoeson to tie them stronglytogelhertin the familiar rr'ay' Simil-arty,both classical and relativisticmechanicsview rime and spaie as concePruallydistinct; the two rheories diffcr in what *rey then say about tlre notiors - whetherthey are inde' pendent(the classical case) or intimatelyrclated (relativistic)' io th. n"*t section I will claim that the frnt and secondfactors of rhought should be intimately ard constantlyrelated.but it doesn't follow that I rhink they are the samething. How, then, does the second lesson relate to thc first? just Becausewe'rc now talking about theoriesof logic. not penicuIn aboutlogic iaelf, things gci a linle bit complicatcd. lar, sinCe *rese theories Qike rhe logica.l syslems tbcy arc abour) are themselves intentional phcnomena,we haYe rwo tTheclaimthat formaliry,in rhecnd.rcduces to firstfactornotiors by B'-C' book in.prepantioo a in .tgu.d is of physicalrcalizabiliry bv Thc Sm'irh.cntided Is CompurarionForma?(To bc Publishcd MA.) MIT Press/ABndford Book. Cambridge. 216 voL. l. l9E? lr.,TELL. co\{PLrT. of representa' (tu'ospecies andmodellinS 3, Tharlaneuage iifferently' Thg completely rion,I uke iil sh'ouldbe lreated malter primary subject the is relation reference lineuisric familiar such engendering nontransitive, to be strictly asiumed of meta' hierarchies distinction, astheuse/menrion conslructs (of otherhand on the T' Modelling, andconvention languages, quadruples' of sets the-soi that treatsTuring machinesas iruth andFalsiryas0 and l. andso forrh)is not onJytakento "frcc," in thesenset'hatyou buialso bctransitive, 3P 4l-o*:d io rr. t modelof X in placcof X irself(evento identifythe two) with theoreticalabandon' "Tarskian I don'tkrow exactly*'hat McDermonmeansby sincehe ctearlyintendsit to be broadenoughto semantics," (on languages of programming analyses includedenotational whichmorebelow),butltakeittomeanroughlyasemantic At anyrate accountthatadoptsall threeof theseassumPtions' herc. I'll usethatdefinition 2. Human cognitionand logic'sassumptions positioncannowbe statedin termsof thefint McDermon's what two assumptions: of much with agrce thatlogicmakesthem'thatAI that I indicated he recognizes already I'vc pas' of this over consequences deduction purc mustrejecir-hem,and thai therheorerical McDermon siys: thatrationalityisn't of aspectinextricable use an is that aredaunting' rejecdon ti* itgi""l forrnulae, for the strengthof his all tharstuff. But let'sgo a linle slowly' to seejust i "gr... I've also got real synrPathy f"o*iJg", 'llet go of ' or a]tered'as a to go back let's easily be In can lead. theie of agreements none rcact[n: Panicular, whcrc th-esc if ,h", *"r" a minor adjustmentto what remainsbasicallya oit --erin historY. "logical" enterprise. soassumpmodcrn of Thereare foundational tugg.tted at the outset,the originators furely "TJrmal" logic - Frege,Russell,Whitehead'Carnap' iions,with all thatthat imPlies. called To makethat concrete,a few more wordsabouteach'To ^nJ to on - ',".rc-primarilylxploring(or at leastmotivated conthings All mathematics' of foundations thc in start with, the central intuition underlyingthe -so-called ty) issues iisiru.t.d';language projectat CSLI involves andcomputation sidercd,rheYdid an excellentjob. a co.m' descentheir We, exactoPPo.site: with its the linr assumption replacing Unfortunarely,though,theyalsodied' we're to develop is goal $tat The use. achievement langu.g. by their *in.o ^io dircctfocuson dcnts,have beenso impressed mustbc of thinkingthattheydehnedwhatsemantics ne* rheori.sand semanti.J fome*orks that analyzeindii;;;g.t question the to related thecrucialroleof circumsunce andembrace a t,ro-stag"'ptoblem, raises "idualunerances, itit iif.. and parricular andcontext.To takejust oneexample,thisinvolvesdiagnos*i*O ", rhe ourselabourihc-relationberween their facton: accePt to liable we are of all peninentcontextud ine the relevantstRrcture unir"oA insigbts.At fint blush, "lt's 4:00 why us wherc tcll to place (such the around ai them having facts too-glibly, not t i"urnt background o.*Jt 'tlr;;;it (thal strucrure djscourse theieciiions theydid. Then,oncewe discover*rat p.r.;' o* tl"id;, pt"tuppositions, we'rc for our Purposes' i,elpresolvePronouns,foiexample),factsaboutthelanguage O.ir p"ni.ular choicesarc untenable of rlrCsubjectmaneror describcd in a^ig.t of throwingthe wholething away,babycumbath- U.i'ngus"d, ih. tt..tut which linguisticand cognitivcprowater, " situriion,with respectro that werc made' that choices murualbclicf stnrctures all the canvass strucrured, to rurn'be ";" in i"ti,.t rhantrying ;;;; cogniforabout facts traditional internd underlying said, can't be assumPtions and three can liit what explain lcimc simply of internal for AI' The following'in t.i"i ^Ict,i,..trc that penain to the interprctadon rnJ togi. tf,"t I U.ti"". areunienable just onesetof is this and so on ind so forrh.And structures, otlrer words, arc tenetswc must,gq,t: dozensof had has CSLI be considered, iit".i ,rt"i have to l. That use can be ignorcd' This premise leads logic to yean so far - and' from 4 project for this on people oorking 'riiitt. ignore agents and procesiing, to setaside-context,andto focus In peripectiue''lttrint it's makinggoodprogress' It unerances' individual or on sentence rypes tnsteadoftokens subsomething bc nJ, rot thananother4 yean thercshoutd also suggeststh"t a s.nt.nce must rePresentis whole content stantidto show. could sincc no other resourcesare licensed that .ifu"iif!, - thatfirst andsecondfaclorsarc Tlre secondassumPtion nafural from -"'f." ",i.,it contribudons. This is quite different just asdeeP;I.alsothinkthisis the goes IocJiy ina.p"ndent where dynamic and contixrual factorsoften implic' i*gu"g", He claimsthattheculprit sobereJM.D.*on. so l"r ifirt has for itf .o""ttiuute to the content (the time of unerance' only to prov' (inference "nowl')' deduction of notion its is loeicism of "a""t.q"ences), for the word "i'^*pf", provides an inter,preration connection semantical but if a different rUf. Z. if,"i iocaily the two symbolic factors can be treatedindca way such in defineable (sav. were semantically abduciion) pcnOentty. even though (as suggestedabove) they must ultisemanfrom cleft be (fint could factor) role procedural ii'"i it. marely bc globally related. In panicular, proof. theory or senseto tical imporr(secondfactor),thenit would still make (secono or semantlcs model theory (fint and factor) inferencc putathe second "at the ;"'"-,hi"Bt iown first, andbuildProgTms factor) are tied together, for any given system',only are subtleties: thesc But enrttprise' theorems'From step r;". .ts.n-.. of thelogicist end,'i with soundnessand completeness and I don't believe indepindencc, l6cal assume 'formai" Jo"t i"gi. proceto step, in a Proof, the (first-factor) inference 'thousht that. like is dure iannot dependon or affect (second-factor)semanticinterthisone,hasbeento usecomin attacking "formal" is taken to mean, by V! o*n strategy, pretation. (In iact this is what asa andintrospection puuiionallyinreiialnotionsof reflection I notjust one' Lotsof peoplehave s)'slemsto consider, -Jresrtea symbol o'irh ho* theyrelate:fromTanki, in settinguPP.rc T, to Quinc' worrylng conditionson satisryingConvention But theovcrall translation' of "Uort tf,. radicalindereiminacy second The ' enough ilear is factori*.*t -ron,rr, of rheconnection of content the (for eiamplc' accounr rhe rheoretical of complete yd thc include musr papc'r) first 1963 Kripk.'s' (the underinvestiSdtion of thesysrem triLna Torrordime'nsions in logic' modal of ."n,"*. p*f rclations.andmodelstructures is theory the that just say whatir is to That's l{i';k.'I;;.). abin the systemunderinvestigation' --inougt, intricacies'What malrersherc is that a single'uniof both factors'I takethis mustprovideaccounts thcory ficd pointto vantage ,."oenition of the'needfor a singletheoretical great contributions' bc d;'csecondof logic's l-et's turn to McDermon' So much for background. theorisrs as diverse as Fodor and Martin-Lof.) \ +r i t in u'hichto u,orkour virblealrcm;rrives. Thepoinr'of ffi,Wf-lUte t |i€-Lisp.Smith (1981)in prrticular, wasto showrhatevcn\r,ar" t horseprogramminglaneuages arebestundentoodin termsof l o c a l l y i n t c n w i n e d f a c t o r s .A s i r h a p p c n s 2. - L i s p i g n o r e dc o n ( textual aspecls of use (by design) - thereby drawing somethin,eof a distinction between the l'int and secondassumptions t - bur at the same tirne maUn_sits architecturelessgenenliz. I ablc than one mi-ehtlike. I hardly needadd tharlots more *'ork is necessary herc. I' Similarly the third - that lan_euage and modellingare caret gorically distinct. Although McDermon doesn't addressthis t premisc explicitly, it stands in equal need of reconsrruction. c ttglg_lg th eqgcs--gI-IePrEsenIation' a nd c orresno nd ence w i ll t, bc required (see Smith 19E7 for some inirial ana.lyses). Therc a arc two reasons this revamping is urgent: panly to e.rplain c computational practice (see below), and parrJyro clarify our standardtheoretical apparatus.In particular,althoughpromisri cuous modclling may be helpful in answering large-scaleand n hencc rather coarse-grainedquestions(such as whethera given n formula is true, decidable, computable),it can be pernicious Ir whcn one asks fine-grained questionsabout control, inrena sional identity, and the use olfinire rcsources.Also, currenr u computational systems involve representationalstnrcruresof i r a.ll kinds. ranging continuously from linguisricexpressions to b vimrdly iconic isomorphisms like bit maps and simulation structurcs- This is a large area where rhe panicular assumpc tions of math,ematicallogic haveled ro unrenablemethodologiir cal practices (for AI theorisrs),as well as to unrenableclaims ? on our primaly subject maner. All in all- in other words. I agreewirh McDermon drarrhc consequenccs of rejecting these assumptionsare enormous. sr And yes, rhey certainJy underminethe coherenceof the "logift cist" program. Writing knowledgedown in advance,rvirhout P rcgand to use, is a conceptual error doomed to failurc. b 3. \4tat then? and this is really whereI've beendriving- whararc But wc do insread?Insteadof abandoning to hopcandreveninsro rl unconstraincd symbolmongering. surelythi raskis ro develop h si aJternarivetheoreticalframeworks. Now McDerrnon doesn't really argue for pure symbol tc "procedunlisr" paraci mongering,but he doessuggestthatthc digm thc only game is other in rown, insinuaring tharrhere e couldn'l Why be any others. should rhat be true? For example, tc flesh n r why shouldn't we develop a full-scalerheoryof usc Y u l out lhc projecr that the philosophyof scicncehas only jusr rhatmust re staned, for example - and uncovcrthe regularities ol undcrlie integratedconrenrand behaviour?Fromthe factthar rt usc and co{rrentare inexrricablylinked it doesn'tfollow that r-h mtionaliry is random. And if ir isn't random,we can undercreed). d i standit (at lcast thar seemslikc a plausibleinrellecrual Sce, this is rcally whar I think's got McDermon'sgoat. t^ Compurationalpracdce whatour programsacruallydo, not what - doesn't honour logic's threc we about sav them AI as ot assumprionslaid out above;it mixesbehaviourandconrent rh richly and rhickly as we do. The only rigoroussemanrical m theorieswe have, on the otherhand.do-makethosercsrrictive areunlenm assumptions-McDermott seesdratthe assumptions able. and correcdy notes that there proposals aren't orher any in around (" - . - it must have a semantics: so it musrhavea Tarskiansemanrics.because thereis no orhercandidarc"). So he's forccd ro taud pracrice.Bur then in virruallythe same brcathhe admirs rhai rhat makeshim uncomforrible. So he Rq H Obcysthrcc c l a s s i c aa l ssumptions Thcore ticrl fnmcwork Logic and scr rhcory Spcifrc applicat ion A x i o m a r i z a r i oonfs l i q u i d sr.r a f f i c , . . . Rlquircd rorni I, c o S n r t t v sc c i e n c e Qualirrtivcmodelsof physicd rcrsoning Frc. l. Appropriare theorcrical fnmcwo*s. endsup somervhat confused. Whathe shouldendonc (it sayshere)isn.tpracticeitself, bvt theoreticalframero.rks that do justicero'tharpractice'. Thatis, what we want is a conceptuat backdrop in'te.msoi whichto undentandForbus'swork in thesamewayrhatlogic and model theory form a conceprualbackdropio, ttrylsi re.search_on-rhe-o.nto-logy of liquids.This siruation is picruredin Fig. l. To be fair, McDermonsaysa lor of thingsiu.ssestin; thathe agreeswirh the generalthmstof rhisdiagram::.tl,.ri arelargeclassesof programsrharlack any kind of theoretical "AI progralns underpinning:." arenororious for beingimpen_ etnblycomplex. . ., bul a modeltharwe don.rundei:rand is nota modelar all," "whar's reallybothering meis rharthese (diagnosis) programembodytacittheories ofabduction,"erc. Wbathedoesn'rsuggesri at leastsufficienrJy cxplicitly,is thar we needrheoriesto do jusricero programsin jusr rheway rhat logicprovidestheoriesthardo justiceto lmirhemaricalj scntences. The question,that is. is how we are goingro fill in rhe missing quadranr. It seemsrharthere,r* tio o=buious sugges_ rions.We could uke loeic and set theory,andtry ro molify them.Or we could rhrow away logic and s"r iheory, "ni simplysrudyrhe pracdceirself, like enromologisrs srudying becs. With respecr to modifyingiogic andsertheory,I've a,lready saida UnleaboutwharI rhinkwouldbe required ftuild in the opposire of rherhreeassumprions listedabove).And I.ve said it will be hand.I agreewi$ Isracl(and,I takcit, McDermon) that incremenralvariantslike nonmonoticlogic are nothing likestrongenough(Israel1980).Theproblem is rharonccyoi startrcvampingthis muchof logic's foundarions, ir's notciear whatremains.It's easyto say t-hatone mustundersrand just whataspecrs of classicallogic are parricular(i.c., spccifii ro metamathematics), which are universal- bur thai doesn't male it easyro do. So far I've only suggested rwo universal lessons: external(noncomputed!) conrent.anda singletheoreticalvanragepoinr. But therc'sa lot morework to do. I-alsohavegreatrcspeclfor the orlrersuggesrion: srudying and.reconstrucring pracrice.We shoutd..-,rlinty unden-tni architecrurc, physicalembodiment,rcsourceallocarion- all thc usualsruff, In facr rhis is wheremostof my own workhas concentrated. But it is also wberc my origina.lworry comes backro roosr:rhe worry r}at logic's two giar lessoniwill bc lost.ln factrhis worry can be seenas a triplcthrcar. - Fint, becauseit porenriallyconfusespracticeirself wirh t-heories thatdojusticeto suchpracticc,I'm afraidrhatMcDermon's papcr will lead peoplero discardlogic's r-heoredcal stancecomplerely,and focustoo muchon thepracricalside. To be honest.I don'r rhink McDermonhimselfwill do this (he'stoo unreminingly theorerical), but a casualreadercould mistake his inrenrions. Second, if youjust look at programs, andtry ro makesense \ C O N I P U Tl .^ " T E L L ,\ ' O L . l . l 9 E ? 218 Spccificationvicw: (rypical in theorctical computcr science) , Program ,.,nantiar*l I Ingrcdicnt view: (typical in AI) Frc. 2. Two readings of "prognm". out of what they are doing, you will be liable to focussolelyon l'irst-factor aspectsof systems, for a simple reason:The first factor. as we said above, is tlre one that needsto be realizedin thc machinc. And since the point of programsis to conjure up an otherwise unorganized state machine into appropriateform to exhibit reasoning, the program only needsto concentrateon first-factor problems: strucrures, operations, behaviour. ln general, as.we saw at the outset, the contenl isn't in the machine at oll. The third problem arisesfrom a curious fact abouthow practice is currently undentood. It's hard to tell exactly rvhat "program" and "denonMcDermon means, but the words tional semanlics," when uttered in one sentence,ineviubly bring to mind the denotationalsemanticstraditionin compuler science, as illustrated by Gordon, Plotkin, etc. (see, for cxample, Gordon 1979). Now I firmly believe that all current computational systems - from Amord to Zetalisp - blend both factors we talked about earlier. The only factot of compu'ational systems that.computer scienceulks about,however,is ,c Frrst: procedural role, I believe this is true, curiously "semailfcs." What is cnough, even when people usethe lenn "denotational semantics" calleil in computerscienceis in fact a mo'del-theoreticanalysis of first-factor proceduralmlc, for a rcason that depends on an ambiguiry in the use of the word "program," suggestedin Fig. 2. I bave tried to explainthese in Smith (1987); diagrams, and some of their consequences, the main insight is that programs, as understoodin theoretical computer science, arc rypically vieu'ed as specificationsof computational behaviour, whereasAI (McDermon is clearlyof this view) takes them to be constituenrswithin it. On the fint (computer science) view, the content of a prograrnis the computation specified;hence, the computationitselfis what denotational semantics takes its subject mauer lo bc. Note, in passing,that all sortsof otherotherwiseodd things are explained by this analvsis.To begin with, it explainswhy pmgramming language semanticistsuse term models,initial a-l-ecbras,and all the resr: they individuatetheir semantical models (remember the rhird assumpdonl)so finely because they are really modelling (the behaviourof; computational processesthemselves.Iryaddirion. since 1,ouarc supposedto be /'/' a b l er o i n s r a n t i a raec o m p u t a t i o ns. i v e na p r o g r a ms p c c i f fi n gi r " p r e s c r i p t i o n s . "a s u ' e l l a s " d e s c r i p t r o n s . " ( p r o g r a m sa r e ro u s eN 1 ' g a a r d 'psh r a s e )t,h e s e m a n t i c ar le l a t i o ni s c o n s t n i n e dr o be effccrivein a u,ay that the contentrelation,for naturallan. s u a g e a n d A I . i s c l e a r l i ' n o t ( r e m e m b e rt h e f i n t l e s s o n l l . F u r t h e r m o r e ,o p e r a t i o n a la n d d e n o t a t i o n a ls e m a n t i c s ,i n theoreticalcomputerscience.are t\\'o differentkinds of analrsis of rhe samerelation;Ihat's u'hy 1'ouseeequivalenceproofs betweenthem (tle distinctionbetweenoperetionaland denoktional semantics,in otlrer words, is completelydifferentfrom the one we've been talking about betu'eenfirst and secondfacton). Finally. note tlat, under this view, a computerreally does interprct a program, in the philosopher'ssense! Bur enough about someoneelse's worries. Thc imponant point hereis that the contentrelationthat AI needsto study,as opposed to thc case just considered, is the one that holds berweenthe computational process and the world outsidc it. "ingredient" view of programs,this wouldjust be the On rhe semanricsof the program itself; on the specifrcation(computer science)view, it would be lhe semanticsof the senunticsoJ the program (i.c., two levels of reference).Whicheverway you go, threethings should be rememberedabout it: it will (at leasi in general)reachoutsidethe machine,it won'I (again,in general) bc effectivc, and it won't ever be computed. trr'i get back to McDermon. We had noted that a thorough' eoing reconstruclion from first principles was an eno[nous rheoreticaltask, and were looking at lhe other way of proceeding - by reconstructin-ePractice. With respectto the latter alernativc, I had three u'orries. First, if (as a casualreaderof McDermon)you just endone practice,you are liableto remain a-rheoretical.Second, if you try to reconstructPracticcde nol,o. you arcTdl-only faced with an enorrnoustask, but art liable to see only first'factor asPects,sincc those are the onJy onesthat arc implemented(content.remember,doesn'taPPcar in rhe program at all). And tlren third. the rwister: if you bonow techniquesfrom theoreticalcomputerscience,you will be led to focus on the wrons relationcompletely.Funhermore, not only doesn't it focus on the semanticalrelationwe'rc intcresredin, for somewhat -sraruitousreasons(the fact that programsarc prescriptive); it also ignoreslogic's first lesson:the ineducibiliryof contentto form. No matterhow you do it. in olher words, there'sa danger that you'll miss out on logic's rwo grcatlessons.And that- I hopg McDermon will agrcc - would be tragic. of prognmminglan' description Gonpox.M. l9?9. The dcnoutiona.l Ncw Yo*' NY. Springer-Vedag, an introduction. -euagcs: logic?ProccedIsn,rrt,D. 1980.Whal'su'rongwith non-monotonic on Anificial Intelliingsof thc Fint Annualn'ationalConfcrencc g c n c cS. u n f o r dC , A. pp.99-101. on modallogic' Acta Krupxl, S. 1963.Scmanticalconsiderations Fennica,l6: 83-9a. Philosophica in Lisp. Confcrcncc Surrn. B. C. l9&4. Reflcctionand scmantics of on Principlcs Rccordof rhc ElevcnthAnnualACM Symposium Salt [:]ic City, UT, pp. 23-35. Also Prognmmingl-angua-ecs, CenterIntelligcntSystcms availablcas Xcrox PaloAho Research TechnicalRcponISL-S,PaloAIto, CA, 1984. Laboratory asPedsof /n Theorctical t986. Varicriesof self-rcfcrcncc. about knowledee.Editedb.r JosephHalpcrn,Mor-ean reasoning in Anificial Los Altos.CA. pp. l9-{3. Alsoto appcar Kaufman, Inrclligence' continuum.CSU Tcchnical 1987.Thc correspondencc ReponCSLI'8?'71. Sunfod Univeniry,Sunford,CA. Also of the SixrhCanrdianConference wirh thc Procecdings appcarcd Intclligence. )"lontreal. on-Artificial Qucbec,\{arch1986,andto in AnificialIntclligcnce. appear