/LMAR=0/XLINE=3/FONT#0=BASL30/FONT#1=BASB30/FONT#3=BASI30/FONT#4=NGR40/FONT#5=NGR25/FONT#6=SUB/FONT#7=GRFX35/TMAR=10/PMAR=2650/BMAR=5
␈↓ β ␈↓␈↓ ¬g␈↓∧␈↓&Abstract␈↓)αβ␈↓

␈↓ β ␈↓Knowledge␈α≡may␈α≡be␈α≡organized␈α≡ as␈α≡ a␈α≡community␈α≥ of
␈↓ β ␈↓interacting␈α∃ modules.␈α∀Each␈α∃module␈α∀is␈α∃granted␈α∃a␈α∀complex
␈↓ β ␈↓structure,␈α⊗to␈α⊗simulate␈α⊗a␈α⊗particular␈α⊗expert␈α⊗in␈α⊗some␈α∃small
␈↓ β ␈↓domain.␈α∃ An␈α∃extended␈α∃analogy␈α∃is␈α∃drawn␈α∃to␈α∃a␈α∃group␈α∀of
␈↓ β ␈↓cooperating␈α∩human␈α∩specialists.␈α⊃ Based␈α∩on␈α∩this,␈α∩an␈α⊃internal
␈↓ β ␈↓constraint␈α
is␈α∞imposed␈α
on␈α
the␈α∞modules:␈α
Their␈α∞structure␈α
 must
␈↓ β ␈↓be␈α standard␈α over␈α the␈α entire␈αcommunity.␈α Some␈αadvantages
␈↓ β ␈↓of␈α→ a␈α→ uniform␈α~ formalism␈α→are␈α→thereby␈α~preserved.␈α→ An
␈↓ β ␈↓experimental␈α∀ community␈α∃was␈α∀implemented␈α∀ for␈α∃ the␈α∀task
␈↓ β ␈↓domain␈α⊗of␈α↔automatic␈α⊗programming.␈α↔ It␈α⊗has␈α↔managed␈α⊗to
␈↓ β ␈↓synthesize␈α≠a␈α≠few␈α~inductive␈α≠inference␈α≠ LISP␈α~ programs,
␈↓ β ␈↓nonformally,␈α→ from␈α→ speci≡c␈α_ restricted␈α→dialogues␈α→with␈α_a
␈↓ β ␈↓human user.

␈↓ β ␈↓␈↓ ∧s␈↓∧␈↓&1. Experts and Beings␈↓)αβ␈↓

␈↓ β ␈↓␈↓ βPConsider␈α∂an␈α∂interdisciplinary␈α∂enterprise,␈α∂attempted␈α∂by␈α∂a
␈↓ β ␈↓community␈α∂of␈α⊂human␈α∂experts␈α∂who␈α⊂are␈α∂specialists␈α∂in␈α⊂--␈α∂and
␈↓ β ␈↓only␈α∞in␈α
--␈α∞their␈α
own␈α∞≡elds.␈α
 What␈α∞modes␈α
of␈α∞interactions␈α
will
␈↓ β ␈↓be␈α⊃productive?␈α∩ The␈α⊃dominant␈α∩paradigm␈α⊃might␈α∩well␈α⊃settle
␈↓ β ␈↓into␈α∪␈↓βquestioning␈α∪and␈α∪answering␈↓␈α∪each␈α∪other.␈α∪ Instead␈α∀of␈α∪a
␈↓ β ␈↓chairman,␈α⊃suppose␈α⊃the␈α⊃group␈α⊃adopts␈α⊃rules␈α⊃for␈α⊃gaining␈α⊂the
␈↓ β ␈↓∨oor,␈α∞what␈α
a␈α∞speaker␈α
may␈α∞do,␈α
 and␈α∞how␈α
to␈α∞resolve␈α
disputes.
␈↓ β ␈↓When␈α
a␈α
topic␈αis␈α
being␈α
considered,␈αone␈α
or␈α
two␈α
experts␈αmight
␈↓ β ␈↓recognize␈α∞it␈α
and␈α∞speak␈α
up.␈α∞In␈α
the␈α∞course␈α
of␈α∞their␈α
exposition
␈↓ β ␈↓they␈αmight␈α
need␈αto␈αcall␈α
on␈αother␈αspecialists.␈α
This␈αmight␈αbe␈α
by
␈↓ β ␈↓name,␈α∞by␈α∞specialty,␈α∞or␈α∞simply␈α∞by␈α∞posing␈α∞a␈α∂new␈α∞sub-question
␈↓ β ␈↓and␈α
hoping␈α
someone␈α
could␈α
recognize␈α
his␈α
own␈α
relevance␈α
and
␈↓ β ␈↓volunteer␈α_a␈α_suggestion.␈α_ Such␈α_transfers␈α_would␈α→be␈α_more
␈↓ β ␈↓common␈α
at␈α∞the␈α
beginning,␈α
when␈α∞the␈α
task␈α
is␈α∞(by␈α
assumption)
␈↓ β ␈↓too␈α∩general␈α∩for␈α∩any␈α∩one␈α∩member␈α∩to␈α∩comprehend.␈α∪ As␈α∩the
␈↓ β ␈↓questions␈α∪focus␈α∪on␈α∪more␈α∪speci≡c␈α∪issues,␈α∪single␈α∪individuals
␈↓ β ␈↓should␈αbe␈αable␈αto␈αsupply␈αcomplete␈αsolutions.␈α If␈αthe␈αtask␈αis␈αto
␈↓ β ␈↓construct␈αsomething,␈αthen␈αthe␈αactivities␈αof␈αthe␈αexperts␈αshould
␈↓ β ␈↓not␈αbe␈αstrictly␈αverbal.␈α
 Often,␈αone␈αwill␈αrecognize␈αhis␈α
relevance
␈↓ β ␈↓to␈α∂the␈α∞current␈α∂situation␈α∞and␈α∂ask␈α∞to␈α∂␈↓βdo␈↓␈α∞something:␈α∂clarify␈α∞or
␈↓ β ␈↓modify or (rarely) create.

␈↓ β ␈↓␈↓ βPWhat␈α∀would␈α∀it␈α∃mean␈α∀to␈α∀␈↓βsimulate␈↓␈α∀the␈α∃above␈α∀activity?
␈↓ β ␈↓Imagine␈α~several␈α~little␈α→programs,␈α~each␈α~one␈α~modelling␈α→a
␈↓ β ␈↓di≥erent␈α
expert.␈α
What␈α
should␈α
each␈α
program,␈α
called␈α
a␈α
␈↓βBeing␈↓,
␈↓ β ␈↓be␈αcapable␈αof?␈α
 It␈αmust␈αpossess␈αa␈α
corpus␈αof␈αspeci≡c␈α
facts␈αand
␈↓ β ␈↓strategies␈α⊃for␈α∩its␈α⊃designated␈α⊃speciality.␈α∩It␈α⊃must␈α∩interact␈α⊃via
␈↓ β ␈↓questioning␈αand␈αanswering␈αother␈αBeings.␈αEach␈α
Being␈αshould
␈↓ β ␈↓be␈α
able␈αto␈α
recognize␈αwhen␈α
it␈αis␈α
relevant.␈α It␈α
must␈αset␈α
up␈αand
␈↓ β ␈↓alter structures, just as the human specialists do.

␈↓ β ␈↓␈↓ βPLet␈α∞us␈α∞return␈α∞to␈α∞our␈α∞meeting␈α∞of␈α∞human␈α∞experts.␈α∂ To␈α∞be
␈↓ β ␈↓more␈α⊃concrete,␈α⊃suppose␈α⊂their␈α⊃task␈α⊃is␈α⊂to␈α⊃design␈α⊃and␈α⊃code␈α⊂a
␈↓ β ␈↓large␈α_computer␈α_program:␈α↔a␈α_concept␈α_formation␈α↔system[2].
␈↓ β ␈↓Experts␈α⊃who␈α⊃will␈α⊃be␈α⊃useful␈α⊃include␈α⊃scienti≡c␈α⊂programmers,
␈↓ β ␈↓non-programming␈α)psychologists,␈α(system␈α)hackers,␈α(and
␈↓ β ␈↓management␈α~personnel.␈α→ What␈α~happens␈α→in␈α~the␈α→ensuing
␈↓ β ␈↓session?␈α∃ When␈α∀an␈α∃expert␈α∃participates,␈α∀he␈α∃will␈α∃either␈α∀be
␈↓ β ␈↓aiding␈α⊂a␈α⊃colleague␈α⊂in␈α⊂some␈α⊃di≠culty␈α⊂or␈α⊂else␈α⊃transferring␈α⊂a
␈↓ β ␈↓tiny,␈α
customized␈α
bit␈α
of␈αhis␈α
expertise␈α
(facts␈α
about␈α
his␈α≡eld)␈α
into
␈↓ β ␈↓a␈α
programmed␈α
function␈α
which␈α
can␈α
do␈α
something.␈α
 The␈α
≡nal
␈↓ β ␈↓code␈α
re∨ects␈αthe␈α
members'␈α
knowledge,␈αin␈α
that␈α
sense.␈α One␈α
way
␈↓ β ␈↓the␈α
session␈αmight␈α
proceed␈α
is␈αfor␈α
the␈αspecialists␈α
to␈α
actually␈α␈↓βdo␈↓
␈↓ β ␈↓the␈α⊃concept␈α⊃formation␈α∩task.␈α⊃As␈α⊃they␈α⊃become␈α∩familiar␈α⊃with
␈↓ β ␈↓what␈αpart␈αof␈αtheir␈αown␈α
expertise␈αis␈αbeing␈αcalled␈αupon,␈αand␈α
in
␈↓ β ␈↓what␈α⊃ways,␈α⊃they␈α⊃can␈α⊃begin␈α⊃to␈α⊃isolate␈α⊃it.␈α⊃ When␈α⊃it␈α⊃is␈α⊂clear



␈↓ β ␈↓␈↓¬Lenat␈↓ ¬bJuly 2, 1975␈↓ λkpage 1␈↓
␈↓ β ␈↓precisely␈α⊂what␈α⊂each␈α⊂is␈α∂doing,␈α⊂they␈α⊂can␈α⊂take␈α⊂their␈α∂extracted
␈↓ β ␈↓bits␈α_of␈α_knowledge,␈α_organize␈α_them,␈α_formalize␈α→them,␈α_and
␈↓ β ␈↓program␈α∪them.␈α∪ {A␈α∪conscious␈α∩e≥ort␈α∪along␈α∪these␈α∪lines␈α∩was
␈↓ β ␈↓made␈αin␈α
[8],␈αwhere␈αexperts␈α
gradually␈αreplaced␈α
themselves␈αby
␈↓ β ␈↓programs.␈α⊗ Instead␈α⊗of␈α⊗discussing␈α∃how␈α⊗to␈α⊗write␈α⊗a␈α∃speech
␈↓ β ␈↓program,␈α⊂they␈α∂␈↓βdid␈↓␈α⊂speech␈α∂recognition,␈α⊂until␈α∂each␈α⊂one␈α∂could
␈↓ β ␈↓introspect␈α∀su≠ciently␈α∃into␈α∀his␈α∃own␈α∀activities␈α∃to␈α∀formalize
␈↓ β ␈↓them.}␈α→ For␈α→our␈α→task,␈α_one␈α→expects␈α→the␈α→psychologists␈α_to
␈↓ β ␈↓dominate␈αthe␈α
early␈αdiscussions,␈α
later␈αyielding␈αto␈α
programmers.
␈↓ β ␈↓The␈α⊂project␈α⊂sponsor␈α⊃might␈α⊂be␈α⊂passive,␈α⊂submitting␈α⊃a␈α⊂single
␈↓ β ␈↓speci≡cation␈αorder␈α
for␈αthe␈α
program,␈αor␈α
active,␈αparticipating␈α
in
␈↓ β ␈↓the␈α⊂work␈α⊂as␈α∂a␈α⊂(somewhat␈α⊂priveleged)␈α∂member␈α⊂of␈α⊂the␈α∂team.
␈↓ β ␈↓This␈α∪individual␈α∪is␈α∪the␈α∪one␈α∪who␈α∪wants␈α∪the␈α∀≡nal␈α∪product,
␈↓ β ␈↓hence will be called the ␈↓βuser␈↓.

␈↓ β ␈↓␈↓ βPHow␈α∂could␈α⊂Beings␈α∂do␈α⊂this?␈α∂There␈α⊂would␈α∂be␈α⊂some␈α∂little
␈↓ β ␈↓program␈α∂containing␈α∞information␈α∂about␈α∞␈↓¬CONCEPT-FORMATION␈↓
␈↓ β ␈↓(much␈α⊗more␈α∃than␈α⊗would␈α∃be␈α⊗used␈α∃in␈α⊗writing␈α⊗any␈α∃single
␈↓ β ␈↓concept␈α
formation␈α
program),␈α
another␈α
Being␈α
who␈α
knows␈αhow
␈↓ β ␈↓to␈α⊃manage␈α⊂a␈α⊃group␈α⊂to␈α⊃␈↓¬WRITE-PROGRAMS␈↓,␈α⊂and␈α⊃many␈α⊂lower-
␈↓ β ␈↓level␈α⊂specialists,␈α⊂for␈α⊂example␈α⊂␈↓¬INFO-OBTAINER,␈α⊂TEST,␈α∂MODIFY-
␈↓ β ␈↓¬DATA-STRUCTURE,␈α⊗UNTIL-LOOP,␈α↔VISUAL-PERCEPTION,␈α⊗AVOID-
␈↓ β ␈↓¬CONTRADICTION,␈αPROPOSE-PLAUSIBLE-NAME␈↓.␈α Like␈α
the␈αhuman
␈↓ β ␈↓specialists,␈α the␈α Beings␈α∨would␈α contain␈α far␈α too␈α∨much
␈↓ β ␈↓information,␈α
far␈αtoo␈α
ine≠ciently␈αrepresented,␈α
to␈αbe␈α
able␈αto␈α
say
␈↓ β ␈↓"we␈α⊂ourselves␈α⊂constitute␈α⊂the␈α⊂desired␈α⊂program!"␈α⊂They␈α⊂would
␈↓ β ␈↓have␈α
to␈αdiscuss,␈α
and␈α
perhaps␈αcarry␈α
out,␈α
the␈αconcept␈α
formation
␈↓ β ␈↓task.␈α⊃They␈α∩would␈α⊃write␈α∩specialized␈α⊃versions␈α∩of␈α⊃themselves,
␈↓ β ␈↓programs␈α⊃which␈α⊃could␈α⊃do␈α∩exactly␈α⊃what␈α⊃the␈α⊃Beings␈α∩did␈α⊃to
␈↓ β ␈↓carry␈α⊂out␈α⊂the␈α⊂task,␈α⊂no␈α∂more␈α⊂nor␈α⊂less␈α⊂(although␈α⊂they␈α∂would
␈↓ β ␈↓hopefully␈α⊂take␈α⊃much␈α⊂less␈α⊂time,␈α⊃be␈α⊂more␈α⊃customized).␈α⊂ This
␈↓ β ␈↓activity␈α
is␈α
referred␈α
to␈α
in␈α
the␈α
sequel␈α
as␈α
␈↓βautomatic␈α
programming␈↓.
␈↓ β ␈↓Some␈α≡Beings␈α∨(e.g.,␈α≡␈↓¬TEST␈↓)␈α∨may␈α≡have␈α∨several␈α≡distinct,
␈↓ β ␈↓streamlined␈α∀fractions␈α∀of␈α∀themselves␈α∀in␈α∀the␈α∃≡nal␈α∀program.
␈↓ β ␈↓Beings␈α∃(e.g.,␈α∀␈↓¬PROPOSE-PLAUSIBLE-NAME␈↓)␈α∃which␈α∃only␈α∀aided
␈↓ β ␈↓other␈α↔Beings␈α⊗may␈α↔not␈α⊗have␈α↔␈↓βany␈↓␈α⊗correlates␈α↔in␈α↔the␈α⊗≡nal
␈↓ β ␈↓synthesized code.

␈↓ β ␈↓␈↓ βPAn␈α≠experimental␈α≠system,␈α≠PUP6,␈α≠was␈α≤designed␈α≠and
␈↓ β ␈↓partially␈α
implemented.␈αPUP6␈α
synthesized␈α
a␈αconcept␈α
formation
␈↓ β ␈↓program␈α∞(similar␈α
to␈α∞[7]),␈α
but␈α∞the␈α
user,␈α∞who␈α
is␈α∞human,␈α
 must
␈↓ β ␈↓come␈αup␈αwith␈α
certain␈αspeci≡c␈αanswers␈α
to␈αsome␈αof␈α
the␈αBeings'
␈↓ β ␈↓critical␈α∪queries.␈α∪ A␈α∪grammatical␈α∪inference␈α∪program␈α∪and␈α∪a
␈↓ β ␈↓simple␈αproperty␈αlist␈αmaintenance␈αroutine␈αwere␈αalso␈αgenerated.
␈↓ β ␈↓A␈α∞few␈α∞new␈α∞Beings␈α∞had␈α∞to␈α∞be␈α∞added␈α∞to␈α∞PUP6␈α∂to␈α∞synthesize
␈↓ β ␈↓them.

␈↓ β ␈↓␈↓ βPThe␈α∞next␈α∞section␈α∞illustrates␈α
how␈α∞the␈α∞experts␈α∞might␈α
have
␈↓ β ␈↓cooperated␈α∀on␈α∀the␈α∀task␈α∀of␈α∀writing␈α∀the␈α∀concept␈α∪formation
␈↓ β ␈↓program.␈α∪Section␈α∪3␈α∪describes␈α∪the␈α∪program␈α∪they␈α∩produced.
␈↓ β ␈↓Next␈α∀comes␈α∃the␈α∀Being␈α∃hypothesis:␈α∀complex␈α∃but␈α∀standard
␈↓ β ␈↓anatomy.␈α∞ Later␈α∞sections␈α∞explain␈α∞this,␈α∞both␈α∞theoretically␈α
and
␈↓ β ␈↓by␈α
examining␈α
the␈α
behavior␈α
of␈α
the␈α
actual␈α
PUP6␈α
pool␈α
of␈α
100
␈↓ β ␈↓Beings.  The appendix exhibits a typical Being.

␈↓ β ␈↓␈↓ ∧v␈↓∧␈↓&2. Experts Interacting␈↓)αβ␈↓

␈↓ β ␈↓␈↓ βPThe␈α≤input/output␈α≤behavior␈α≠of␈α≤the␈α≤desired␈α≠concept
␈↓ β ␈↓formation␈α_program␈α↔is␈α_speci≡ed␈α↔in␈α_this␈α↔section,␈α_and␈α↔we
␈↓ β ␈↓eavesdrop␈αon␈αa␈αsimulated␈α
group␈αof␈αspecialists␈α as␈αthey␈α
get␈αto
␈↓ β ␈↓work␈α∀on␈α∀writing␈α∀it.␈α∃ As␈α∀the␈α∀presentation␈α∀of␈α∃the␈α∀experts'
␈↓ β ␈↓activities␈α
becomes␈α
more␈α∞speci≡c,␈α
the␈α
reader's␈α∞currently␈α
vague
␈↓ β ␈↓conception␈α∞of␈α∞Beings␈α∞will␈α∞be␈α∞made␈α∞less␈α∞amorphous␈α∞(because
␈↓ β ␈↓Beings␈α∂are␈α∞constrained␈α∂to␈α∂carry␈α∞on␈α∂approximately␈α∂the␈α∞same
␈↓ β ␈↓discussion as the experts below do).



␈↓ β ␈↓␈↓¬Lenat␈↓ ¬bJuly 2, 1975␈↓ λkpage 1␈↓
␈↓ β ␈↓␈↓ βPExternally,␈αthe␈αconcept␈αformation␈αtask␈αcan␈αbe␈αspeci≡ed␈αas
␈↓ β ␈↓follows:␈αpictures␈αof␈αstructures␈α(built␈αout␈αof␈αsimple␈αgeometrical
␈↓ β ␈↓shapes)␈α∂will␈α∂be␈α∂presented␈α⊂one␈α∂after␈α∂another.␈α∂For␈α⊂each␈α∂such
␈↓ β ␈↓scene,␈α
the␈α
concept␈α
formation␈αprogram,␈α
call␈α
it␈α
CF,␈α
must␈αguess
␈↓ β ␈↓its␈αname.␈α
The␈αpresenter␈α
will␈αthen␈αreveal␈α
the␈αcorrect␈α
name␈αof
␈↓ β ␈↓the␈α⊗structure.␈α⊗CF␈α↔must␈α⊗quickly␈α⊗learn␈α⊗to␈α↔identify␈α⊗simple
␈↓ β ␈↓structures␈α∀(ARCH,␈α∀TOWER),␈α∃and␈α∀must␈α∀never␈α∃make␈α∀the
␈↓ β ␈↓same␈α⊂mistake␈α∂twice␈α⊂in␈α⊂a␈α∂row.␈α⊂ Assume,␈α∂as␈α⊂given,␈α⊂a␈α∂process
␈↓ β ␈↓which extracts a description of a visual scene.

␈↓ β ␈↓␈↓ βPOur␈αgroup␈αof␈αexperts␈αare␈αgiven␈αthis␈αspeci≡cation␈αfor␈αCF.
␈↓ β ␈↓Assume␈α
that␈α
the␈α
user␈α
(the␈α
≡nancial␈α
sponsor)␈α
is␈α
available␈α
for
␈↓ β ␈↓resolving␈αimportant␈αquestions,␈αvia␈αmessenger,␈αand␈αhe␈αmay␈αin
␈↓ β ␈↓fact␈αask␈αquestions␈αof␈αthe␈αgroup.␈α Whenever␈αan␈αexpert␈α
speaks,
␈↓ β ␈↓almost␈α∞all␈α∞the␈α∞others␈α∞in␈α∂the␈α∞room␈α∞hear␈α∞him.␈α∞Usually␈α∂only␈α∞a
␈↓ β ␈↓few␈α⊂can␈α⊂bene≡t␈α⊂from␈α⊂what␈α∂he␈α⊂says,␈α⊂and␈α⊂fewer␈α⊂still␈α⊂care␈α∂to
␈↓ β ␈↓react.␈α
 The␈αconversation␈α
in␈α
the␈αroom␈α
might␈α
go␈αsomething␈α
like
␈↓ β ␈↓the␈α∩following␈α⊃(the␈α∩suggestive␈α⊃names␈α∩of␈α⊃the␈α∩experts␈α∩are␈α⊃of
␈↓ β ␈↓course coincidental):
␈↓ β ␈↓GENL-MANAGER:␈α⊂Can␈α⊂anybody␈α⊂here␈α⊂≡gure␈α⊂out␈α⊂what␈α∂to
␈↓ β ␈↓␈↓ βPdo,␈α↔what␈α↔the␈α↔user's␈α↔saying?␈α↔(waves␈α↔the␈α⊗input/output
␈↓ β ␈↓␈↓ βPspeci≡cations in the air)
␈↓ β ␈↓PGM-MANAGER:␈α
I␈α
can.␈α
He␈α
wants␈α
a␈α
computer␈α∞program␈α
to
␈↓ β ␈↓␈↓ βPbe␈α∂written.␈α∂If␈α∂somebody␈α∂will␈α∂explain␈α∂the␈α⊂task␈α∂"con-cept-
␈↓ β ␈↓␈↓ βPfor-ma-tion"␈α∩to␈α⊃me␈α∩a␈α⊃little␈α∩more␈α⊃clearly,␈α∩I'll␈α∩delegate␈α⊃it
␈↓ β ␈↓␈↓ βPproperly.
␈↓ β ␈↓PSYCH:␈α∞Permit␈α∞me␈α∂to␈α∞assist␈α∞you.␈α∂I␈α∞know␈α∞all␈α∂about␈α∞concept
␈↓ β ␈↓␈↓ βPformation.  In fact, my master's thesis...
␈↓ β ␈↓PGM-MANAGER:␈αWait,␈αthe␈αuser␈αmust␈αbe␈αtold␈αthat␈αwe'll␈αbe
␈↓ β ␈↓␈↓ βPable to handle the job for him.
␈↓ β ␈↓MESSENGER:␈α∂Here,␈α∂I␈α∂can␈α⊂take␈α∂that␈α∂message.␈α∂Go␈α⊂on␈α∂with
␈↓ β ␈↓␈↓ βPyour work.
␈↓ β ␈↓PGM-MANAGER:␈α⊗We␈α⊗need␈α↔a␈α⊗name␈α⊗for␈α↔this␈α⊗program.
␈↓ β ␈↓␈↓ βPSomebody get one, please.
␈↓ β ␈↓NAMER:␈α
How␈αabout␈α
"CONCEPT"?␈αMaybe␈α
just␈α
"CF".␈αLet's
␈↓ β ␈↓␈↓ βPask the user to decide.
␈↓ β ␈↓MESSENGER:␈α_(panting)␈α_I␈α_just␈α_came␈α_back␈α_from␈α_there!
␈↓ β ␈↓␈↓ βPAlright, I'm going...  User says to call it "CF".
␈↓ β ␈↓PGM-MANAGER:␈α∞Now␈α∞then,␈α∞I␈α∞have␈α∞four␈α∞people␈α∞in␈α
mind,
␈↓ β ␈↓␈↓ βPone␈α∞of␈α∞whom␈α∞must␈α∞take␈α
over␈α∞now␈α∞in␈α∞an␈α∞important␈α
way.
␈↓ β ␈↓␈↓ βPEach of them always wants to do something di≥erent.
␈↓ β ␈↓CHOOSER:␈α∂Give␈α⊂me␈α∂their␈α⊂names␈α∂and␈α⊂I'll␈α∂choose␈α⊂the␈α∂best
␈↓ β ␈↓␈↓ βPone for you.
␈↓ β ␈↓PGM-MANAGER:␈α∃They␈α∃are␈α∃INFO-OBTAINER,␈α∀INFO-
␈↓ β ␈↓␈↓ βPUSER, PGMMER, and ANALYST.
␈↓ β ␈↓CHOOSER:␈α
You␈α∞four␈α
experts␈α
each␈α∞must␈α
justify␈α
to␈α∞me␈α
why
␈↓ β ␈↓␈↓ βPyou should take control now.
␈↓ β ␈↓␈↓ ∧ INFO-OBTAINER:␈α
We␈α
already␈α
have␈α
some␈α
unused
␈↓ β ␈↓␈↓ βPinformation, so I guess I can wait.
␈↓ β ␈↓␈↓ ∧ INFO-USER:␈α
Let␈α∞me␈α
have␈α
the␈α∞∨oor!␈α
I␈α∞know␈α
what
␈↓ β ␈↓␈↓ βPto do with the user's speci≡cation.
␈↓ β ␈↓␈↓ ∧ PGMMER:␈α⊃Well,␈α⊃I␈α⊃suppose␈α⊂I␈α⊃could␈α⊃set␈α⊃up␈α⊂some
␈↓ β ␈↓␈↓ βP"empty" top-level function, er...
␈↓ β ␈↓␈↓ ∧ ANALYST:␈α∃Marginal␈α∃pro≡ts␈α∃accruing␈α∃from␈α∀my
␈↓ β ␈↓␈↓ βPworking at this point in time do not warrant...
␈↓ β ␈↓CHOOSER:␈α⊂Okay,␈α⊂INFO-USER,␈α⊂you␈α⊂seem␈α⊂to␈α⊂be␈α⊂the␈α∂only
␈↓ β ␈↓␈↓ βPrelevant expert.
␈↓ β ␈↓INFO-USER:␈α⊂First␈α⊃thing␈α⊂is␈α⊃for␈α⊂PSYCH␈α⊂to␈α⊃tell␈α⊂us␈α⊃how␈α⊂to
␈↓ β ␈↓␈↓ βPactually ␈↓βdo␈↓ concept formation.
␈↓ β ␈↓PSYCH:␈α
There␈α
are␈α
several␈αdecisions␈α
to␈α
be␈α
made,␈αdepending
␈↓ β ␈↓␈↓ βPon␈α$what␈α$your␈α$independent␈α$variables␈α$and␈α$your
␈↓ β ␈↓␈↓ βPexperimental␈αsetup␈αare.␈αFor␈α
example,␈αare␈αwe␈αstudying␈α
one
␈↓ β ␈↓␈↓ βPindividual,␈αor␈αmust␈αour␈αprogram␈αadapt␈αto␈αsimulate␈αmany
␈↓ β ␈↓␈↓ βPdi≥erent␈α∞subjects?␈α∞Must␈α∞stimulus␈α∞items␈α∞be␈α∂just␈α∞classi≡ed,
␈↓ β ␈↓␈↓ βPor␈α⊂classi≡ed␈α∂and␈α⊂ordered,␈α⊂or␈α∂classi≡ed␈α⊂and␈α⊂ordered␈α∂and
␈↓ β ␈↓␈↓ βPmetrized?␈α
Are␈α
the␈α
scenes␈α
left␈α
in␈α
view␈α
inde≡nitely,␈α
or␈αjust
␈↓ β ␈↓␈↓ βPthe␈α
current␈α
scene,␈αor␈α
is␈α
even␈αthat␈α
one␈α
just␈α
∨ashed␈αbefore
␈↓ β ␈↓␈↓ βPthe␈α∩subject␈α⊃for␈α∩a␈α⊃limited␈α∩time?␈α⊃May␈α∩the␈α∩subject␈α⊃make
␈↓ β ␈↓␈↓ βPwritten␈α
notes?␈αCan␈α
he␈αlearn␈α
from␈αpositive␈α
instances␈αof␈α
the
␈↓ β ␈↓␈↓ βPconcepts?␈α⊃ from␈α⊂negative␈α⊃instances?␈α⊃ Is␈α⊂there␈α⊃to␈α⊃be␈α⊂any
␈↓ β ␈↓␈↓ βPpositive transfer e≥ect...?
␈↓ β ␈↓DEFER:␈αHold␈αon␈αhere;␈αwe␈αcan't␈αswamp␈αthe␈αuser␈αwith␈αlots␈αof
␈↓ β ␈↓␈↓ βPunnecessary␈α
questions.␈α Let's␈α
go␈α
through␈αeach␈α
one␈αand␈α
see



␈↓ β ␈↓␈↓¬Lenat␈↓ ¬bJuly 2, 1975␈↓ λkpage 2␈↓
␈↓ β ␈↓␈↓ βPif␈α∪it␈α∩can␈α∪wait.␈α∩ Consider␈α∪the␈α∩ decision␈α∪about␈α∪the␈α∩task
␈↓ β ␈↓␈↓ βPbeing␈αclassi≡catory,␈αcomparative,␈αor␈αmetrical.␈αWhat␈α
is␈αthe
␈↓ β ␈↓␈↓ βP≡rst situation where that decision really matters?
␈↓ β ␈↓WHEN-NEXT:␈α∩This␈α∪involves␈α∩a␈α∪list␈α∩of␈α∪alternatives.␈α∩Who
␈↓ β ␈↓␈↓ βPknows about that?
␈↓ β ␈↓ALTERNATIVES:␈α∀I␈α∀know␈α∪a␈α∀few␈α∀tricks␈α∪here.␈α∀If␈α∀all␈α∪the
␈↓ β ␈↓␈↓ βPalternative␈α⊂sequences␈α⊂of␈α∂activities␈α⊂have␈α⊂the␈α⊂same␈α∂initial
␈↓ β ␈↓␈↓ βPsubsequence,␈α⊃then␈α⊃do␈α⊃that␈α⊃common␈α⊃subsequence␈α⊃before
␈↓ β ␈↓␈↓ βPmaking the decision.
␈↓ β ␈↓UTILIZE:␈α∪In␈α∪this␈α∪case,␈α∩all␈α∪three␈α∪begin␈α∪by␈α∪partitioning␈α∩a
␈↓ β ␈↓␈↓ βPdomain of elements.
␈↓ β ␈↓DEFER:␈α⊃Temporarily,␈α⊃replace␈α⊃the␈α⊃task␈α⊃CF␈α⊃by␈α⊃the␈α∩task␈α⊃of
␈↓ β ␈↓␈↓ βPpartitioning␈α⊃a␈α∩domain.␈α⊃ ␈↓β(The␈α⊃user␈α∩is␈α⊃informed␈α∩of␈α⊃this.)␈↓
␈↓ β ␈↓␈↓ βPNow␈αconsider␈αthe␈αdecision␈αabout␈αsubject-speci≡c␈αbehavior
␈↓ β ␈↓␈↓ βPbeing required.
␈↓ β ␈↓PSYCH:␈α
This␈α
involves␈α
periodically␈α
inputting␈α
a␈α
description␈α
of
␈↓ β ␈↓␈↓ βPthe human subject.
␈↓ β ␈↓CODER:␈α∞That␈α∞would␈α∞mean␈α∞adjusting␈α∞the␈α∂algorithms␈α∞based
␈↓ β ␈↓␈↓ βPon a vector of parameters.
␈↓ β ␈↓WHEN-NEXT: This would probably a≥ect the entire code.
␈↓ β ␈↓DEFER: I can't defer this decision. Someone resolve it.
␈↓ β ␈↓RESOLVE: Ask the user about it.
␈↓ β ␈↓ASK-USER:␈α
Phrase␈α
this␈α
as␈α
a␈α
yes/no␈α
question.␈α
Explain␈α
each
␈↓ β ␈↓␈↓ βPalternative to the user.
␈↓ β ␈↓MESSENGER:␈αThat's␈αmy␈αjob...␈α␈↓β<types␈αon␈αteletype>␈↓␈αUser␈αsays
␈↓ β ␈↓␈↓ βPno, don't simulate di≥erent people.
␈↓ β ␈↓DEFER: The next decision...
␈↓ β ␈↓␈↓ ∧ ␈↓βEleven␈αdecisions␈αare␈α
ultimately␈αproposed␈αby␈α
PSYCH,
␈↓ β ␈↓β␈↓ ∧ and all but one are deferred.␈↓
␈↓ β ␈↓INFO-USER:␈α
I␈αhave␈α
no␈αobjections␈α
now␈αif␈α
someone␈αwants␈α
the
␈↓ β ␈↓␈↓ βP∨oor.
␈↓ β ␈↓PGM-MANAGER:␈α∀Do␈α∀any␈α∀of␈α∀the␈α∀other␈α∀three␈α∀experts␈α∀I
␈↓ β ␈↓␈↓ βPmentioned earlier want to speak now?
␈↓ β ␈↓PGMMER:␈α∞Yes.␈α∞I␈α∞think␈α∞that␈α∞the␈α∞top-level␈α∞function␈α∂CF␈α∞can
␈↓ β ␈↓␈↓ βPnow be coded.
␈↓ β ␈↓CODER: Give me the arguments and the body of the code.
␈↓ β ␈↓PGMMER:␈α∞There␈α
are␈α∞no␈α
known␈α∞arguments.␈α
The␈α∞body␈α∞is␈α
a
␈↓ β ␈↓␈↓ βPcall on ␈↓βPARTITION-DOMAIN␈↓.
␈↓ β ␈↓CODER:␈α↔Okay.␈α↔I␈α↔will␈α↔precede␈α⊗that␈α↔with␈α↔a␈α↔call␈α↔to␈α⊗an
␈↓ β ␈↓␈↓ βP␈↓βINITIALIZE␈↓␈α∩function,␈α∩and␈α∩follow␈α∪it␈α∩with␈α∩a␈α∩call␈α∪to␈α∩a
␈↓ β ␈↓␈↓ βP␈↓βFINALIZE␈↓␈αfunction,␈αwhich␈α
are␈αboth␈αde≡ned␈αas␈α
NIL␈αfor
␈↓ β ␈↓␈↓ βPnow.␈α∪ Is␈α∩␈↓βPARTITION-DOMAIN␈↓␈α∪simple␈α∩enough␈α∪to␈α∩be
␈↓ β ␈↓␈↓ βPcomposed right now and ≡lled in here?
␈↓ β ␈↓MATHEMATICIAN: No way. Any realizate...
␈↓ β ␈↓CODER: Uh, thanks. There. The function CF is de≡ned as:
␈↓ β ␈↓␈↓ β@␈↓¬(LAMBDA () (INITIALIZE) (PARTITION-DOMAIN) (FINALIZE)).␈↓
␈↓ β ␈↓ANALYST:␈α⊃Remind␈α⊃me␈α⊂to␈α⊃examine␈α⊃the␈α⊃initialization␈α⊂and
␈↓ β ␈↓␈↓ βP≡nalization␈α∩functions␈α⊃at␈α∩the␈α⊃end␈α∩of␈α⊃our␈α∩task.␈α∩If␈α⊃either
␈↓ β ␈↓␈↓ βPfunction is still null, it will be deleted.
␈↓ β ␈↓WARNER:␈α
I␈αhave␈α
just␈α
put␈αthat␈α
note␈α
into␈αthe␈α
code␈α
for␈αCF,␈α
as
␈↓ β ␈↓␈↓ βPa comment of type "demon".
␈↓ β ␈↓PGMMER:␈α⊂Can␈α∂someone␈α⊂advise␈α∂me␈α⊂of␈α∂what␈α⊂else␈α∂to␈α⊂do␈α∂to
␈↓ β ␈↓␈↓ βP≡nish de≡ning this function?
␈↓ β ␈↓PGM-MANAGER:␈α∃Each␈α∀function␈α∃should␈α∀have␈α∃a␈α∀proper
␈↓ β ␈↓␈↓ βPname.␈αShow␈αthe␈αuser␈αthe␈αnames␈αyou␈αhave␈αpicked,␈αand␈αlet
␈↓ β ␈↓␈↓ βPhim choose other ones if he prefers.
␈↓ β ␈↓MESSENGER: Okay... The user agrees to all 3 names.
␈↓ β ␈↓INFO-USER:␈α∪Somebody,␈α∩please␈α∪tell␈α∩the␈α∪group␈α∩how␈α∪to␈α∩␈↓βdo␈↓
␈↓ β ␈↓␈↓ βPpartitioning of a space of examples.

␈↓ β ␈↓␈↓ βPA␈α∪complete␈α∪script,␈α∪like␈α∩the␈α∪above,␈α∪was␈α∪constructed␈α∩by
␈↓ β ␈↓hand.␈α
 In␈α
the␈α∞sequel,␈α
this␈α
will␈α
be␈α∞referred␈α
to␈α
as␈α∞the␈α
␈↓βprotocol␈↓.
␈↓ β ␈↓In␈α∃all,␈α∃87␈α∃di≥erent␈α∃experts␈α∃were␈α∃called␈α∃for:␈α∃17␈α∀speci≡cly
␈↓ β ␈↓dealing␈α∞with␈α∞inductive␈α∂inference␈α∞tasks,␈α∞and␈α∞70␈α∂dealing␈α∞with
␈↓ β ␈↓programming,␈α⊃managing␈α⊃workers,␈α⊃and␈α⊃communicating␈α⊃with
␈↓ β ␈↓the␈α⊂user.␈α⊃ Near␈α⊂the␈α⊂end␈α⊃of␈α⊂the␈α⊂protocol,␈α⊃the␈α⊂user␈α⊃is␈α⊂asked
␈↓ β ␈↓which␈α
of␈α
the␈α
three␈α
types␈α
of␈α
concept␈α
formation␈α
CF␈αis␈α
supposed
␈↓ β ␈↓to␈α⊗do.␈α⊗He␈α↔responds␈α⊗"␈↓βCLASSIFICATORY␈α⊗only␈↓",␈α↔and␈α⊗the
␈↓ β ␈↓experts␈αdiscover␈αthat␈αthey␈αare␈α≡nished.␈αAll␈αthe␈αnewly␈αcreated
␈↓ β ␈↓code␈α∩is␈α∩dumped␈α∩out␈α∪onto␈α∩a␈α∩fresh␈α∩≡le:␈α∩After␈α∪hundreds␈α∩of
␈↓ β ␈↓pages,␈α_a␈α_concept␈α↔formation␈α_program␈α_meeting␈α_the␈α↔user's
␈↓ β ␈↓speci≡cations␈α→had␈α→been␈α~written.␈α→ The␈α→next␈α~section␈α→will
␈↓ β ␈↓describe that program in detail.



␈↓ β ␈↓␈↓¬Lenat␈↓ ¬bJuly 2, 1975␈↓ λkpage 3␈↓
␈↓ β ␈↓␈↓ β}␈↓∧␈↓&3. The Program the Experts Wrote␈↓)αβ␈↓

␈↓ β ␈↓␈↓ βPOne␈α∞of␈α∞the␈α∞experts␈α∞at␈α∞the␈α∞simulated␈α∞meeting␈α∂must␈α∞have
␈↓ β ␈↓read␈αP.␈α
Winston's␈αdissertation[7],␈α
because␈αCF,␈αthe␈α
synthesized
␈↓ β ␈↓concept␈α⊃formation␈α⊃program,␈α⊂was␈α⊃remarkably␈α⊃similar␈α⊃to␈α⊂the
␈↓ β ␈↓one␈α↔therein␈α_described.␈α↔CF␈α↔has␈α_a␈α↔much␈α_simpler␈α↔graph-
␈↓ β ␈↓matching␈α
algorithm,␈αand␈α
relations␈α
on␈αrelations␈α
are␈α
stored␈αin␈α
a
␈↓ β ␈↓di≥erent␈αway␈αthan␈αsimple␈αrelations␈αon␈αobjects.␈α Since␈αCF␈αwas
␈↓ β ␈↓later␈α
synthesized␈α
by␈α
PUP6,␈α
the␈α
programmed␈α
pool␈α
of␈α
Beings,
␈↓ β ␈↓it is worth detailing here.

␈↓ β ␈↓␈↓ βPCF␈α∂repeatedly␈α∂scans␈α∂a␈α∂scene␈α∂and␈α∂tries␈α∂to␈α∂name␈α∂it.␈α∂As␈α∞a
␈↓ β ␈↓≡rst␈α
step,␈αthe␈α
scene␈αis␈α
broken␈αinto␈α
a␈α
set␈αof␈α
objects␈αand␈α
a␈αset␈α
of
␈↓ β ␈↓features␈α∂(relations␈α∂on␈α⊂those␈α∂objects).␈α∂CF␈α∂maintains␈α⊂a␈α∂model
␈↓ β ␈↓for␈αeach␈αdi≥erently-named␈αscene␈αit␈αhas␈αencountered.␈αA␈αmodel
␈↓ β ␈↓contains␈α∂a␈α∂description␈α∂of␈α∂the␈α∂objects␈α∂one␈α∂expects␈α∂in␈α∂such␈α∞a
␈↓ β ␈↓structure,␈α⊃a␈α∩set␈α⊃of␈α∩features␈α⊃which␈α⊃␈↓βmust␈↓␈α∩be␈α⊃present␈α∩in␈α⊃any
␈↓ β ␈↓scene␈α
having␈αthis␈α
name,␈αa␈α
set␈αof␈α
features␈αwhich␈α
␈↓βmust␈α
not␈↓␈αbe
␈↓ β ␈↓present␈α∪if␈α∪the␈α∪scene␈α∪is␈α∀to␈α∪have␈α∪this␈α∪name,␈α∪and␈α∪a␈α∀set␈α∪of
␈↓ β ␈↓features␈αwhich␈α␈↓βmay␈↓␈αbe␈αpresent␈αor␈αabsent.␈αThus␈αa␈αmodel␈αis␈αan
␈↓ β ␈↓archetypical␈αscene␈αplus␈αa␈αname.␈α For␈αexample,␈αpart␈αof␈αa␈α
scene
␈↓ β ␈↓might be described as:
␈↓ β ␈↓␈↓ β@␈↓↓OBJECTS␈↓␈↓ ∧@␈↓ ¬a,b,c,d
␈↓ β ␈↓␈↓ β@␈↓↓RELATIONS␈↓␈↓ ¬(Green a) (Blue c) (Touches c d)
␈↓ β ␈↓␈↓ β@␈↓ ∧@␈↓ ¬(Supports a c) (Supports b c)
␈↓ β ␈↓CF's current model for an arch might be:
␈↓ β ␈↓␈↓ β@␈↓↓NAME␈↓␈↓ ∧@␈↓ ¬Arch
␈↓ β ␈↓␈↓ β@␈↓↓OBJECTS␈↓␈↓ ∧@␈↓ ¬a,b,c
␈↓ β ␈↓␈↓ β@␈↓↓MUST␈↓␈↓ ∧@␈↓ ¬(Supports a c) (Supports b c)
␈↓ β ␈↓␈↓ β@␈↓↓MUSTNOT␈↓␈↓ ¬(Touches a b)
␈↓ β ␈↓␈↓ β@␈↓↓MAY␈↓␈↓ ∧@␈↓ ¬(Green a) (Wedge c) (Prism a) (Block b)
␈↓ β ␈↓␈↓ β@␈↓ ∧@␈↓ ¬(Parallel a b) (Red a) (Red b)

␈↓ β ␈↓␈↓ βPEach␈αtime␈αit␈αis␈αconfronted␈αby␈αa␈αnew␈αscene,␈αCF␈α must␈αscan
␈↓ β ␈↓its␈α
models␈α
until␈αit␈α
≡nds␈α
one␈α
which␈αmatches␈α
it.␈α
A␈α
model␈αis␈α
said
␈↓ β ␈↓to␈α∂match␈α∂a␈α∂scene␈α∂if␈α∞all␈α∂the␈α∂MUST␈α∂features␈α∂associated␈α∞with
␈↓ β ␈↓that␈α≥model␈α≥are␈α≥observed␈α≤in␈α≥the␈α≥scene,␈α≥and␈α≥all␈α≤the
␈↓ β ␈↓MUSTNOT␈α∀ features␈α∪ are␈α∀absent␈α∪ from␈α∀ the␈α∀  scene.␈α∪CF
␈↓ β ␈↓informs␈αthe␈αuser␈αof␈αthis␈αguess,␈αand␈αaccepts␈αthe␈αproper␈α name.
␈↓ β ␈↓If␈α⊗it␈α∃ guessed␈α⊗ incorrectly,␈α⊗ CF␈α∃modi≡es␈α⊗its␈α⊗models.␈α∃The
␈↓ β ␈↓wrong-guess␈αmodel␈αmay␈αhave␈αfeatures␈αadded␈αto␈αits␈αMUST␈α
or
␈↓ β ␈↓MUSTNOT␈α∀sets.␈α∪ This␈α∀is␈α∪su≠cient␈α∀to␈α∪prevent␈α∀CF␈α∪from
␈↓ β ␈↓making␈α∃the␈α∃same␈α∃wrong␈α∀guess␈α∃twice␈α∃in␈α∃succession.␈α∀ The
␈↓ β ␈↓correct-name␈α∂model␈α⊂may␈α∂have␈α⊂ to␈α∂be␈α∂ modi≡ed␈α⊂or␈α∂(if␈α⊂it's␈α∂a
␈↓ β ␈↓new␈α⊂name)␈α⊂created␈α⊂and␈α⊂inserted␈α⊂into␈α⊂the␈α⊂list␈α⊂of␈α⊂models,␈α∂to
␈↓ β ␈↓ensure␈αthat␈αCF␈αwill␈αeventually␈αlearn␈αthat␈αconcept.␈α A␈α␈↓βconcept␈↓
␈↓ β ␈↓here␈α⊂simply␈α⊂means␈α⊂a␈α⊂model;␈α∂i.e.,␈α⊂all␈α⊂scenes␈α⊂having␈α⊂a␈α∂given
␈↓ β ␈↓name.

␈↓ β ␈↓␈↓ βPSuppose␈α∪ that␈α∪the␈α∪target␈α∩program␈α∪reads␈α∪in␈α∪the␈α∩above
␈↓ β ␈↓scene␈α
fragment␈α
and␈α
tries␈α
to␈α
match␈α
it␈α
to␈α
 the␈α∞ above␈α
 ARCH
␈↓ β ␈↓model.␈α The␈α
 MUST␈αrelations␈α
 should␈α all␈α
 be␈α present.␈α
  Yes,
␈↓ β ␈↓the␈α
 scene␈α does␈α
contain␈α
(SUPPORTS␈αa␈α
c)␈αand␈α
(SUPPORTS
␈↓ β ␈↓b␈α
c).␈α
Next,␈α
the␈α
MUSTNOT␈α
 relations␈α
 must␈α
be␈α
absent␈α
from
␈↓ β ␈↓the␈αscene.␈αSure␈αenough,␈α(TOUCHES␈αa␈αb)␈αisn't␈αthere.␈α So␈αthe
␈↓ β ␈↓model␈α
and␈α
scene␈α∞are␈α
consistent,␈α
and␈α
the␈α∞program␈α
announces
␈↓ β ␈↓that␈α
its␈α
guess␈α
is␈α
ARCH.␈α
 If␈α
the␈α
user␈α
veri≡es␈α
this␈α
guess,␈αthen
␈↓ β ␈↓the␈α⊂MAY␈α⊂set␈α⊂of␈α⊂the␈α⊂ARCH␈α⊂model␈α⊂is␈α⊂augmented␈α⊂with␈α∂the
␈↓ β ␈↓relations␈α≤(BLUE␈α≥c)␈α≤and␈α≤(TOUCHES␈α≥c␈α≤d),␈α≥and␈α≤the
␈↓ β ␈↓OBJECTS set is augmented with "d."

␈↓ β ␈↓␈↓ βPIf␈α∂the␈α∂user␈α∂denies␈α∞that␈α∂the␈α∂scene␈α∂is␈α∞an␈α∂arch,␈α∂CF␈α∂sees␈α∞if
␈↓ β ␈↓there␈α
are␈α
any␈α
relations␈α
in␈α
the␈α
ARCH␈α
model's␈α
MAY␈αset␈α
which
␈↓ β ␈↓do␈α→not␈α→occur␈α→in␈α→the␈α~scene.␈α→If␈α→so,␈α→one␈α→of␈α~them␈α→(e.g.,
␈↓ β ␈↓(PARALLEL␈αa␈αb))␈αwill␈αbe␈αtransferred␈αfrom␈αthe␈αMAY␈αto␈αthe
␈↓ β ␈↓MUST␈α∂set.␈α∂ If␈α∂no␈α∞such␈α∂feature␈α∂existed,␈α∂the␈α∂program␈α∞would
␈↓ β ␈↓look␈αfor␈αa␈α
feature␈αpresent␈αin␈αthe␈α
scene␈αbut␈αnot␈α
mentioned␈αin
␈↓ β ␈↓any␈α⊂set␈α⊂of␈α⊂the␈α⊃ARCH␈α⊂model␈α⊂(e.g.,␈α⊂(TOUCHES␈α⊂c␈α⊃d)),␈α⊂and
␈↓ β ␈↓insert␈α∂it␈α⊂into␈α∂the␈α⊂MUSTNOT␈α∂set.␈α∂ In␈α⊂either␈α∂case,␈α⊂the␈α∂user
␈↓ β ␈↓would␈α⊃be␈α⊃asked␈α⊂what␈α⊃the␈α⊃true␈α⊂name␈α⊃was,␈α⊃and␈α⊃that␈α⊂model
␈↓ β ␈↓would␈α
have␈α
its␈αMAY␈α
set␈α
augmented␈αby␈α
any␈α
new␈α
features␈αin
␈↓ β ␈↓the␈α∩scene␈α∩and␈α∩by␈α∩any␈α∩features␈α∩on␈α∩ the␈α∩true-name␈α∩model's
␈↓ β ␈↓MUST or MUSTNOT sets which contradicted the scene.



␈↓ β ␈↓␈↓¬Lenat␈↓ ¬bJuly 2, 1975␈↓ λkpage 4␈↓
␈↓ β ␈↓␈↓ βf␈↓∧␈↓&4. Anatomy of Synergetic Cooperation␈↓)αβ␈↓

␈↓ β ␈↓␈↓ βPConsider␈α⊃the␈α∩birth␈α⊃of␈α⊃one␈α∩small␈α⊃idea␈α⊃necessary␈α∩in␈α⊃the
␈↓ β ␈↓writing␈αof␈αCF␈α(e.g.,␈αthat␈αof␈αclassifying␈αa␈αmodel's␈αfeatures␈αinto
␈↓ β ␈↓three␈α∀categories␈α∀(MUST,␈α∀MUSTNOT,␈α∀MAY)).␈α∀No␈α∀single
␈↓ β ␈↓specialist␈αat␈αthe␈αmeeting␈αcould␈α
have␈αhad␈αthis␈αidea␈αby␈α
himself.
␈↓ β ␈↓How␈α⊂do␈α⊂intellects␈α⊂mesh,␈α⊂e≥ectively␈α⊂communicate,␈α⊃and␈α⊂unite
␈↓ β ␈↓their␈αpowers?␈α A␈α
tentative␈αmechanism,␈αwhich␈αbarely␈α
scratches
␈↓ β ␈↓the␈αsurface␈αof␈αthis␈αmystery,␈αwill␈αbe␈αhypothesized.␈αThe␈αBeings
␈↓ β ␈↓in␈α⊂PUP6␈α⊂ embody␈α⊂this␈α⊂concept,␈α⊂and␈α⊂are␈α⊂able␈α⊂to␈α∂reproduce
␈↓ β ␈↓both the experts' discussion and the ≡nal CF program.

␈↓ β ␈↓␈↓ βPViewing␈α⊃the␈α⊃group␈α⊃of␈α∩experts␈α⊃as␈α⊃a␈α⊃single␈α∩entity,␈α⊃what
␈↓ β ␈↓makes␈αit␈αproductive?␈αThe␈αmembers␈αmust␈αbe␈αvery␈αdi≥erent␈αin
␈↓ β ␈↓abilities,␈α
in␈α
order␈αto␈α
handle␈α
such␈αa␈α
complex␈α
task,␈α
yet␈αsimilar
␈↓ β ␈↓in␈α
basic␈α
cognitive␈α
structure␈α
(in␈α
the␈α
anatomy␈α
of␈α
their␈α
minds)␈α
to
␈↓ β ␈↓permit␈α∪facile␈α∪communications␈α∀to␈α∪∨ow.␈α∪ For␈α∀example,␈α∪each
␈↓ β ␈↓specialist␈α∞knows␈α∞how␈α
to␈α∞direct␈α∞a␈α
programmer␈α∞to␈α∞do␈α∞some␈α
of
␈↓ β ␈↓the␈α∞things␈α∞he␈α∞can␈α∞do,␈α∞but␈α∞the␈α∞speci≡c␈α∞facts␈α∞each␈α∂expert␈α∞has
␈↓ β ␈↓under␈α∩this␈α⊃category␈α∩must␈α⊃be␈α∩quite␈α⊃unique.␈α∩Similarly,␈α⊃each
␈↓ β ␈↓member␈αmay␈αhave␈αa␈αset␈αof␈αstrategies␈αfor␈αrecognizing␈αhis␈αown
␈↓ β ␈↓relevance␈α⊃to␈α⊂a␈α⊃proposed␈α⊂question,␈α⊃but␈α⊂the␈α⊃␈↓βcontents␈↓␈α⊃of␈α⊂that
␈↓ β ␈↓knowledge␈α≠varies␈α≠from␈α≠individual␈α≠to␈α≠individual.␈α≠ The
␈↓ β ␈↓hypothesis␈α⊂is␈α∂that␈α⊂all␈α∂the␈α⊂experts␈α∂can␈α⊂be␈α∂said␈α⊂to␈α⊂consist␈α∂of
␈↓ β ␈↓categorized␈α∞information,␈α
where␈α∞the␈α∞set␈α
of␈α∞categories␈α∞is␈α
fairly
␈↓ β ␈↓standard,␈αand␈αindicates␈αthe␈α␈↓βtypes␈↓␈αof␈αquestions␈αany␈αexpert␈αcan
␈↓ β ␈↓be␈αexpected␈αto␈αanswer.␈α
An␈αexpert␈αis␈αconsidered␈α
␈↓βequivalent␈↓␈αto
␈↓ β ␈↓his␈α∞answers␈α∞to␈α∞several␈α∞standard␈α∞questions.␈α∞ Each␈α∞expert␈α
has
␈↓ β ␈↓the␈α
same␈αmental␈α
"parts",␈α
it␈αis␈α
only␈α
the␈αvalues␈α
stored␈α
in␈αthese
␈↓ β ␈↓parts, their contents, which distinguish him as an individual.

␈↓ β ␈↓␈↓ βPArmed␈α∀with␈α∀this␈α∃dubious␈α∀view␈α∀of␈α∀intelligence,␈α∃let␈α∀us
␈↓ β ␈↓return␈αto␈αthe␈αdesign␈αof␈αBeings.␈αEach␈αBeing␈αshall␈α
have␈αmany
␈↓ β ␈↓parts,␈α
each␈αpossessing␈α
a␈α
name␈α(a␈α
question␈α
it␈αdeals␈α
with)␈αand␈α
a
␈↓ β ␈↓value␈α⊗(a␈α∃procedure␈α⊗capable␈α∃of␈α⊗answering␈α⊗that␈α∃question).
␈↓ β ␈↓Henceforth,␈α
"␈↓βpart␈↓"␈α
will␈α
be␈αused␈α
in␈α
this␈α
technical␈αsense.␈α
 When
␈↓ β ␈↓a␈α⊃Being␈α⊂asks␈α⊃a␈α⊂question,␈α⊃it␈α⊃is␈α⊂really␈α⊃just␈α⊂one␈α⊃part␈α⊃who␈α⊂is
␈↓ β ␈↓asking.␈α
In␈α
fact,␈αit␈α
must␈α
be␈α
that␈αthe␈α
␈↓βvalue␈↓␈α
subpart␈α
of␈αsome␈α
part
␈↓ β ␈↓can't␈αanswer␈α␈↓βhis␈↓␈α
question␈αwithout␈αfurther␈αassistance.␈α
He␈αmay
␈↓ β ␈↓not␈α
know␈α
enough␈α∞to␈α
call␈α
on␈α
speci≡c␈α∞other␈α
Beings␈α
(so␈α∞he␈α
lets
␈↓ β ␈↓anyone␈α⊃respond␈α⊃who␈α⊃feels␈α⊂relevant),␈α⊃but␈α⊃he␈α⊃should␈α⊂␈↓βalways␈↓
␈↓ β ␈↓specify␈αwhat␈αBeing␈α␈↓βpart␈↓␈αthe␈αquestion␈αshould␈αbe␈αanswered␈αby.
␈↓ β ␈↓By␈αanalogy␈α
with␈αthe␈α
experts,␈αeach␈αBeing␈α
will␈αhave␈α
the␈αsame
␈↓ β ␈↓"universal"␈α
set␈α
of␈α
types␈α
of␈α
parts␈α
(will␈α
answer␈α
the␈α
same␈αkinds
␈↓ β ␈↓of␈α↔queries),␈α↔and␈α_this␈α↔uniformity␈α↔should␈α_permit␈α↔painless
␈↓ β ␈↓intercommunication.␈α⊃Since␈α⊃the␈α∩paradigm␈α⊃of␈α⊃the␈α∩meeting␈α⊃is
␈↓ β ␈↓questioning␈α⊂and␈α⊂answering,␈α⊃the␈α⊂names␈α⊂of␈α⊂the␈α⊃parts␈α⊂should
␈↓ β ␈↓cover␈α∪all␈α∩the␈α∪types␈α∪of␈α∩questions␈α∪one␈α∩expert␈α∪wants␈α∪to␈α∩ask
␈↓ β ␈↓another.␈αEach␈αpart␈αof␈αeach␈αBeing␈αwill␈αhave␈αimplicit␈αaccess␈αto
␈↓ β ␈↓this␈αlist:␈αit␈αmay␈αask␈αonly␈αthese␈αtypes␈αof␈αquestions.␈αEach␈αBeing
␈↓ β ␈↓should␈α␈↓βnot␈↓␈αhave␈αaccess␈αto␈αthe␈αlist␈αof␈αall␈αBeings␈αin␈αthe␈αsystem:
␈↓ β ␈↓requests␈α
should␈α
be␈α
phrased␈αin␈α
terms␈α
of␈α
what␈α
is␈αwanted;␈α
rarely
␈↓ β ␈↓is␈α∃the␈α∃name␈α∃of␈α⊗the␈α∃answerer␈α∃speci≡ed␈α∃in␈α⊗advance.␈α∃ (By
␈↓ β ␈↓analogy:␈αthe␈αhuman␈αspeaker␈αis␈αnot␈αaware␈αof␈αprecisely␈α
who␈αis
␈↓ β ␈↓in␈αthe␈αroom;␈αwhen␈αhe␈αfeels␈αinadequate,␈αhe␈αasks␈αfor␈αhelp␈αand
␈↓ β ␈↓hopes␈α
someone␈α∞responds).␈α
 Another␈α
point␈α∞is␈α
that␈α∞Beings␈α
are
␈↓ β ␈↓not␈α∂a␈α∂recursive␈α∂concept␈α∂(like␈α∂ACTORs[3]␈α∂are):␈α∂a␈α∂part␈α⊂of␈α∂a
␈↓ β ␈↓Being␈αis␈αa␈αbrief␈αcollection␈αof␈αknowledge␈α(usually␈αprocedural),
␈↓ β ␈↓not␈α∃another␈α∃Being;␈α∀a␈α∃collection␈α∃of␈α∀Beings␈α∃(also␈α∃called␈α∀a
␈↓ β ␈↓community,␈αa␈αpool,␈αthe␈αsystem,␈αor␈αa␈αgroup)␈αis␈αalso␈αnot␈αitself␈αa
␈↓ β ␈↓Being. There are no ␈↓βstructured␈↓ clusters of Beings.

␈↓ β ␈↓␈↓ βPOnce␈αagain:␈αthe␈αconcept␈αof␈αa␈αpool␈αof␈αBeings␈αis␈αthat␈αmany
␈↓ β ␈↓entities␈α⊃coexist,␈α⊃each␈α⊃having␈α⊃a␈α⊃complex␈α⊃structure,␈α⊃but␈α⊃that
␈↓ β ␈↓structure␈αdoes␈αnot␈αvary␈αfrom␈αBeing␈αto␈αBeing.␈α This␈αidea␈αhas
␈↓ β ␈↓analogues␈α
in␈αmany␈α
≡elds:␈αtransactional␈α
analysis␈αin␈α
psychology,
␈↓ β ␈↓anatomy in medicine, modular design in architechture.

␈↓ β ␈↓␈↓ βPHow␈αcan␈α
we␈αtest␈α
out␈αthis␈αidea?␈α
We␈αmust␈α
build␈αa␈α
pool␈αof
␈↓ β ␈↓Beings,␈α
a␈α
modular␈α
program␈α
which␈α
will␈α
interact␈α
with␈αa␈α
human
␈↓ β ␈↓user␈α
and␈α
generate␈α
the␈α
CF␈α
program.␈α
 Recasting␈α
the␈α∞idea␈α
into
␈↓ β ␈↓operational␈α∞terms,␈α∞we␈α∞arrive␈α
at␈α∞this␈α∞procedure␈α∞for␈α∞writing␈α
a
␈↓ β ␈↓pool of Beings:



␈↓ β ␈↓␈↓¬Lenat␈↓ ¬bJuly 2, 1975␈↓ λkpage 5␈↓
␈↓ β ␈↓(1)␈αStudy␈αthe␈αtask␈αwhich␈αthe␈αpool␈αis␈αto␈αdo.␈αSee␈αwhat␈αkinds␈αof
␈↓ β ␈↓questions are asked by simulated experts.

␈↓ β ␈↓(2)␈α∞Distill␈α
this␈α∞into␈α
a␈α∞core␈α∞of␈α
simple␈α∞questions,␈α
Q,␈α∞in␈α∞such␈α
a
␈↓ β ␈↓way␈αthat␈αeach␈αinter-expert␈αquestion␈αor␈αtransfer␈αof␈αcontrol␈α
can
␈↓ β ␈↓be␈αrephrased␈αin␈αterms␈αof␈αQ.␈α
The␈αsize␈αof␈αQ␈αis␈αvery␈α
important.
␈↓ β ␈↓If␈αQ␈αis␈αtoo␈αlarge,␈αaddition␈αof␈αnew␈αBeings␈αwill␈αdemand␈αeither
␈↓ β ␈↓great␈αe≥ort␈αor␈αgreat␈αintelligence␈α(an␈αexample␈αof␈αa␈αsystem␈αlike
␈↓ β ␈↓this␈αis␈αACTORS).␈αIf␈αQ␈αis␈αtoo␈αsmall,␈αall␈αthe␈αnon-uniformity␈α
is
␈↓ β ␈↓simply␈α∂pushed␈α∂down␈α∞into␈α∂the␈α∂values␈α∞of␈α∂one␈α∂or␈α∂two␈α∞general
␈↓ β ␈↓catchall questions (all ≡rst-order logical languages do this).

␈↓ β ␈↓(3)␈α
List␈α
all␈α
the␈α
Beings␈α
who␈α
will␈α
be␈α
present␈α
in␈α
the␈α
pool,␈αand␈α
≡ll
␈↓ β ␈↓in␈α∪their␈α∪parts.␈α∩The␈α∪time␈α∪to␈α∩encode␈α∪knowledge␈α∪into␈α∩many
␈↓ β ␈↓simple␈α∂representation␈α∂schemes␈α∂is␈α∂proportional␈α∂to␈α∂the␈α∞square
␈↓ β ␈↓of␈α⊃(occasionally␈α⊃exponential␈α⊃in)␈α⊃the␈α⊃amount␈α⊃of␈α⊃interrelated
␈↓ β ␈↓knowledge␈α
(e.g.,␈α
consider␈α∞the␈α
frame␈α
problem).␈α
 The␈α∞≡lling␈α
in
␈↓ β ␈↓of␈α⊃a␈α⊃new␈α⊂Being␈α⊃is␈α⊃ ␈↓βindependent␈↓␈α⊂of␈α⊃the␈α⊃number␈α⊃of␈α⊂Beings
␈↓ β ␈↓already␈α∩in␈α∩the␈α∩pool,␈α⊃because␈α∩Beings␈α∩can␈α∩communicate␈α⊃via
␈↓ β ␈↓nondeterministic␈αgoal␈αmechanisms,␈αand␈αnot␈αhave␈αto␈αknow␈αthe
␈↓ β ␈↓names␈α∩of␈α⊃the␈α∩Beings␈α⊃who␈α∩will␈α⊃answer␈α∩their␈α∩queries.␈α⊃This
␈↓ β ␈↓≡lling␈αin␈αis␈α ␈↓βlinear␈↓␈αin␈αthe␈αnumber␈αof␈αBeing␈αparts␈αlisted␈αin␈αQ;
␈↓ β ␈↓all parts of all Beings must be (at least, should be) ≡lled in.

␈↓ β ␈↓(4)␈α∀The␈α∀human␈α∀user␈α∀interacts␈α∀with␈α∀the␈α∃completed␈α∀Being
␈↓ β ␈↓community, until the desired task is complete.

␈↓ β ␈↓␈↓ βPSection␈α⊂5␈α⊂clari≡es␈α⊂the␈α⊂e≥ects␈α⊂of␈α⊂constraining␈α⊂that␈α⊂Q␈α⊂be
␈↓ β ␈↓constant␈α∪(over␈α∪all␈α∪the␈α∪Beings␈α∪in␈α∪the␈α∪system).␈α∩ Theoretical
␈↓ β ␈↓aspects␈α
of␈α
Being␈α
systems␈α
follow,␈α
in␈α
section␈α
6.␈α
Next␈α
comes␈α
an
␈↓ β ␈↓evaluation␈αof␈α
PUP6's␈αbehavior.␈α The␈α
uses␈αand␈α
the␈αproblems
␈↓ β ␈↓with Beings are summed up in the ≡nal section.

␈↓ β ␈↓␈↓ ∧D␈↓∧␈↓&5. Internal Details of Beings␈↓)αβ␈↓

␈↓ β ␈↓␈↓ βPA␈α!set␈α!of␈α!29␈α!ubiquitous␈α!questions␈α!were␈α!chosen,
␈↓ β ␈↓representing␈α→everything␈α→one␈α→expert␈α→might␈α→want␈α~to␈α→ask
␈↓ β ␈↓another.␈α⊂At␈α⊂least,␈α⊂they␈α⊂naturally␈α⊂encompass␈α⊂those␈α∂questions
␈↓ β ␈↓which␈α↔were␈α↔asked␈α↔during␈α↔the␈α↔simulated␈α_meeting,␈α↔hence
␈↓ β ␈↓should␈αbe␈α
su≠cient␈αfor␈α
generating␈αCF.␈α
 Q,␈αthis␈α
universal␈αset
␈↓ β ␈↓of␈αBeing␈αparts,␈αis␈αlisted␈αin␈αAppendix␈α1.␈α The␈αreader␈αis␈αurged
␈↓ β ␈↓to␈αglance␈αat␈αthis␈αnow,␈α
and␈αrefer␈αto␈αit␈αwhenever␈αa␈α
Being␈αpart
␈↓ β ␈↓is speci≡cally mentioned.

␈↓ β ␈↓␈↓ βPEach␈α⊃of␈α⊃the␈α⊃100␈α⊃Beings␈α⊃in␈α⊃PUP6␈α⊃should␈α⊃have␈α⊃had␈α⊃a
␈↓ β ␈↓value␈α∞for␈α∞each␈α∞part␈α∞(in␈α∂reality,␈α∞only␈α∞40%␈α∞of␈α∞these␈α∂2900␈α∞slots
␈↓ β ␈↓were␈α⊂≡lled␈α⊂in;␈α⊂only␈α∂30%␈α⊂were␈α⊂actually␈α⊂necessary␈α⊂to␈α∂generate
␈↓ β ␈↓CF).␈αA␈α
value␈αfor␈α
a␈αpart␈αis␈α
simply␈αa␈α
LISP␈αprogram␈αwhich␈α
can
␈↓ β ␈↓answer␈α
that␈α
question,␈α
often␈α∞by␈α
asking␈α
questions␈α
of␈α∞the␈α
same
␈↓ β ␈↓Being,␈α⊂of␈α⊂other␈α⊃Beings,␈α⊂and␈α⊂of␈α⊂the␈α⊃user.␈α⊂A␈α⊂part␈α⊃may␈α⊂also
␈↓ β ␈↓assert␈α∂some␈α∞fact,␈α∂create␈α∂or␈α∞modify␈α∂some␈α∂structure␈α∞(including
␈↓ β ␈↓demons,␈α
Beings,␈α
and␈α
parts␈α
of␈α
Beings).␈α
 Appendix␈α
1␈αshows␈α
the
␈↓ β ␈↓values␈α∂stored␈α∞under␈α∂each␈α∞part␈α∂for␈α∞the␈α∂typical␈α∂Being␈α∞named
␈↓ β ␈↓"␈↓¬INFO-OBTAINER␈↓".

␈↓ β ␈↓␈↓ βPThe␈α∂set␈α∂of␈α∂parts␈α∂breaks␈α∂into␈α∂three␈α∂rough␈α⊂categories:␈α∂(1)
␈↓ β ␈↓those␈α∂parts␈α⊂which␈α∂are␈α∂useful␈α⊂in␈α∂deciding␈α∂which␈α⊂Being␈α∂gets
␈↓ β ␈↓control,␈α∃(2)␈α∃those␈α∃which␈α⊗are␈α∃used␈α∃once␈α∃the␈α⊗Being␈α∃gains
␈↓ β ␈↓control,␈α↔and␈α↔(3)␈α⊗those␈α↔useful␈α↔only␈α⊗to␈α↔answer␈α↔the␈α⊗user's
␈↓ β ␈↓questions␈αand␈αkeep␈αhim␈αoriented.␈α The␈αnext␈αsection␈αdescribes
␈↓ β ␈↓categories␈α
1␈αand␈α
2;␈αthe␈α
section␈αafter␈α
that␈αexplicates␈α
the␈αthird
␈↓ β ␈↓category of Being parts.

␈↓ β ␈↓␈↓↓␈↓&5.1. Control in the PUP6 System␈↓)αβ␈↓

␈↓ β ␈↓␈↓ βPAt␈αthe␈αhumans'␈αmeeting,␈αonly␈αone␈αexpert␈αspoke␈αat␈αa␈αtime;
␈↓ β ␈↓in␈αthe␈αBeings␈α
community,␈αonly␈αone␈α
Being␈αhas␈αcontrol␈α
at␈αany
␈↓ β ␈↓given␈α⊂moment.␈α⊂He␈α∂uses␈α⊂his␈α⊂parts␈α∂to␈α⊂do␈α⊂things␈α⊂(ask,␈α∂create,
␈↓ β ␈↓modify),␈α∀and␈α∀yields␈α∀control␈α∀either␈α∀voluntarily␈α∀or␈α∀through
␈↓ β ␈↓interruption.

␈↓ β ␈↓␈↓ βPIn␈α∃slightly␈α∀more␈α∃procedural␈α∃terms,␈α∀the␈α∃scenario␈α∃is␈α∀as



␈↓ β ␈↓␈↓¬Lenat␈↓ ¬bJuly 2, 1975␈↓ λkpage 6␈↓
␈↓ β ␈↓follows.␈α∞One␈α∞part␈α∞of␈α∞a␈α∞Being␈α∞senses␈α∞its␈α∞relevance␈α∞(often␈α∞the
␈↓ β ␈↓IDEN␈α⊂or␈α⊂EFFECTS␈α∂parts,␈α⊂which␈α⊂are␈α∂united␈α⊂with␈α⊂all␈α∂such
␈↓ β ␈↓parts␈αto␈αform␈αa␈αlarge␈αproduction␈αsystem[5]).␈αIf␈αmore␈αthan␈αone
␈↓ β ␈↓Being␈α∨wants␈α∨control␈α at␈α∨any␈α∨time,␈α∨a␈α special␈α∨Being,
␈↓ β ␈↓CHOOSER,␈α≠seizes␈α≠control␈α≠momentarily.␈α≠He␈α≤asks␈α≠each
␈↓ β ␈↓competing␈α∩Being␈α∩to␈α⊃evaluate␈α∩its␈α∩WHEN␈α⊃part,␈α∩to␈α∩see␈α⊃how
␈↓ β ␈↓seriously␈α
it␈α
needs␈α
to␈α
go␈α
immediately.␈α
If␈α
some␈α
Beings␈α
are␈αstill
␈↓ β ␈↓tied␈α≤for␈α≥≡rst␈α≤place,␈α≥he␈α≤asks␈α≥them␈α≤to␈α≥evaluate␈α≤their
␈↓ β ␈↓COMPLEXITY␈α∂parts,␈α∂to␈α∂see␈α⊂which␈α∂is␈α∂the␈α∂simplest.␈α⊂If␈α∂any
␈↓ β ␈↓␈↓βstill␈↓␈α⊃tie␈α⊃for␈α⊃top,␈α⊃one␈α⊂is␈α⊃randomly␈α⊃chosen.␈α⊃In␈α⊃any␈α⊃case,␈α⊂the
␈↓ β ␈↓winner␈α∀is␈α∀then␈α∪passed␈α∀control.␈α∀Once␈α∪in␈α∀control,␈α∀a␈α∪Being
␈↓ β ␈↓arranges␈αsome␈αof␈αits␈αparts␈αin␈αsome␈αorder␈αand␈αevaluates␈αthem.
␈↓ β ␈↓For␈α⊂example,␈α⊂the␈α⊂ARGS␈α⊂part␈α⊂might␈α⊂be␈α⊂≡rst;␈α⊂if␈α⊂it␈α⊂asks␈α∂for
␈↓ β ␈↓some␈α∀arguments␈α∀which␈α∃no␈α∀Being␈α∀has␈α∀supplied,␈α∃then␈α∀the
␈↓ β ␈↓whole␈α→Being␈α→might␈α~decide␈α→to␈α→fail.␈α→Some␈α~ parts,␈α→when
␈↓ β ␈↓evaluated,␈α⊃might␈α⊃create␈α⊃a␈α⊃new␈α⊃Being,␈α⊃might␈α∩ask␈α⊃questions
␈↓ β ␈↓which␈α⊂require␈α⊂this␈α⊂whole␈α⊂process␈α⊂to␈α⊂repeat␈α⊃recursively,␈α⊂etc.
␈↓ β ␈↓This␈α
"asking"␈α
really␈α
means␈α
broadcasting␈α
a␈α
request␈α
to␈α
one␈α
or
␈↓ β ␈↓two␈αparts␈αof␈αevery␈αBeing;␈αfor␈αexample␈α"Is␈αthere␈αa␈αknown␈αfast
␈↓ β ␈↓way␈α∂of␈α∂gronking␈α∂toves?"␈α∂would␈α∂be␈α∂asked␈α∂as␈α∂a␈α∂search␈α⊂for␈α∂a
␈↓ β ␈↓Being␈α⊗whose␈α⊗COMPLEXITY␈α⊗part␈α⊗indicated␈α⊗speed,␈α∃and
␈↓ β ␈↓whose␈α~EFFECTS␈α~part␈α→contained␈α~a␈α~production␈α~with␈α→a
␈↓ β ␈↓template␈αmatching␈α"gronking␈αtoves".␈α A␈αlist␈αof␈αthe␈αresponders
␈↓ β ␈↓would␈α
be␈α
returned.␈α
(Incidentally,␈α
GERUND␈α
would␈αrecognize
␈↓ β ␈↓this,␈α∞but␈α∞later␈α∞give␈α∞up␈α∞when␈α∞no␈α∞one␈α∞could␈α∞recognize␈α
"gronk
␈↓ β ␈↓toves".)␈α↔The␈α⊗questioner␈α↔might␈α⊗pose␈α↔some␈α↔new␈α⊗questions
␈↓ β ␈↓directly␈α∩to␈α∩these␈α∩Beings,␈α∩might␈α∩turn␈α∩control␈α∩over␈α∩to␈α∩them
␈↓ β ␈↓directly,␈α⊗etc.␈α⊗One␈α⊗way␈α⊗or␈α⊗another,␈α⊗the␈α⊗Being␈α∃eventually
␈↓ β ␈↓relinquishes␈αcontrol.␈αIf␈αit␈αhad␈αno␈αdirect␈αsuccessor␈αin␈αmind,␈αall
␈↓ β ␈↓the␈α∞Beings␈α
are␈α∞asked␈α∞if␈α
they␈α∞want␈α∞to␈α
take␈α∞over.␈α∞ There␈α
will
␈↓ β ␈↓always␈α↔be␈α⊗␈↓βsome␈↓␈α↔Being␈α⊗who␈α↔will␈α⊗take␈α↔over;␈α↔the␈α⊗general
␈↓ β ␈↓management␈αtypes␈αof␈αBeings␈αare␈αalways␈αable␈α --␈αbut␈αreluctant
␈↓ β ␈↓-- to do so.

␈↓ β ␈↓␈↓ βPHow␈α
does␈α
each␈α
Being␈αdecide␈α
which␈α
parts␈α
to␈αevaluate,␈α
and
␈↓ β ␈↓in␈α⊃which␈α⊃order,␈α⊃once␈α⊃it␈α⊃gains␈α⊃control?␈α⊃ The␈α⊃answer␈α⊂might
␈↓ β ␈↓seem␈α⊃to␈α⊃be␈α⊃di≠cult␈α∩or␈α⊃tedious␈α⊃for␈α⊃whoever␈α∩writes␈α⊃Beings,
␈↓ β ␈↓since␈α∞it␈α
might␈α∞vary␈α∞from␈α
Being␈α∞to␈α∞Being.␈α
In␈α∞fact,␈α∞it␈α
doesn't!
␈↓ β ␈↓The␈α_commitment␈α_to␈α↔a␈α_universal␈α_set␈α↔of␈α_Being␈α_parts␈α↔is
␈↓ β ␈↓ine≠cient␈α∞in␈α∞some␈α∞ways␈α∞(each␈α
Being␈α∞␈↓βneeded␈↓␈α∞only␈α∞a␈α∞third␈α
of
␈↓ β ␈↓all␈α∞the␈α∞parts)␈α∞but␈α∂allows␈α∞for␈α∞some␈α∞simpli≡cations␈α∂right␈α∞here.
␈↓ β ␈↓What␈αparts␈αshould␈αbe␈αevaluated,␈αand␈αin␈αwhat␈αorder,␈αwhen␈αa
␈↓ β ␈↓Being␈αgains␈αcontrol?␈αThis␈αdecision␈αdepends␈αprimarily␈α
on␈αthe
␈↓ β ␈↓␈↓βtypes␈↓␈αof␈αparts␈αpresent␈αin␈αthe␈αBeing,␈αnot␈αon␈αtheir␈α␈↓βvalues␈↓.␈α But
␈↓ β ␈↓every␈α∞Being␈α∞has␈α∞the␈α
same␈α∞anatomy,␈α∞so␈α∞one␈α∞single␈α
algorithm
␈↓ β ␈↓can␈α∀assemble␈α∪any␈α∀Being's␈α∀parts␈α∪into␈α∀an␈α∀executable␈α∪LISP
␈↓ β ␈↓function.␈α∀Moreover,␈α∀this␈α∀assemby␈α∪can␈α∀be␈α∀done␈α∀when␈α∪the
␈↓ β ␈↓system␈α
is␈α
≡rst␈α∞loaded␈α
(or␈α
when␈α
a␈α∞new␈α
Being␈α
is␈α∞≡rst␈α
created),
␈↓ β ␈↓and␈αneed␈αonly␈αbe␈αredone␈α
for␈αa␈αBeing␈αwhen␈αthe␈αvalues␈α
of␈αits
␈↓ β ␈↓parts␈α∂change.␈α⊂Such␈α∂changes␈α⊂are␈α∂rare:␈α⊂experts␈α∂are␈α⊂not␈α∂often
␈↓ β ␈↓open-minded.␈α∂The␈α∞precise␈α∂algorithm␈α∞is␈α∂sketched␈α∞in␈α∂the␈α∞box
␈↓ β ␈↓below.␈α∀The␈α∀parts␈α∪useful␈α∀here␈α∀include␈α∀ARGS,␈α∪DEMONS,
␈↓ β ␈↓META-CODE,␈α3COMMENTS,␈α4ARG-CHECK,␈α3and
␈↓ β ␈↓REQUISITES.
␈↓ β ␈↓π␈↓ βP⊂ααααααααααααααααααααααααααααααααααααααααααααα␈↓ 	0⊃
␈↓ β ␈↓¬␈↓ β@␈↓ βi␈↓↓␈↓&Assembling a Being into an executable function␈↓)αβ␈↓¬
␈↓ β ␈↓¬␈↓ β@When␈αa␈αBeing␈α␈↓βB␈↓¬␈αfirst␈αgains␈αcontrol,␈αits␈αEXPLICIT-ARGS␈α
are
␈↓ β ␈↓¬␈↓ β@bound.␈α∂The␈α∞IMPLICIT-ARGS␈α∂are␈α∞initialized,␈α∂the␈α∞name␈α∂␈↓βB␈↓¬␈α∞is
␈↓ β ␈↓¬␈↓ β@pushed␈α∪onto␈α∪the␈α∪Being␈α∩control␈α∪stack,␈α∪and␈α∪any␈α∩newly-
␈↓ β ␈↓¬␈↓ β@activated␈αDEMONS␈αare␈αso␈αtagged.␈αThe␈αBeing␈αwho␈αcalled␈α␈↓βB␈↓¬
␈↓ β ␈↓¬␈↓ β@should␈α∀have␈α∀explained␈α∪his␈α∀reasons␈α∀by␈α∀assigning␈α∪some
␈↓ β ␈↓¬␈↓ β@phrase␈α∪to␈α∩the␈α∪variable␈α∪BECAUSE.␈α∩ This␈α∪reason␈α∪is␈α∩now
␈↓ β ␈↓¬␈↓ β@stored␈α∀as␈α∀a␈α∀special␈α∀sub-part␈α∪of␈α∀the␈α∀WHY␈α∀part␈α∀of␈α∪␈↓βB␈↓¬.
␈↓ β ␈↓¬␈↓ β@BECAUSE␈α∞is␈α∞rebound␈α∞periodically␈α∞in␈α∞the␈α∂META-CODE␈α∞and
␈↓ β ␈↓¬␈↓ β@COMMENTS␈αparts,␈α
to␈αkeep␈α
current␈αthe␈α
explanation␈αof␈α
each
␈↓ β ␈↓¬␈↓ β@call␈α≥that␈α≥␈↓βB␈↓¬␈α≥makes.␈α≥Each␈α≥ARG-CHECK␈α≡predicate␈α≥is
␈↓ β ␈↓¬␈↓ β@evaluated.␈α⊃If␈α⊃any␈α∩returns␈α⊃NIL,␈α⊃the␈α⊃entire␈α∩Being␈α⊃reports
␈↓ β ␈↓¬␈↓ β@that␈α∃it␈α∀has␈α∃failed;␈α∀otherwise,␈α∃the␈α∃PRE-REQUISITES␈α∀are
␈↓ β ␈↓¬␈↓ β@examined.␈α∞Effort␈α∞is␈α∞expended␈α
to␈α∞make␈α∞them␈α∞true,␈α∞if␈α
they
␈↓ β ␈↓¬␈↓ β@are␈α∂currently␈α∂not␈α∞satisfied.␈α∂Each␈α∂COMMENT␈α∂is␈α∞evaluated,
␈↓ β ␈↓¬␈↓ β@then␈α∀the␈α∪CO-REQUISITES,␈α∀META-CODE,␈α∪and␈α∀the␈α∪current
␈↓ β ␈↓¬␈↓ β@demons␈α∃are␈α∀executed␈α∃in␈α∀pseudo-parallel.␈α∃ Each␈α∀POST-
␈↓ β ␈↓¬␈↓ β@REQUISITE␈αis␈αthen␈αexamined,␈αand␈αan␈αeffort␈αmade␈αto␈αsatisfy
␈↓ β ␈↓¬␈↓ β@it.␈α⊗ The␈α↔newly-activated␈α⊗demons␈α↔are␈α⊗exorcized,␈α↔␈↓βB␈↓¬␈α⊗is


␈↓ β ␈↓␈↓¬Lenat␈↓ ¬bJuly 2, 1975␈↓ λkpage 7␈↓
␈↓ β ␈↓¬␈↓ β@popped␈α∃from␈α∃the␈α⊗Being␈α∃control␈α∃stack,␈α∃and␈α⊗the␈α∃value
␈↓ β ␈↓¬␈↓ β@computed␈α∨by␈α∨the␈α∨META-CODE␈α∨is␈α∨returned.␈α≡ Some
␈↓ β ␈↓¬␈↓ β@heuristics␈α
were␈α
devised␈α
to␈αtake␈α
advantage␈α
of␈α
the␈αfact␈α
that
␈↓ β ␈↓¬␈↓ β@the␈αBeings␈α
often␈αdidn't␈αneed␈α
many␈αof␈αthe␈α
standard␈αparts.
␈↓ β ␈↓¬␈↓ β@For␈α∞example,␈α∞INFO-OBTAINER␈α∞has␈α∞no␈α∞new␈α∞demons␈α∂or␈α∞co-
␈↓ β ␈↓¬␈↓ β@requisites, so no parallel processing need be simulated.
␈↓ β ␈↓π%αααααααααααααααααααααααααααααααααααααααααααααααα␈↓ 	0$

␈↓ β ␈↓␈↓↓␈↓&5.2. Keeping the User Informed␈↓)αβ␈↓

␈↓ β ␈↓␈↓ βPIn␈αthe␈αearlier␈αconversation␈αexcerpts,␈αthe␈αsimulated␈α
human
␈↓ β ␈↓user␈α_had␈α_no␈α_trouble␈α_whatever␈α_understanding␈α_what␈α↔the
␈↓ β ␈↓simulated␈α∃experts␈α∃asked␈α∃him.␈α∃In␈α∃the␈α⊗actual␈α∃programmed
␈↓ β ␈↓PUP6␈α∩system,␈α∩the␈α∩human␈α∩who␈α∩was␈α∩sitting␈α∩at␈α∩the␈α∩teletype
␈↓ β ␈↓quite␈α
␈↓βrarely␈↓␈α
understood␈α
what␈αwas␈α
wanted␈α
by␈α
the␈α
Beings.␈αHe
␈↓ β ␈↓frequently␈α∩had␈α∩to␈α∩interrupt␈α⊃them␈α∩and␈α∩ask␈α∩them␈α⊃questions
␈↓ β ␈↓about␈α∂who␈α∂was␈α⊂in␈α∂control,␈α∂why,␈α∂what␈α⊂he␈α∂was␈α∂trying␈α⊂to␈α∂do,
␈↓ β ␈↓what␈α⊗had␈α⊗recently␈α⊗transpired,␈α⊗etc.␈α⊗These␈α⊗ideally␈α⊗can␈α∃be
␈↓ β ␈↓phrased␈α⊂as␈α⊂simple␈α⊂retrievals␈α⊂and␈α⊂EVALs␈α⊂of␈α⊂active␈α∂Beings'
␈↓ β ␈↓parts.␈α The␈αBeing␈αparts␈αmost␈αoften␈αcalled␈αfor␈αby␈αthe␈αuser␈αare
␈↓ β ␈↓the␈α∀simple␈α∀one-line␈α∀"orientation"␈α∀templates.␈α∀These␈α∪include
␈↓ β ␈↓WHAT,␈α⊗HOW,␈α↔WHY,␈α⊗and␈α⊗AFFECTS.␈α↔ For␈α⊗theoretical
␈↓ β ␈↓reasons␈α∩explained␈α⊃later,␈α∩the␈α⊃synthesized␈α∩program,␈α∩CF,␈α⊃was
␈↓ β ␈↓writen␈α∞as␈α∞a␈α∞pool␈α∞of␈α∞Beings␈α∞itself␈α∞(by␈α∞PUP6,␈α∞but␈α∞not␈α
during
␈↓ β ␈↓the␈α⊂protocol.␈α⊃Actually,␈α⊂a␈α⊂fortuitous␈α⊃"bug"␈α⊂in␈α⊃PUP6␈α⊂created
␈↓ β ␈↓this␈α∀intriguing␈α∪situation.)␈α∀Although␈α∀its␈α∪question-answering
␈↓ β ␈↓ability␈α⊃is␈α⊃inferior␈α⊃to␈α⊃PUP6,␈α∩the␈α⊃fact␈α⊃that␈α⊃it␈α⊃has␈α∩␈↓βany␈↓␈α⊃such
␈↓ β ␈↓power␈αwas␈αsurprising␈α
to␈αthe␈αauthor.␈α
In␈αother␈αwords,␈α
one␈αcan
␈↓ β ␈↓interrupt␈α↔the␈α⊗target␈α↔program␈α⊗as␈α↔it␈α⊗is␈α↔running␈α↔and␈α⊗ask
␈↓ β ␈↓questions.␈α
Any␈α
Being␈α
on␈α
the␈α
control␈α
stack␈α
will␈α
provide␈α
fully
␈↓ β ␈↓instantiated␈α⊂answers␈α⊂to␈α⊂any␈α⊂of␈α⊂its␈α⊂29␈α⊂allowable␈α⊃queries␈α⊂(its
␈↓ β ␈↓parts);␈α~all␈α→other␈α~Beings␈α→will␈α~provide␈α~only␈α→hypothetical
␈↓ β ␈↓answers.␈α⊃As␈α⊃an␈α⊃example,␈α⊂consider␈α⊃this␈α⊃actual␈α⊃excerpt␈α⊃of␈α⊂a
␈↓ β ␈↓human␈α
using␈αthe␈α
CF␈α
program␈αsynthesized␈α
by␈α
PUP6.␈α (Some
␈↓ β ␈↓liberty␈α
has␈α∞been␈α
taken␈α∞with␈α
the␈α
English;␈α∞e.g.,␈α
the␈α∞user␈α
really
␈↓ β ␈↓types␈α
␈↓βWHAT?␈↓,␈α∞not␈α
␈↓βWhat␈α∞are␈α
you␈α∞doing?␈↓)␈α
"???"␈α∞simply␈α
means
␈↓ β ␈↓"guess␈α→the␈α→name␈α→of␈α_the␈α→scene␈α→with␈α→these␈α→objects␈α_and
␈↓ β ␈↓relations".  CF types in ␈↓βitalics␈↓, the user in ␈↓↓boldface.␈↓
␈↓ β ␈↓π⊂αααααααααααααααααααααααααααααααααααααααααααααααα␈↓ 	0⊃
␈↓ β ␈↓β␈↓ β@CF: Ready to accept brand new scene.
␈↓ β ␈↓β␈↓ β@␈↓↓User: (??? (A B) (Block A) (Wedge B) (Touches A B))␈↓β
␈↓ β ␈↓β␈↓ β@CF: Not an Arch. Not a Tower. Might be a Row.
␈↓ β ␈↓β␈↓ β@␈↓ ¬:My guess is Row.   What is it Really?
␈↓ β ␈↓β␈↓ β@␈↓↓User:  Nothing␈↓β
␈↓ β ␈↓β␈↓ β@CF: Transfer from the May to the Must set of the
␈↓ β ␈↓β␈↓ β@␈↓ ¬lRow model, the feature: (Block  B)
␈↓ β ␈↓β␈↓ β@CF: Ready to accept brand new scene.
␈↓ β ␈↓π%αααααααααααααααααααααααααααααααααααααααααααααααα␈↓ 	0$

␈↓ β ␈↓␈↓ βPHere␈α
is␈α∞the␈α
␈↓βsame␈↓␈α
excerpt,␈α∞but␈α
with␈α
the␈α∞user␈α
interrupting
␈↓ β ␈↓occasionally.␈α∩ "␈↓↓U:␈↓"␈α∪signi≡es␈α∩he␈α∪has␈α∩interrupted,␈α∪and␈α∩"␈↓↓OK.␈↓"
␈↓ β ␈↓signi≡es "␈↓↓Go back to work now, CF.␈↓"
␈↓ β ␈↓β␈↓ ε1       ␈↓↓U:  What are you doing?␈↓β
␈↓ β ␈↓βCF:  Divide a domain into subdomains,
␈↓ β ␈↓β     using specialized version of PARTITION-DOMAIN.
␈↓ β ␈↓β␈↓ λ3       ␈↓↓How?␈↓β
␈↓ β ␈↓βCF:  By building up the partition gradually,
␈↓ β ␈↓β     by adding one <scene, name> pair.   ␈↓ 	
␈↓↓OK.␈↓β
␈↓ β ␈↓βCF: Ready to accept brand new scene.
␈↓ β ␈↓β␈↓↓User: (??? (A B) (Block A) (Wedge B) (Touches A B))␈↓β
␈↓ β ␈↓β␈↓ ε1       ␈↓↓U:  What are you doing?␈↓β
␈↓ β ␈↓βCF:  Break OBJECTS piece o≥ input scene to examine.
␈↓ β ␈↓β␈↓ βe       ␈↓↓What code is the current Being going to execute?␈↓β
␈↓ β ␈↓βCF:  The meta-code is (CADR arg1). Evals to (A B). ␈↓ 	
␈↓↓OK.␈↓β
␈↓ β ␈↓βCF: Not an Arch.␈↓ λC␈↓↓U:  Why?␈↓β
␈↓ β ␈↓βCF: The ARCH MUST-relation (SUPPORTS A B)
␈↓ β ␈↓β    is contradicted in the SCENE.␈↓ 	
␈↓↓OK.␈↓β
␈↓ β ␈↓βCF: Not a Tower. Might be a Row.  My guess is Row.
␈↓ β ␈↓β      What is it really?␈↓ εs␈↓↓U:  Why are you asking?␈↓β
␈↓ β ␈↓βCF: I must resolve it now, but can't, so I ask User.
␈↓ β ␈↓β␈↓ ¬→       ␈↓↓OK. What you wanted is:   Nothing␈↓β
␈↓ β ␈↓β␈↓ ¬p       ␈↓↓U:  What are you doing now?␈↓β
␈↓ β ␈↓βCF: Go through space of concept names, and make user


␈↓ β ␈↓␈↓¬Lenat␈↓ ¬bJuly 2, 1975␈↓ λkpage 8␈↓
␈↓ β ␈↓β    aware of changes.␈↓ ε≡␈↓↓What Beings might this a≥ect?␈↓β
␈↓ β ␈↓βCF: MESSAGE is possibly called; some version of TEST
␈↓ β ␈↓β    is possibly called.␈↓ 	
␈↓↓OK.␈↓β
␈↓ β ␈↓βCF: Transfer from the May to the Must set of the
␈↓ β ␈↓β    Row model, the feature: (Block B)␈↓ λM␈↓↓U: Why?␈↓β
␈↓ β ␈↓βCF: Change model of ROW until it contradicts
␈↓ β ␈↓β    the previous inputted Scene.␈↓ 	
␈↓↓OK.␈↓β
␈↓ β ␈↓βCF: Ready to accept brand new scene.

␈↓ β ␈↓␈↓ ∧β␈↓∧␈↓&6. Theory of Pure Beings Systems␈↓)αβ␈↓

␈↓ β ␈↓␈↓ βPWe␈α∩now␈α∩discuss␈α∩the␈α∩constraints␈α∩each␈α∩Being,␈α∩and␈α⊃each
␈↓ β ␈↓group␈α∂of␈α∂Beings,␈α∂must␈α∞conform␈α∂to.␈α∂Hopefully,␈α∂ideas␈α∂will␈α∞be
␈↓ β ␈↓separated␈α≠from␈α≠implementation␈α≠details,␈α≠prejudices␈α≠from
␈↓ β ␈↓plausible features.

␈↓ β ␈↓␈↓ βPIt␈αwould␈αbe␈αaesthetically␈αpleasing␈αto␈αrestrict␈αall␈αentities␈αin
␈↓ β ␈↓the␈α∀system␈α∃to␈α∀be␈α∃Beings.␈α∀However,␈α∃this␈α∀would␈α∃cause␈α∀an
␈↓ β ␈↓in≡nite␈αregress,␈αas␈αeach␈αpart␈αof␈αeach␈αBeing␈αwould␈αhave␈αparts
␈↓ β ␈↓which␈α∞had␈α
parts...␈α∞To␈α
stop␈α∞this,␈α
one␈α∞can␈α
assert␈α∞that␈α∞at␈α
some
␈↓ β ␈↓≡nite␈α~level,␈α→all␈α~constructs␈α→are␈α~primitive.␈α~ACTORs,␈α→for
␈↓ β ␈↓example,␈αset␈αthis␈αlevel␈αto␈αzero;␈αBeings␈αset␈αit␈αto␈αone.␈αACTORs
␈↓ β ␈↓themselves␈α⊂are␈α⊂primitive,␈α∂but␈α⊂only␈α⊂␈↓βparts␈↓␈α∂of␈α⊂Beings␈α⊂can␈α∂be.
␈↓ β ␈↓For␈α∂this␈α∞reason,␈α∂Beings␈α∂can␈α∞not␈α∂be␈α∞viewed␈α∂as␈α∂a␈α∞convergent
␈↓ β ␈↓recursive de≡nition, as can ACTORs.

␈↓ β ␈↓␈↓ βPSuppose␈α
it␈α
were␈α
decreed␈α
that␈α
the␈α
only␈αautonomous␈α
entities
␈↓ β ␈↓possessing␈α∪control␈α∀abilities␈α∪were␈α∀Beings.␈α∪In␈α∀particular,␈α∪we
␈↓ β ␈↓forbid␈α∃any␈α⊗plain␈α∃␈↓βfunctions␈↓␈α⊗to␈α∃exist.␈α∃ In␈α⊗the␈α∃case␈α⊗of␈α∃an
␈↓ β ␈↓automatic␈αprogramming␈αtask,␈αthe␈αBeings␈αwould␈αhave␈αto␈α
write
␈↓ β ␈↓new␈α⊃Beings,␈α⊃not␈α⊃new␈α⊃LISP␈α⊃functions.␈α⊃The␈α⊃target␈α⊂program
␈↓ β ␈↓would␈α
thus␈α
itself␈αbe␈α
a␈α
community␈α
of␈αBeings.␈α
In␈α
order␈α
to␈α≡ll␈α
in
␈↓ β ␈↓all␈α
the␈αparts,␈α
a␈α
vast␈αamount␈α
of␈α
super∨uous␈αinformation␈α
would
␈↓ β ␈↓be␈α∞collected.␈α∂These␈α∞supplementary␈α∞facts␈α∂can␈α∞be␈α∞viewed␈α∂as␈α∞a
␈↓ β ␈↓standardized,␈α∂organized␈α∂body␈α∞of␈α∂␈↓βdocumentation␈↓,␈α∂a␈α∞formatted
␈↓ β ␈↓system of comments tacked onto each Being produced.

␈↓ β ␈↓␈↓ βPWhich␈α∩Beings␈α∪would␈α∩write␈α∪the␈α∩new␈α∪Beings?␈α∩ Looking
␈↓ β ␈↓back␈αat␈αour␈αinterdisciplinary␈αexperts,␈αwe␈αsee␈αthat␈αeach␈αexpert
␈↓ β ␈↓is␈α∩responsible␈α∩for␈α∩distilling␈α∩his␈α∩own␈α∩essential␈α∩contribution,
␈↓ β ␈↓which␈α∀is␈α∪then␈α∀encoded␈α∪by␈α∀a␈α∪programmer.␈α∀ Perhaps␈α∪each
␈↓ β ␈↓Being␈α≤should␈α≠be␈α≤able␈α≤to␈α≠direct␈α≤construction␈α≤of␈α≠new,
␈↓ β ␈↓specialized␈αBeings␈αwhich␈αrelate␈αto␈αit.␈αIf␈αno␈αBeing␈αrelates␈αto␈αa
␈↓ β ␈↓task,␈α∞then␈α∞it␈α∞can't␈α∂be␈α∞coded;␈α∞if␈α∞several␈α∞respond,␈α∂they␈α∞should
␈↓ β ␈↓cooperate.␈αThis␈αability␈αis␈αin␈αreality␈αthe␈αSPECIALIZATIONS
␈↓ β ␈↓part␈α∪of␈α∪each␈α∪Being␈α∪(see␈α∪Appendix␈α∪1).␈α∪The␈α∀Being␈α∪which
␈↓ β ␈↓actually␈αdoes␈αthe␈αcreation␈α(␈↓¬CODER␈↓)␈αin␈αthe␈αexperimental␈αsystem
␈↓ β ␈↓is␈α
almost␈α
trivial,␈α∞getting␈α
very␈α
precise␈α
instructions␈α∞from␈α
other
␈↓ β ␈↓Beings.

␈↓ β ␈↓␈↓ βPSince␈α∩the␈α∩pool␈α∩must␈α⊃communicate␈α∩with␈α∩the␈α∩user,␈α⊃some
␈↓ β ␈↓Beings␈α∪must␈α∪translate␈α∪quasi-English␈α∪phrases␈α∪into␈α∀calls␈α∪on
␈↓ β ␈↓Beings.␈α∞ Drawing␈α∞again␈α∂on␈α∞our␈α∞experts␈α∞analogy,␈α∂we␈α∞require
␈↓ β ␈↓that␈α
each␈α
Being␈α
recognize␈α
his␈α
own␈α
relevance.␈α
So␈α
translation␈α
is
␈↓ β ␈↓merely␈αthe␈αact␈α
of␈αasking␈αthe␈α
whole␈αpool␈α"Who␈α
can␈αrecognize
␈↓ β ␈↓this...",␈α∪collecting␈α∪the␈α∩responders,␈α∪having␈α∪␈↓βthem␈↓␈α∪decide␈α∩who
␈↓ β ␈↓should␈α
take␈αcontrol,␈α
and␈αletting␈α
the␈αwinner␈α
do␈αthe␈α
translation.
␈↓ β ␈↓Most␈α∃communication␈α∃is␈α∃done␈α⊗as␈α∃if␈α∃it,␈α∃too,␈α∃were␈α⊗such␈α∃a
␈↓ β ␈↓translation activity.

␈↓ β ␈↓␈↓ βPOne␈αbias␈αis␈αthe␈αrejection␈αof␈αdebugging␈αas␈αa␈αfundamental
␈↓ β ␈↓programming␈α∞tool.␈α∞ It␈α∞is␈α∞felt␈α∞to␈α∞be␈α∞worth␈α∞the␈α∞extra␈α∂e≥ort␈α∞to
␈↓ β ␈↓make␈αthe␈αsystem's␈αinternal␈αmodel␈αof␈αthe␈αcurrent␈αpartial␈αtarget
␈↓ β ␈↓program␈α!␈↓βcorrect␈↓.␈α"Debugging␈α!demands␈α"detective␈α!work,
␈↓ β ␈↓examing␈αone's␈αearlier␈αe≥orts␈αfor␈α∨aws,␈αfor␈αdetails␈αwhich␈αhave
␈↓ β ␈↓been␈α∀overlooked.␈α∀ Any␈α∃tireless␈α∀system␈α∀ should␈α∃not␈α∀ignore
␈↓ β ␈↓details,␈α⊃but␈α⊃rather␈α⊂defer␈α⊃them,␈α⊃asserting␈α⊂a␈α⊃warning␈α⊃to␈α⊂this
␈↓ β ␈↓e≥ect␈α∂when␈α∂it␈α∂does␈α∂so.␈α∂Procrastination␈α∂is␈α∂quite␈α⊂valuable;␈α∂in
␈↓ β ␈↓PUP6,␈α↔much␈α↔e≥ort␈α↔is␈α↔spent␈α↔deferring␈α_any␈α↔unresolvable
␈↓ β ␈↓decision.␈α∂ Undeferrable␈α∂unresolvable␈α∞decisions␈α∂must␈α∂cause␈α∞a
␈↓ β ␈↓backtrack␈α∞ point␈α
to␈α∞be␈α∞reluctantly␈α
set␈α∞up.␈α∞Another␈α
prejudice
␈↓ β ␈↓is␈α∀that␈α∀most␈α∪carelessness␈α∀bugs␈α∀can␈α∪be␈α∀eliminated␈α∀by␈α∪this
␈↓ β ␈↓deferral,␈α⊂feed-forward,␈α⊂and␈α⊂precise␈α⊂record-keeping.␈α⊂Humans



␈↓ β ␈↓␈↓¬Lenat␈↓ ¬bJuly 2, 1975␈↓ λkpage 9␈↓
␈↓ β ␈↓depend␈αon␈αtheir␈αadaptability␈αto␈αcompensate␈αfor␈αlimitations␈αin
␈↓ β ␈↓their␈α
brain␈α
hardware,␈α
but␈α
there␈α
is␈α
no␈α
need␈α
for␈α
an␈α␈↓βautomatic␈↓
␈↓ β ␈↓programming␈α
system␈α
to␈α
do␈α
so.␈α
 These␈α
biasses␈α
are␈αnot␈α
inherent
␈↓ β ␈↓in␈α∩the␈α∩Beings␈α∩formulation,␈α∩but␈α⊃only␈α∩in␈α∩the␈α∩design␈α∩of␈α⊃the
␈↓ β ␈↓PUP6 system (and in the mind of the author).

␈↓ β ␈↓␈↓ βPTo␈α_clarify␈α_what␈α↔Beings␈α_are␈α_and␈α↔are␈α_not,␈α_they␈α↔are
␈↓ β ␈↓contrasted␈α
with␈α
some␈αother␈α
ideas.␈α
FRAMES[4]␈αare␈α
su≠ciently
␈↓ β ␈↓amorphous␈αto␈αsubsume␈αBeings.␈αIn␈αphilosophy,␈αFRAMES␈αare
␈↓ β ␈↓meant␈αto␈αmodel␈αperception,␈αand␈αintentionally␈αrely␈αon␈αimplicit
␈↓ β ␈↓default␈α⊂values;␈α∂Beings␈α⊂intentionally␈α∂avoid␈α⊂making␈α∂decisions
␈↓ β ␈↓by␈α∂default.␈α∂ This␈α∂is␈α∞also␈α∂the␈α∂di≥erence␈α∂ between␈α∞HACKER
␈↓ β ␈↓and␈α↔PUP6.␈α↔ Since␈α↔PUP6␈α↔writes␈α↔structured␈α↔programs,␈α⊗it
␈↓ β ␈↓should␈α→be␈α~distinguished␈α→from␈α→macro␈α~expansion.␈α→Macro
␈↓ β ␈↓procedures␈α
expand␈α
mechanically:␈α
␈↓↓expand(sequence␈α  m␈↓ε1␈↓↓␈α
 m␈↓ε2␈↓↓)
␈↓ β ␈↓↓=␈α∪(sequence␈α∩ expand(m␈↓ε1␈↓↓)␈α∪ expand(m␈↓ε2␈↓↓)))␈↓.␈α∩Beings␈α∪could␈α∩use
␈↓ β ␈↓information␈αgleaned␈αduring␈αexpansion␈αof␈αm␈↓ε1␈↓␈αto␈α
improve␈αthe
␈↓ β ␈↓way␈α∞m␈↓ε2␈↓␈α
was␈α∞handled.␈α
 ACTORs[3],␈α∞unlike␈α
Beings,␈α∞have␈α
no
␈↓ β ␈↓≡xed␈αstructure␈αimposed,␈αand␈αdo␈αnot␈αbroadcast␈αtheir␈αmessages
␈↓ β ␈↓(they␈α→ specify␈α→who␈α→gets␈α~each␈α→message,␈α→by␈α→name,␈α~to␈α→a
␈↓ β ␈↓bureaucracy).

␈↓ β ␈↓␈↓ βPBeings␈α
subsume␈α
(ine≠ciently)␈α
many␈α
popular␈α
AI␈α
features;
␈↓ β ␈↓the␈α∞demonstration␈α∞will␈α∞be␈α
brief:␈α∞A␈α∞␈↓βdemon␈↓␈α∞could␈α∞be␈α
replaced
␈↓ β ␈↓by␈α≥a␈α≥Being␈α≥whose␈α≥ARG-CHECK␈α≥predicate␈α≥was␈α≥the
␈↓ β ␈↓triggering␈αpredicate,␈αwhose␈αWHEN␈αpart␈αwas␈αhigh␈αenough␈αto
␈↓ β ␈↓ensure␈α∪frequent␈α∪attention,␈α∪and␈α∪whose␈α∪META-CODE␈α∩part
␈↓ β ␈↓was␈α∂the␈α∞body␈α∂of␈α∂the␈α∞demon.␈α∂An␈α∞␈↓βassertion␈↓␈α∂in␈α∂an␈α∞associative
␈↓ β ␈↓data␈αnetwork␈α
could␈αbe␈α
a␈αBeing␈α
with␈αonly␈α
an␈αIDEN␈αpart␈α
≡lled
␈↓ β ␈↓in;␈α⊗when␈α∃it␈α⊗recognizes␈α⊗its␈α∃relevance,␈α⊗a␈α⊗fully␈α∃instantiated
␈↓ β ␈↓assertion␈α⊃is␈α⊃returned.␈α⊃A␈α⊃␈↓βfunction␈↓␈α⊃is␈α⊃equivalent␈α⊃to␈α⊃a␈α⊃Being
␈↓ β ␈↓with␈αonly␈αa␈α
META-CODE,␈αARGS,␈αand␈αEVAL-ARGS␈α
parts;
␈↓ β ␈↓one␈αknows␈α
almost␈αnothing␈αabout␈α
it␈αbefore␈αexecuting␈α
it.␈α The
␈↓ β ␈↓ine≠ciencies␈α∩should␈α∩be␈α∩clear:␈α⊃whenever␈α∩a␈α∩Being␈α∩throws␈α⊃a
␈↓ β ␈↓question␈α
open␈α
to␈α
the␈α
∨oor,␈α"Who␈α
can...",␈α
it␈α
takes␈α
an␈αamount␈α
of
␈↓ β ␈↓time␈α⊃proportional␈α⊃to␈α⊃the␈α⊃number␈α⊃of␈α⊃Beings␈α⊃in␈α⊃the␈α⊃system.
␈↓ β ␈↓One␈αwould␈α
introduce␈αthis␈αhuge␈α
time␈αfactor␈αby␈α
replacing␈αany
␈↓ β ␈↓of the above mechanisms by Beings.

␈↓ β ␈↓␈↓ βPThe␈α␈↓βnumber␈↓␈αof␈α
Being␈αparts␈α seems␈α
to␈αindicate␈αthe␈α
balance
␈↓ β ␈↓between␈α⊂uniformity␈α⊃and␈α⊂structure␈α⊃in␈α⊂the␈α⊃community.␈α⊂ This
␈↓ β ␈↓was␈αtouched␈α
on␈αas␈αpoint␈α
(2)␈αof␈αSection␈α
4.␈α A␈α
small␈αuniversal
␈↓ β ␈↓set␈α∀of␈α∀Being␈α∀parts␈α∃is␈α∀necessary␈α∀to␈α∀preserve␈α∀some␈α∃of␈α∀the
␈↓ β ␈↓advantages␈αof␈αuniformity␈α
(easy␈αaddition␈αof␈αknowledge␈α
to␈αthe
␈↓ β ␈↓system,␈α
easy␈α
inter-Being␈α
communication).␈α
This␈α
demands␈α
that
␈↓ β ␈↓the␈αnumber␈αof␈αparts␈αof␈αeach␈αBeing␈αbe,␈αsay,␈αunder␈α100.␈αBut␈αit
␈↓ β ␈↓is␈α⊂the␈α⊂complex␈α⊂structure␈α∂of␈α⊂a␈α⊂Being␈α⊂which␈α⊂makes␈α∂complex
␈↓ β ␈↓behaviors␈α
feasable,␈α∞including␈α
∨exible␈α
communication␈α∞as␈α
well
␈↓ β ␈↓as␈α∂viable␈α⊂≡nal␈α∂products.␈α⊂So␈α∂each␈α⊂Being␈α∂should␈α⊂have␈α∂many
␈↓ β ␈↓parts,␈α∞say␈α
at␈α∞least␈α∞ten.␈α
 This␈α∞range,␈α
␈↓¬10␈α∞≤␈α∞||Q||␈α
≤␈α∞100␈↓,␈α∞is␈α
wide
␈↓ β ␈↓for␈αthe␈α
domain␈αof␈α
automatic␈αprogramming.␈α
In␈αother␈α
domains,
␈↓ β ␈↓it␈αmay␈α
be␈αnarrow␈α
or␈αdisappear␈α
altogether;␈αthis␈αwould␈α
indicate
␈↓ β ␈↓that Beings could ␈↓βnot␈↓ be used e≥ectively for those tasks.

␈↓ β ␈↓␈↓ ∧g␈↓∧␈↓&7. Experimental Results␈↓)αβ␈↓

␈↓ β ␈↓␈↓ βPTwo␈α⊗of␈α↔the␈α⊗most␈α↔signi≡cant␈α⊗questions␈α↔for␈α⊗automatic
␈↓ β ␈↓programming␈α∀systems␈α∪are␈α∀what␈α∪programs␈α∀are␈α∪synthesized,
␈↓ β ␈↓and what kinds of things the user must tell the system.

␈↓ β ␈↓␈↓↓␈↓&7.1. Character of the Dialogue between User and PUP6␈↓)αβ␈↓

␈↓ β ␈↓␈↓ βPBy␈α⊂lumping␈α⊂all␈α⊂the␈α⊂parts␈α⊂of␈α⊂the␈α⊂PUP6␈α⊂system␈α∂together
␈↓ β ␈↓conceptually,␈α
the␈α∞interaction␈α
is␈α∞seen␈α
as␈α∞a␈α
␈↓βdialogue␈↓␈α∞between␈α
a
␈↓ β ␈↓human␈α∩user␈α⊃and␈α∩an␈α⊃automatic␈α∩programming␈α∩system.␈α⊃ The
␈↓ β ␈↓level␈α
of␈α
detail␈α
maintained␈α
in␈α
the␈α
earlier␈α
polylogue␈α
(discussions
␈↓ β ␈↓among␈α∂the␈α∂experts)␈α∂obscured␈α∂what␈α∂the␈α∂human␈α∂actually␈α∂saw
␈↓ β ␈↓and␈α∞responded.␈α∞Below␈α
is␈α∞the␈α∞same␈α
excerpt␈α∞as␈α∞was␈α
presented
␈↓ β ␈↓in␈αSection␈α2,␈αbut␈αas␈αthe␈α␈↓βuser␈↓␈αsitting␈αat␈αa␈αteletype␈αactually␈αsaw
␈↓ β ␈↓it.␈α⊂ These␈α⊂were␈α⊂the␈α⊃initial␈α⊂few␈α⊂lines␈α⊂of␈α⊂an␈α⊃actual␈α⊂300-page
␈↓ β ␈↓long␈α⊂dialogue.␈α⊂In␈α⊂section␈α⊃2,␈α⊂this␈α⊂excerpt␈α⊂was␈α⊃presented␈α⊂the
␈↓ β ␈↓way experts (the same way the PUP6 Beings) discussed it.



␈↓ β ␈↓␈↓¬Lenat␈↓ ¬bJuly 2, 1975␈↓ λ\page 10␈↓
␈↓ β ␈↓PUP: I want any task.
␈↓ β ␈↓USER: Write a program which does concept formation
␈↓ β ␈↓PUP: I am about to write a program for the task:
␈↓ β ␈↓           ␈↓¬specialized kind of CONCEPT-FORMATION␈↓.
␈↓ β ␈↓PUP: Give me a proper name for this program.
␈↓ β ␈↓                (I suggest CF or CONCEPT)
␈↓ β ␈↓USER: CF
␈↓ β ␈↓PUP: Must CF mimic di≥erent human subjects?
␈↓ β ␈↓USER: NO
␈↓ β ␈↓PUP: I am temporarily replacing the task  "␈↓¬specialized
␈↓ β ␈↓¬        kind of CONCEPT-FORMATION␈↓" by the simpler
␈↓ β ␈↓        task  "␈↓¬specialized kind of PARTITION-A-DOMAIN␈↓".

␈↓ β ␈↓␈↓ βPFrom␈α⊂the␈α⊃above,␈α⊂one␈α⊂gets␈α⊃the␈α⊂true␈α⊂impression␈α⊃that␈α⊂the
␈↓ β ␈↓dialogue␈α⊃was␈α⊃fairly␈α⊃tedious:␈α⊃choosing␈α⊃names␈α∩and␈α⊃resolving
␈↓ β ␈↓design␈α
decisions␈α
␈↓βwas␈↓␈α
the␈α
user's␈α
main␈α
role.␈α
There␈α
were,␈α
in␈α
sum,
␈↓ β ␈↓about␈α∪300␈α∪such␈α∀responses␈α∪involved␈α∪in␈α∪generating␈α∀CF.␈α∪13
␈↓ β ␈↓crucial␈αresponses␈αwere␈αalso␈αnecessary,␈αwhere␈αthe␈αuser␈αsupplies
␈↓ β ␈↓needed␈α∞information␈α
that␈α∞was␈α∞not␈α
explicitly␈α∞requested␈α∞by␈α
the
␈↓ β ␈↓Beings.␈α A␈αfew␈αof␈αthese␈αkey␈αlines␈αare:␈α␈↓βWrite␈αa␈αprogram␈αwhich
␈↓ β ␈↓βdoes␈αconcept␈αformation;␈αAn␈α"element␈αof␈αthe␈αdomain␈αof␈αstimuli"
␈↓ β ␈↓βis␈αa␈αstatic␈αscene;␈αThe␈αguessing␈αprocess␈αis␈αone␈αof␈αsearching,␈αnot
␈↓ β ␈↓βcomputing;␈α↔Terminate␈α↔the␈α_loop␈α↔i≥␈α↔any␈α↔model␈α_feature␈α↔is
␈↓ β ␈↓βincompatible with the scene's features␈↓.

␈↓ β ␈↓␈↓ βPThe␈α
translation␈α∞mechanism␈α
is␈α
simply␈α∞the␈α
IDEN␈α∞parts␈α
of
␈↓ β ␈↓all␈α∂the␈α⊂Beings.␈α∂Each␈α∂such␈α⊂part␈α∂contains␈α∂a␈α⊂few␈α∂productions,
␈↓ β ␈↓which␈α⊂are␈α⊂united␈α⊂with␈α⊃those␈α⊂of␈α⊂the␈α⊂other␈α⊂Beings␈α⊃into␈α⊂one
␈↓ β ␈↓large␈α∂production␈α∂system.␈α∂The␈α∂left␈α∞side␈α∂of␈α∂a␈α∂production␈α∂is␈α∞a
␈↓ β ␈↓template␈α⊃which␈α⊃is␈α∩matched␈α⊃against␈α⊃the␈α⊃given␈α∩phrase.␈α⊃The
␈↓ β ␈↓right␈α∩side␈α⊃is␈α∩a␈α∩small␈α⊃program␈α∩to␈α∩be␈α⊃executed␈α∩in␈α∩case␈α⊃the
␈↓ β ␈↓phrase␈α
is␈α
recognized.␈α
Generally,␈α
that␈α
program␈α
simply␈α
says␈αto
␈↓ β ␈↓translate␈α
some␈αsubphrase␈α
of␈αthe␈α
phrase,␈αand␈α
then␈α(probably)
␈↓ β ␈↓pass␈αcontrol␈αto␈αthe␈αBeing␈αwhich␈αthis␈αis␈αin␈αthe␈αIDEN␈αpart␈αof.
␈↓ β ␈↓While␈α∂not␈α∂up␈α∂to␈α∂the␈α⊂"state␈α∂of␈α∂the␈α∂art"␈α∂in␈α⊂natural␈α∂language
␈↓ β ␈↓processing,␈αthis␈αis␈αboth␈αadequate␈αand␈αfaithful␈αto␈αthe␈αspirit␈αof
␈↓ β ␈↓distributing␈α∂problems␈α∂over␈α∂all␈α∂the␈α∂Beings,␈α∂by␈α∂making␈α∂them
␈↓ β ␈↓␈↓βparts␈↓, so it is each Being's duty to know a little about them.

␈↓ β ␈↓␈↓ βPDuring␈α∞the␈α∞course␈α∂of␈α∞the␈α∞dialogue,␈α∞PUP6␈α∂originates␈α∞the
␈↓ β ␈↓idea␈α∃of␈α∃a␈α∃MUST/MUSTNOT/MAY␈α∃structuring␈α⊗of␈α∃each
␈↓ β ␈↓model's␈α
features.␈α
Here␈α
is␈α
how␈α
the␈α
user␈α
sees␈α
that␈α
portion␈α
of␈α
the
␈↓ β ␈↓synthesis.␈α⊃Notice␈α⊃that␈α⊃only␈α⊃the␈α⊃≡rst␈α⊃user␈α⊃comment␈α⊃is␈α⊃other
␈↓ β ␈↓than a simple choice.
␈↓ β ␈↓PUP: How do we know when to terminate this loop?
␈↓ β ␈↓USER:␈α~Terminate␈α~i≥␈α→any␈α~feature␈α~in␈α~model-features␈α→is
␈↓ β ␈↓␈↓ βPincompatible with scene-features
␈↓ β ␈↓PUP:␈α∩PUP␈α∩wants␈α∩USER␈α∩to␈α∩type␈α∩in␈α∩name␈α∪for␈α∩specialized
␈↓ β ␈↓␈↓ βPversion of CONTRADICTS
␈↓ β ␈↓USER: Call it IMPOSS
␈↓ β ␈↓PUP:␈α
Move␈αcursor␈α
to␈αIMPOSS␈α
type␈αof␈α
CONTRADICTS.␈αI
␈↓ β ␈↓␈↓ βPcan␈αno␈αlonger␈αdefer␈αthis␈αdecision:␈αA:␈αContradiction␈αarises
␈↓ β ␈↓␈↓ βPfrom␈α∂features␈α⊂who␈α∂have␈α⊂probability=1␈α∂of␈α⊂occurring.␈α∂ B:
␈↓ β ␈↓␈↓ βPContradiction␈α-arises␈α,from␈α-features␈α-who␈α,have
␈↓ β ␈↓␈↓ βPprobability=0␈α
of␈α
occurring.␈α
 C:␈α
Contradiction␈α
arises␈αfrom
␈↓ β ␈↓␈↓ βPfeatures␈α∞who␈α∞have␈α
probability>0&<1␈α∞of␈α∞occurring␈α
Please
␈↓ β ␈↓␈↓ βPtype some of: A,B,C; or type: Defer, ?, or Details.
␈↓ β ␈↓USER: A, B, and C are all relevant.
␈↓ β ␈↓PUP:␈α∞I␈α
recommend␈α∞that␈α∞each␈α
model's␈α∞features␈α∞be␈α
structured
␈↓ β ␈↓␈↓ βP(along␈α⊗1␈α⊗dimension)␈α⊗into␈α⊗these␈α⊗3␈α↔pieces:␈α⊗PRESENT,
␈↓ β ␈↓␈↓ βPwhich␈α∂will␈α∂hold␈α⊂features␈α∂whose␈α∂occurrence␈α⊂should␈α∂have
␈↓ β ␈↓␈↓ βPprobability=1;␈αABSENT,␈α which␈αwill␈αhold␈αfeatures␈αwhose
␈↓ β ␈↓␈↓ βPoccurrence␈α↔should␈α⊗have␈α↔probability=0;␈α↔and␈α⊗EITHER,
␈↓ β ␈↓␈↓ βPwhich␈α∂will␈α∂hold␈α⊂features␈α∂whose␈α∂occurrence␈α⊂should␈α∂have
␈↓ β ␈↓␈↓ βPany␈α_other␈α_probability.␈α_ Please␈α_type␈α_back␈α_YES,␈α↔NO,
␈↓ β ␈↓␈↓ βPUNSURE, DETAILS, and/or 3 new names.
␈↓ β ␈↓USER:␈α8YES.␈α8Call␈α8them␈α8MUST-BE-PRESENT,
␈↓ β ␈↓␈↓ βPMUSTNOT-BE-PRESENT, MAY-BE-PRESENT.
␈↓ β ␈↓PUP: Too long. Nicknamed Must, Mustnot, May.
␈↓ β ␈↓PUP:␈α↔Located␈α↔3␈α↔references␈α↔to␈α↔model's␈α↔features␈α_in␈α↔code
␈↓ β ␈↓␈↓ βPgenerated␈α⊂so␈α⊂far.␈α⊂ In␈α⊂each␈α⊂case,␈α⊂the␈α⊂reference␈α⊃has␈α⊂been
␈↓ β ␈↓␈↓ βPreplaced by "(APPEND Must Mustnot May)".




␈↓ β ␈↓␈↓¬Lenat␈↓ ¬bJuly 2, 1975␈↓ λ\page 11␈↓
␈↓ β ␈↓␈↓ βPThis␈α_is␈α_PUP6␈α_at␈α↔its␈α_best:␈α_making␈α_just␈α_the␈α↔"right"
␈↓ β ␈↓inference␈α∂at␈α∞the␈α∂"right"␈α∞time.␈α∂ The␈α∞hundreds␈α∂of␈α∞inter-Being
␈↓ β ␈↓transfers␈α∩are␈α∩masked:␈α⊃the␈α∩system␈α∩appears␈α⊃to␈α∩type␈α∩out␈α⊃just
␈↓ β ␈↓what␈αis␈αnecessary.␈αThe␈αnext␈αfew␈αparagraphs␈αdispell␈αthis␈αfalse
␈↓ β ␈↓illusion;␈αthere␈α
is␈αno␈α
mystery␈αto␈α
PUP6's␈αdoing␈α
the␈αright␈α
things,
␈↓ β ␈↓and any ␈↓βtypical␈↓ dialogue excerpt would be boringly verbose.

␈↓ β ␈↓␈↓ βPBecause␈α∩of␈α⊃its␈α∩genesis␈α∩from␈α⊃a␈α∩single␈α∩"experts␈α⊃meeting"
␈↓ β ␈↓protocol,␈α∩the␈α∩PUP6␈α∪pool␈α∩of␈α∩Beings␈α∪was␈α∩(i)␈α∩easily␈α∪able␈α∩to
␈↓ β ␈↓reproduce␈αthat␈α"proper"␈αdialogue,␈αbut␈α(ii)␈αincapable␈αof␈αwidely
␈↓ β ␈↓varied␈αdialogues␈αwith␈αthe␈αuser.␈α Two␈αassumptions␈αimplicit␈αin
␈↓ β ␈↓using␈α∪the␈α∩simulated␈α∪experts'␈α∪discussion␈α∩as␈α∪a␈α∪literal␈α∩model
␈↓ β ␈↓turned␈α∞out␈α∂to␈α∞be␈α∂serious:␈α∞the␈α∂abilities␈α∞of␈α∂any␈α∞actual␈α∂user␈α∞of
␈↓ β ␈↓PUP6␈α_must␈α_coincide␈α↔with␈α_those␈α_of␈α↔the␈α_user␈α_who␈α↔was
␈↓ β ␈↓simulated␈α∞in␈α∞the␈α
protocol,␈α∞and␈α∞the␈α
order␈α∞of␈α∞and␈α∞wording␈α
of
␈↓ β ␈↓the␈α
actual␈α
user's␈αcomments␈α
must␈α
closely␈αadhere␈α
to␈α
those␈αof␈α
the
␈↓ β ␈↓user␈α⊂in␈α∂the␈α⊂protocol.␈α∂Such␈α⊂di≠culties␈α∂must␈α⊂be␈α⊂overcome␈α∂in
␈↓ β ␈↓any␈α∞system␈α∞designed␈α∞for␈α
wide␈α∞usership,␈α∞but␈α∞were␈α
considered
␈↓ β ␈↓ignorable␈α⊃in␈α⊂the␈α⊃speci≡c␈α⊂automatic␈α⊃code␈α⊂generation␈α⊃task␈α⊂at
␈↓ β ␈↓hand.

␈↓ β ␈↓␈↓ βPAlso␈α∞as␈α∞a␈α
result␈α∞of␈α∞this␈α
approach␈α∞to␈α∞system␈α
speci≡cation,
␈↓ β ␈↓each␈α∂Being␈α∞had␈α∂only␈α∞those␈α∂parts␈α∞speci≡ed␈α∂which␈α∂it␈α∞actually
␈↓ β ␈↓would␈α∂need␈α⊂in␈α∂the␈α⊂ensuing␈α∂dialogue.␈α⊂ Part␈α∂of␈α⊂the␈α∂di≠culty
␈↓ β ␈↓with␈α∞new␈α∂dialogues␈α∞stemmed␈α∂from␈α∞this␈α∂minimal␈α∞completion.
␈↓ β ␈↓In␈α∂the␈α∂protocol,␈α∂when␈α∂a␈α∂decision␈α∂was␈α∂made␈α∂by␈α⊂experts,␈α∂the
␈↓ β ␈↓knowledge␈α⊃necessary␈α⊂to␈α⊃follow␈α⊂the␈α⊃␈↓βother␈↓␈α⊃alternative␈α⊂branch
␈↓ β ␈↓was␈αnot␈αused,␈αnor␈αwere␈αsuch␈αsuper∨uous␈αfacts␈αsupplied␈αto␈αthe
␈↓ β ␈↓Beings␈αin␈αPUP6.␈αThus␈αthe␈α
user␈αof␈αPUP6␈αmust␈αalmost␈α
always
␈↓ β ␈↓resolve␈αeach␈αchoice␈αthe␈αway␈αthe␈αsimulated␈α(protocol)␈αuser␈α
did.
␈↓ β ␈↓It␈α⊃is␈α⊃felt␈α⊃that␈α⊃if␈α⊃all␈α⊂the␈α⊃parts␈α⊃of␈α⊃all␈α⊃the␈α⊃Beings␈α⊃had␈α⊂been
␈↓ β ␈↓faithfully␈α→≡lled␈α→in,␈α→this␈α→problem␈α→would␈α→have␈α→subsided.
␈↓ β ␈↓Basically,␈α∂the␈α⊂di≠culty␈α∂is␈α⊂one␈α∂of␈α⊂modelling␈α∂all␈α⊂the␈α∂possibly
␈↓ β ␈↓relevant␈αknowledge␈αan␈αexpert␈αhas,␈αrather␈αthan␈α(as␈αwas␈αdone)
␈↓ β ␈↓just␈α∂capturing␈α∂enough␈α∂of␈α∂his␈α∞knowledge␈α∂to␈α∂do␈α∂a␈α∂few␈α∞given
␈↓ β ␈↓tasks.

␈↓ β ␈↓␈↓ βPWhile␈α⊂all␈α⊂the␈α⊂Beings'␈α⊂interactions␈α⊂were␈α⊂invisible␈α⊂to␈α⊂the
␈↓ β ␈↓user,␈αthe␈αsystem␈αstill␈αswamped␈αhim␈αwith␈αdata␈αabout␈αwhat␈α
was
␈↓ β ␈↓going␈α
on.␈α
For␈α
example,␈α
most␈α
of␈α
the␈α
entities␈α
he␈α
was␈α∞asked␈α
to
␈↓ β ␈↓name␈α∂were␈α∂never␈α∞referred␈α∂to␈α∂again␈α∞by␈α∂name.␈α∂The␈α∞converse
␈↓ β ␈↓problem␈α
existed␈α
as␈α
well:␈α∞it␈α
was␈α
necessary␈α
to␈α
include␈α∞a␈α
Being
␈↓ β ␈↓which␈α⊗simulated␈α⊗forgetfulness,␈α⊗to␈α⊗prevent,␈α⊗e.g.,␈α⊗anaphora
␈↓ β ␈↓spanning␈α∞minutes␈α∂of␈α∞real␈α∞time.␈α∂ Orienting␈α∞the␈α∞user␈α∂was␈α∞not
␈↓ β ␈↓solved␈αsatisfactorally.␈α
Pointers␈αinto␈α
a␈αgraph␈α
of␈αgenerated␈α
code
␈↓ β ␈↓were␈αsimulated,␈αbut␈αoften␈αa␈αuser␈αwished␈αto␈αrefer␈αto␈αa␈αpiece␈αof
␈↓ β ␈↓code␈α↔not␈α↔by␈α↔name␈α↔or␈α↔by␈α↔pointing,␈α↔but␈α↔by␈α_some␈α↔brief
␈↓ β ␈↓meaningful (to him only!) phrase.

␈↓ β ␈↓␈↓↓␈↓&7.2. The Range of Programs Synthesized by PUP6␈↓)αβ␈↓

␈↓ β ␈↓␈↓ βPThe␈α∀system,␈α∪PUP6,␈α∀did␈α∪eventually␈α∀synthesize␈α∀CF,␈α∪the
␈↓ β ␈↓target␈α∞concept␈α∞formation␈α∞program.␈α∞ PUP6␈α∞was␈α∞200␈α∞pages␈α
of
␈↓ β ␈↓INTERLISP[6],␈α CF␈αwas␈α30␈αpages␈αlong␈α(6␈αpages␈αwhen␈αcoded
␈↓ β ␈↓by␈α∞hand␈α∂during␈α∞the␈α∂protocol).␈α∞CF␈α∞was␈α∂generated␈α∞in␈α∂60␈α∞cpu
␈↓ β ␈↓minutes␈α(compiled,␈αPDP-10␈α
TENEX).␈αThe␈αdialogue␈α
consisted
␈↓ β ␈↓of␈α∞300K␈α
characters␈α∞typed␈α∞by␈α
PUP6,␈α∞and␈α∞4K␈α
by␈α∞the␈α∞user.␈α
It
␈↓ β ␈↓occupied 300 pages, and ≡ve hours of real time.

␈↓ β ␈↓␈↓ βPDespite␈α⊂the␈α⊃lack␈α⊂of␈α⊃dialogue␈α⊂∨exibility,␈α⊃it␈α⊂␈↓βwas␈↓␈α⊃felt␈α⊂that
␈↓ β ␈↓most␈α∀of␈α∃the␈α∀Beings␈α∃could␈α∀be␈α∃useful␈α∀in␈α∃generating␈α∀other
␈↓ β ␈↓programs.␈α⊃For␈α∩this␈α⊃reason,␈α∩two␈α⊃additional␈α∩target␈α⊃programs
␈↓ β ␈↓were␈α⊂speci≡ed.␈α⊂They␈α⊂were␈α⊂synthesized␈α⊂with␈α⊂little␈α⊃change␈α⊂to
␈↓ β ␈↓PUP6, but only by someone familiar with the system.

␈↓ β ␈↓␈↓ βPThe␈α≠second␈α≠target␈α≠program,␈α≠GI,␈α≠is␈α≤a␈α≠grammatical
␈↓ β ␈↓inference␈α⊃program,␈α⊃which␈α⊃accepts␈α⊃strings␈α∩labelled␈α⊃LEGAL,
␈↓ β ␈↓ILLEGAL,␈α∃or␈α∀??.␈α∃In␈α∃the␈α∀latter␈α∃case,␈α∀GI␈α∃must␈α∃guess␈α∀the
␈↓ β ␈↓legality.␈α
Internally,␈α
potential␈α
sets␈αof␈α
rules␈α
are␈α
maintained.␈αOf
␈↓ β ␈↓the␈α⊃original␈α∩pool,␈α⊃49␈α∩out␈α⊃of␈α∩the␈α⊃87␈α∩ Beings␈α⊃were␈α∩used␈α⊃in
␈↓ β ␈↓synthesizing␈αboth␈αtargets.␈α Four␈αtotally␈αnew␈αBeings␈αhad␈αto␈αbe
␈↓ β ␈↓added,␈αrelated␈αto␈αformal␈αgrammars␈αand␈αrules.␈α Unfortunately,
␈↓ β ␈↓the␈α⊂addition␈α∂of␈α⊂␈↓βany␈↓␈α∂new␈α⊂Beings␈α∂demands␈α⊂that␈α∂the␈α⊂user␈α∂be



␈↓ β ␈↓␈↓¬Lenat␈↓ ¬bJuly 2, 1975␈↓ λ\page 12␈↓
␈↓ β ␈↓acquainted␈α∞with␈α∞the␈α∂format␈α∞conventions␈α∞of␈α∞PUP6.␈α∂ The␈α∞GI
␈↓ β ␈↓program␈αgenerated␈αwas␈α
20␈αpages␈αlong;␈αa␈α
hand-coded␈αversion
␈↓ β ␈↓was one-≡fth that size.

␈↓ β ␈↓␈↓ βPPL␈α∀was␈α∀the␈α∀≡nal␈α∪target␈α∀program␈α∀attempted,␈α∀a␈α∪simple
␈↓ β ␈↓property␈αlist␈αmanipulator.␈α
 It␈αrepeatedly␈αaccepts␈αrequests␈α
from
␈↓ β ␈↓the␈α∀user␈α∀to␈α∀insert,␈α∪inspect,␈α∀or␈α∀delete␈α∀some␈α∀record(s).␈α∪Any
␈↓ β ␈↓unspeci≡ed␈α⊗≡elds␈α⊗are␈α⊗treated␈α⊗as␈α⊗don't-cares,␈α⊗so␈α⊗a␈α⊗simple
␈↓ β ␈↓pattern-matcher␈α∞had␈α∞to␈α∞be␈α∞synthesized.␈α∞ Two␈α∞Beings␈α∞had␈α
to
␈↓ β ␈↓be␈α
␈↓βadded␈↓␈αto␈α
PUP6.␈αThe␈α
important␈αpiece␈α
of␈αdata␈α
is␈αthat␈α
about
␈↓ β ␈↓half␈αof␈αthe␈αoriginal␈αPUP6␈αpool␈αof␈αBeings␈αwere␈αactually␈αused
␈↓ β ␈↓in  ␈↓βall three␈↓ target-synthesizing dialogues.

␈↓ β ␈↓␈↓ βPAs␈α∀proposed␈α∀in␈α∀Section␈α∀6,␈α∀the␈α∀Beings␈α∀generate␈α∀other
␈↓ β ␈↓Beings,␈α→never␈α→plain␈α→functions.␈α→This␈α→explains␈α→the␈α_huge
␈↓ β ␈↓increases␈α_in␈α_target␈α_code␈α↔lengths␈α_in␈α_the␈α_PUP6␈α↔versions
␈↓ β ␈↓compared␈αto␈α
the␈αversions␈αproduced␈α
by␈αhand␈αwhen␈α
simulating
␈↓ β ␈↓the␈α
experts␈α
(who␈α
wrote␈αthe␈α
target␈α
programs␈α
as␈αfunctions).␈α
 CF
␈↓ β ␈↓was␈α
a␈α∞pool␈α
of␈α
56␈α∞brand␈α
new␈α∞Beings,␈α
GI␈α
37,␈α∞and␈α
PL␈α∞24.␈α
As
␈↓ β ␈↓with␈α
PUP6,␈α
one␈α
can␈α
interrupt␈α
the␈α
target␈α
programs␈α
as␈α
they␈α
are
␈↓ β ␈↓running␈α∞and␈α∞ask␈α∞questions.␈α∞Any␈α∞Being␈α∞on␈α∞the␈α∞control␈α∞stack
␈↓ β ␈↓will␈α∃provide␈α∃fully␈α∃instantiated␈α⊗answers␈α∃to␈α∃any␈α∃of␈α⊗its␈α∃29
␈↓ β ␈↓allowable␈α∩queries␈α⊃(its␈α∩parts);␈α⊃all␈α∩other␈α⊃Beings␈α∩will␈α⊃provide
␈↓ β ␈↓only␈α∞hypothetical␈α
answers.␈α∞Recall␈α∞the␈α
excerpt␈α∞from␈α∞CF␈α
itself
␈↓ β ␈↓running, found in section 5.2.

␈↓ β ␈↓␈↓ βPSome␈αof␈αthe␈αdi≠culties␈α
stem␈αfrom␈αthe␈αnature␈αof␈α
the␈αtask.
␈↓ β ␈↓In␈α∪any␈α∩long␈α∪dialogue,␈α∩the␈α∪user␈α∩often␈α∪forgets,␈α∪changes␈α∩his
␈↓ β ␈↓mind,␈α⊃errs,␈α⊃etc.␈α⊂A␈α⊃very␈α⊃sophisticated␈α⊂user␈α⊃model␈α⊃would␈α⊂be
␈↓ β ␈↓necessary␈α⊗to␈α⊗accomodate␈α⊗this␈α∃errorful␈α⊗process␈α⊗in␈α⊗a␈α∃non-
␈↓ β ␈↓debugging␈α⊃system.␈α∩Without␈α⊃such␈α∩abilities,␈α⊃the␈α∩system␈α⊃itself
␈↓ β ␈↓may␈α⊂be␈α∂led␈α⊂into␈α∂error.␈α⊂ While␈α∂most␈α⊂bugs␈α∂␈↓βare␈↓␈α⊂avoidable␈α∂by
␈↓ β ␈↓careful␈α⊗record-keeping,␈α∃it␈α⊗proved␈α∃unrealistic␈α⊗to␈α⊗make␈α∃no
␈↓ β ␈↓provision␈αfor␈αdebugging␈αa␈αnew␈αthirty-page␈αprogram.␈αWhen␈αa
␈↓ β ␈↓few errors did occur in CF, PUP6 itself had to be altered.

␈↓ β ␈↓␈↓ ¬7␈↓∧␈↓&8. Conclusions␈↓)αβ␈↓
␈↓ β ␈↓␈↓↓␈↓&8.1. About PUP6␈↓)αβ␈↓

␈↓ β ␈↓␈↓ βPWhat␈α∪have␈α∪we␈α∪learned␈α∪from␈α∪this␈α∪experimental␈α∩study?
␈↓ β ␈↓The␈α
overall␈α
feasability␈α∞of␈α
Beings␈α
was␈α
demonstrated,␈α∞but␈α
the
␈↓ β ␈↓di≠culties␈α∞of␈α
communicating␈α∞with␈α
the␈α∞user␈α
made␈α∞the␈α
system
␈↓ β ␈↓almost␈αimpossible␈αto␈αwork␈αwith.␈α
The␈αset␈αof␈αquestions␈αthe␈α
user
␈↓ β ␈↓was␈α
expected␈α
to␈α
want␈α∞to␈α
ask␈α
is␈α
the␈α∞same␈α
as␈α
the␈α
set␈α∞that␈α
one
␈↓ β ␈↓Being␈α
can␈α
ask␈α
another:␈αthe␈α
Being␈α
parts.␈α
When␈α
the␈α"nice"␈α
user
␈↓ β ␈↓interrupts,␈αhis␈αquestions␈αare␈αtranslated␈αtrivially␈αinto␈α
a␈αsimple
␈↓ β ␈↓retrieval.␈α⊃Real␈α⊃users␈α⊃are␈α⊃seldom␈α⊃nice;␈α⊃the␈α∩Beings␈α⊃generally
␈↓ β ␈↓misunderstood what users asked.

␈↓ β ␈↓␈↓ βPTo␈α∪modify␈α∪PUP6␈α∪to␈α∩synthesize␈α∪new␈α∪programs,␈α∪it␈α∩was
␈↓ β ␈↓necessary␈α⊃to␈α⊃add␈α∩a␈α⊃few␈α⊃general-purpose␈α∩programming␈α⊃and
␈↓ β ␈↓communication␈α
Beings,␈α
plus␈α
add␈α
several␈α
Beings␈α
speci≡c␈αto␈α
the
␈↓ β ␈↓new␈αprogram's␈α
domain,␈αplus␈αgeneralize␈α
a␈αfew␈αexisting␈α
Beings'
␈↓ β ␈↓parts.␈α⊂ The␈α⊂dialogue␈α⊂to␈α⊃produce␈α⊂the␈α⊂new␈α⊂program␈α⊃may␈α⊂be
␈↓ β ␈↓poorly␈α⊂suited␈α⊂to␈α⊂that␈α∂domain,␈α⊂since␈α⊂most␈α⊂of␈α⊂the␈α∂recognized
␈↓ β ␈↓phrases stem from a single (CF-producing) protocol.

␈↓ β ␈↓␈↓ βPTo␈α⊃improve␈α⊃PUP6's␈α⊃performance,␈α⊃one␈α⊃could␈α⊃add␈α⊂some
␈↓ β ␈↓debugging␈α∂specialist␈α⊂Beings,␈α∂some␈α∂dialogue␈α⊂specialists,␈α∂some
␈↓ β ␈↓sophisticated␈αuser␈αpsychology␈αexperts␈α(why␈αis␈αthe␈αuser␈αasking
␈↓ β ␈↓me␈α
that␈αquestion,␈α
what␈α
needn't␈αI␈α
tell␈α
him,␈αhow␈α
should␈αI␈α
direct
␈↓ β ␈↓his␈α
attention),␈αsome␈α
Beings␈αwhose␈α
task␈α
is␈αto␈α
aid␈αthe␈α
untrained
␈↓ β ␈↓user␈α
in␈α
inserting␈α
new␈α
domain-speci≡c␈α
Beings,␈α
and␈α
perhaps␈αa
␈↓ β ␈↓whole library of varied specialist Beings.

␈↓ β ␈↓␈↓↓␈↓&8.2. About Beings␈↓)αβ␈↓

␈↓ β ␈↓␈↓ βPThe␈α⊂performance␈α⊃of␈α⊂the␈α⊃Beings␈α⊂representation␈α⊃itself␈α⊂in
␈↓ β ␈↓PUP6␈αis␈αmixed.␈α Two␈αadvantages␈αwere␈αhoped␈αfor␈αby␈αusing␈αa
␈↓ β ␈↓uniform␈α∞set␈α
of␈α∞Being␈α
parts.␈α∞ Addition␈α
of␈α∞new␈α
Beings␈α∞to␈α
the
␈↓ β ␈↓pool␈α∂was␈α∂not␈α∂easy␈α∂(for␈α∂untrained␈α∂users)␈α∂but␈α∂communication
␈↓ β ␈↓among␈αBeings␈α␈↓βwas␈↓␈αeasy␈α(fast,␈αnatural).␈αTwo␈αadvantages␈αwere
␈↓ β ␈↓hoped␈α∩for␈α∩by␈α∩keeping␈α∩the␈α∩Beings␈α∩highly␈α∩structured.␈α⊃ The



␈↓ β ␈↓␈↓¬Lenat␈↓ ¬bJuly 2, 1975␈↓ λ\page 13␈↓
␈↓ β ␈↓interactions␈α⊃(especially␈α∩with␈α⊃the␈α⊃user)␈α∩were␈α⊃brittle,␈α∩but␈α⊃the
␈↓ β ␈↓complex tasks put to the pool ␈↓βwere␈↓ successfully completed.

␈↓ β ␈↓␈↓ βPThe␈α∂crippling␈α∞problems␈α∂are␈α∞seen␈α∂to␈α∞be␈α∂with␈α∞user-system
␈↓ β ␈↓communication,␈α≤not␈α≤with␈α≠the␈α≤Beings␈α≤ideas␈α≠themselves.
␈↓ β ␈↓Sophisticated,␈α
bug-free␈α
programs␈α␈↓βwere␈↓␈α
generated,␈α
after␈αhours
␈↓ β ␈↓of␈αfairly␈αhigh␈α
level␈αdialogue␈αwith␈αan␈α
active␈αuser,␈αafter␈αtens␈α
of
␈↓ β ␈↓thousands␈α⊃of␈α⊃messages␈α⊃passed␈α⊃among␈α⊃the␈α⊃Beings.␈α⊃ Part␈α⊃of
␈↓ β ␈↓this␈αsuccess␈αis␈αattributed␈αto␈αdistributing␈αthe␈αresponsibility␈αfor
␈↓ β ␈↓writing␈α∩code␈α⊃and␈α∩for␈α∩recognizing␈α⊃relevance,␈α∩to␈α∩a␈α⊃hundred
␈↓ β ␈↓entities,␈α∩rather␈α∩than␈α∩having␈α∩a␈α∩few␈α∩central␈α∩monitors␈α⊃worry
␈↓ β ␈↓about␈α
everything.␈α
 The␈αstandardization␈α
of␈α
parts␈α
made␈α≡lling
␈↓ β ␈↓in the Beings' contents fairly painless.

␈↓ β ␈↓␈↓ βPWhat␈α∂␈↓βare␈↓␈α⊂Beings␈α∂good␈α⊂for?␈α∂ For␈α⊂which␈α∂tasks␈α⊂won't␈α∂the
␈↓ β ␈↓problems␈α
encountered␈α
in␈α
PUP6␈αrecur?␈α
The␈α
idea␈α
of␈α
a␈α≡xed␈α
set
␈↓ β ␈↓of␈αparts␈α(which␈αdistinguishes␈αthem␈αfrom␈αACTORs)␈α
is␈αuseful
␈↓ β ␈↓if␈α∞the␈α∞mass␈α∞of␈α∞knowledge␈α∞is␈α∞too␈α∞huge␈α∞for␈α∞one␈α∞individual␈α∞to
␈↓ β ␈↓keep␈α⊃"on␈α⊃top"␈α⊃of.␈α⊃ It␈α⊃then␈α⊃should␈α⊃be␈α⊃organized␈α⊃in␈α∩a␈α⊃very
␈↓ β ␈↓uniform␈αway␈α
(to␈αsimplify␈αpreparing␈α
it␈αfor␈α
storage),␈αyet␈αit␈α
must
␈↓ β ␈↓also be highly structured (to speed up retrieval).

␈↓ β ␈↓␈↓ βPFor␈αthese␈αreasons,␈αthe␈α
author␈αis␈αcurrently␈αinvestigating,␈α
as
␈↓ β ␈↓a␈α∀potential␈α∀task␈α∪domain,␈α∀"research␈α∀in␈α∀elementary␈α∪number
␈↓ β ␈↓theory".␈α
 This␈αhas␈α
the␈α
added␈αbene≡t␈α
of␈α
isolating␈αthe␈α
problems
␈↓ β ␈↓in␈α
representation␈α
research␈α
from␈α
the␈α
staggering␈α
complexities␈α
of
␈↓ β ␈↓natural␈α∪language␈α∩handling.␈α∪Beings␈α∪are␈α∩big␈α∪and␈α∪slow,␈α∩but
␈↓ β ␈↓valuable␈α∩for␈α⊃organizing␈α∩knowledge␈α⊃in␈α∩ways␈α∩meaningful␈α⊃to
␈↓ β ␈↓how␈αit␈αwill␈αbe␈αused.␈αIn␈αthe␈αfuture␈αAutomated␈αMathematician
␈↓ β ␈↓system,␈α⊃Beings␈α⊃will␈α∩be␈α⊃one␈α⊃--␈α⊃but␈α∩not␈α⊃the␈α⊃only␈α∩--␈α⊃internal
␈↓ β ␈↓mechanism for representing and manipulating knowledge.

␈↓ β ␈↓␈↓ ¬P␈↓∧␈↓&References␈↓)αβ␈↓

␈↓ β ␈↓[1]␈α∂Green,␈α⊂Waldinger,␈α∂Barstow,␈α∂Elschlager,␈α⊂Lenat,␈α∂McCune,
␈↓ β ␈↓␈↓ βPShaw,␈α→and␈α~Steinberg,␈α→␈↓βProgress␈α→Report␈α~on␈α→Program-
␈↓ β ␈↓β␈↓ βPUnderstanding␈α#Systems␈↓,␈α"Memo␈α#AIM-240,␈α"Stanford
␈↓ β ␈↓␈↓ βPArti≡cial Intelligence Laboratory, August, 1974.

␈↓ β ␈↓[2]␈α
Hempel,␈α
Carl␈α
G.,␈α␈↓βFundamentals␈α
of␈α
Concept␈α
Formation␈αin
␈↓ β ␈↓β␈↓ βPEmpirical Science␈↓, Chicago: U. Chicago, 1952.

␈↓ β ␈↓[3]␈α
Hewitt,␈αCarl,␈α
␈↓βA␈α
Universal␈αModular␈α
ACTOR␈αFormalism␈α
for
␈↓ β ␈↓β␈↓ βPArti≡cial␈α,Intelligence␈↓,␈α,Third␈α-International␈α,Joint
␈↓ β ␈↓␈↓ βPConference on Arti≡cial Intelligence, 1973, pp. 235-245.

␈↓ β ␈↓[4]␈α
Minsky,␈αMarvin,␈α
␈↓βFrames␈↓,␈α
in␈α(P.␈α
Winston,␈α
ed.)␈α␈↓βPsychology
␈↓ β ␈↓β␈↓ βPof Computer Vision␈↓, New York: McGraw Hill, 1975.

␈↓ β ␈↓[5]␈α≠Newell,␈α~A.,␈α≠␈↓βProduction␈α~Systems:␈α≠Models␈α≠of␈α~Control
␈↓ β ␈↓β␈↓ βPStructures␈↓,␈α≡in␈α≥(W.␈α≡Chase,␈α≥ed.),␈α≡␈↓βVisual␈α≥Information
␈↓ β ␈↓β␈↓ βPProcessing␈↓, New York: Academic Press, 1973.

␈↓ β ␈↓[6]␈α↔Teitelman,␈α↔Warren,␈α↔␈↓βINTERLISP␈α↔Reference␈α⊗Manual␈↓,
␈↓ β ␈↓␈↓ βPXEROX PARC, Palo Alto, Ca., 1974.

␈↓ β ␈↓[7]␈α∂Winston,␈α∞Patrick,␈α∂␈↓βLearning␈α∞Structural␈α∂Descriptions␈α∞from
␈↓ β ␈↓β␈↓ βPExamples␈↓, MIT Project MAC, TR-231, September, 1970.

␈↓ β ␈↓[8]␈α∪Woods,␈α∀W.A.,␈α∪and␈α∀Makhoul,␈α∪J.,␈α∀␈↓βMechanical␈α∪Inference
␈↓ β ␈↓β␈↓ βPProblems␈α⊗in␈α⊗Continuous␈α⊗Speech␈α⊗Understanding␈↓,␈α∃Third
␈↓ β ␈↓␈↓ βPIJCAI, 1973, pp. 200-207.

␈↓ β ␈↓Many␈αhours␈αof␈αcreative␈αdiscussions␈αwere␈αquite␈αimportant.␈α In
␈↓ β ␈↓particular,␈αthe␈αauthor␈α
 acknowledges␈αC.␈αGreen,␈αR.␈α
Waldinger,
␈↓ β ␈↓D.␈α∂Shaw,␈α∂and␈α∂E.␈α∞Sacerdoti.␈α∂ Computer␈α∂time␈α∂was␈α∞generously
␈↓ β ␈↓provided by the Arti≡cial Intelligence Center of SRI.









␈↓ β ␈↓␈↓¬Lenat␈↓ ¬bJuly 2, 1975␈↓ λ\page 14␈↓
␈↓ β ␈↓␈↓ ∧:␈↓∧␈↓&Appendix 1: A Typical Being␈↓)αβ␈↓

␈↓ β ␈↓␈↓ βPWe␈α$consider␈α$␈↓¬INFO-OBTAINER␈↓,␈α$a␈α$Being␈α%which␈α$is
␈↓ β ␈↓independent␈α∞of␈α∞task␈α∞domain.␈α
 Below␈α∞is␈α∞listed,␈α∞for␈α∞each␈α
part,
␈↓ β ␈↓its␈α∂abbreviated␈α∂name␈α∂(in␈α∂bold),␈α∂an␈α∂English␈α∂question␈α∂that␈α∞it
␈↓ β ␈↓might␈α
(try␈αto)␈α
answer,␈αand␈α
the␈αstored␈α
program␈α
which␈αshould
␈↓ β ␈↓try␈αto␈αanswer␈αit␈α(often␈αa␈αsimple␈αtemplate␈αor␈αa␈αconstant).␈α The
␈↓ β ␈↓percentages␈α
given␈α
indicate␈α
how␈αmany␈α
of␈α
the␈α
(ultimately␈α100)
␈↓ β ␈↓Beings␈αin␈αPUP6␈αactually␈αused␈αthat␈αpart␈αduring␈αthe␈αsynthesis
␈↓ β ␈↓of one of the three target programs.

␈↓ β ␈↓¬␈↓↓WHAT␈↓¬␈↓ ∧⊂What do you do? Summarize your basic idea.␈↓ 	⊃82%
␈↓ β ␈↓¬  (OBTAIN SOME INFORMATION WHICH CAN BE USED)
␈↓ β ␈↓¬␈↓↓WHY␈↓¬␈↓ ∧⊂Justification? Summarize your motivation.␈↓ 	⊃77%
␈↓ β ␈↓¬  (PUP HAS NO MORE INFO THAT IT CAN USE TO PROGRESS)
␈↓ β ␈↓¬␈↓↓HOW␈↓¬␈↓ ∧⊂Global strategy? Summarize your method.␈↓ 	⊃72%
␈↓ β ␈↓¬  (OBTAIN NEW FACTS ABOUT OLD INFO, OR BRAND NEW INFO)
␈↓ β ␈↓¬␈↓↓IDEN␈↓¬  Do you recognize: "Find out more about frob
␈↓ β ␈↓¬  gyrnation"?      (if I see either phrase:
␈↓ β ␈↓¬  (INFO-OBTAINER any1) or (FIND OUT MORE ABOUT any1),
␈↓ β ␈↓¬  then I return: (INFO-OBTAINER (TRANSLATE any1)))␈↓ 	⊃54%
␈↓ β ␈↓¬␈↓↓EXPLICIT-ARGS␈↓¬␈↓ ¬     What argument(s) do you take?  (U)␈↓ 	⊃63%
␈↓ β ␈↓¬␈↓↓EVAL-ARGS␈↓¬␈↓ ∧P  Which are quoted, not evaluated?  NIL␈↓ 	!4%
␈↓ β ␈↓¬␈↓↓IMPLICIT-ARGS␈↓¬␈↓ ¬     What local variables are needed?  NIL␈↓ 	⊃11%
␈↓ β ␈↓¬␈↓↓WHEN␈↓¬␈↓ ∧⊂When should you take control (justify your answer)?
␈↓ β ␈↓¬((if T then add in -10 because ␈↓ 	⊃19%
␈↓ β ␈↓¬ (I AM EXPONENTIALLY-GROWING, GENERALLY UNDESIRABLE))
␈↓ β ␈↓¬(if New-Info-List then add in (Plus 100 (Length New-Info-List))
␈↓ β ␈↓¬because (WE SHOULD WORK ON UNASSIMILATED NEW
␈↓ β ␈↓¬INFORMATION IF THERE IS ANY)))
␈↓ β ␈↓¬␈↓↓REQUISITES␈↓¬  What must you ensure is true just before
␈↓ β ␈↓¬  (pre) and after (post) you go?    NIL␈↓ 	⊃10%
␈↓ β ␈↓¬␈↓↓DEMONS␈↓¬␈↓ ∧    What demons should you activate while you're
␈↓ β ␈↓¬  in control?    NIL␈↓ 	!7%
␈↓ β ␈↓¬␈↓↓META-CODE␈↓¬␈↓ ∧P  What happens when you are in control?␈↓ 	⊃70%
␈↓ β ␈↓¬  (DO  (CHOOSE-FROM  ((GET-NEW-INFORMATION U)
␈↓ β ␈↓¬␈↓ ∧⊂               (TRANSLATE U)
␈↓ β ␈↓¬␈↓ ∧⊂               (ANALYZE-IMPLICATIONS U)
␈↓ β ␈↓¬␈↓ ∧⊂               (EXTRACT-RELEVANT-SUBSET U)))
␈↓ β ␈↓¬    BECAUSE  (WE CAN ONLY TRY TO OBTAIN USABLE
␈↓ β ␈↓¬        INFO IN ONE WAY AT A TIME))
␈↓ β ␈↓¬␈↓↓COMMENTS␈↓¬   Do you have any special hints for filling
␈↓ β ␈↓¬  in undefined subparts of this Being?   NIL␈↓ 	⊃16%
␈↓ β ␈↓¬␈↓↓STRUCTURE␈↓¬   Viewing this Being as a data structure,
␈↓ β ␈↓¬  what can you do to it?  NIL␈↓ 	!4%
␈↓ β ␈↓¬␈↓↓MAIN-EFFECTS␈↓¬  Can you cause this goal to occur:
␈↓ β ␈↓¬  "Usable information exists"?␈↓ 	⊃27%
␈↓ β ␈↓¬  ((to get (NEW INFORMATION any1) or to get
␈↓ β ␈↓¬    (USABLE INFORMATION any1), do (INFO-OBTAINER any1)))
␈↓ β ␈↓¬␈↓↓AFFECTS␈↓¬␈↓ ∧   What other Beings might you call on directly?␈↓ 	⊃14%
␈↓ β ␈↓¬  ((CHOOSE-FROM is called)
␈↓ β ␈↓¬   (call on some Being who can satisfy the goal:
␈↓ β ␈↓¬      (AWARE USER (ABOUT TO OBTAIN USABLE INFO))
␈↓ β ␈↓¬   (GET-NEW-INFORMATION possibly is called)
␈↓ β ␈↓¬   (TRANSLATE possibly is called)
␈↓ β ␈↓¬   (ANALYZE-IMPLICATIONS possibly is called)
␈↓ β ␈↓¬   (EXTRACT-RELEVANT-SUBSET possibly is called) )
␈↓ β ␈↓¬␈↓↓GENERALIZATIONS␈↓¬  What Beings are more general than
␈↓ β ␈↓¬  you?  (WRITE-PROGRAM  SERVE-THE-USER)␈↓ 	⊃27%
␈↓ β ␈↓¬␈↓↓ALTERNATIVES␈↓¬␈↓ ∧p       What Beings are similar to you, to try
␈↓ β ␈↓¬  in case you fail?␈↓ ¬w16%    (USE-INFORMATION, OPTIMIZE,
␈↓ β ␈↓¬   FIX-INCORRECT-PIECE, and FILL-IN-UNDEFINED-SECTION)
␈↓ β ␈↓¬␈↓↓ENCODABLE␈↓¬     Any special constraints on what order
␈↓ β ␈↓¬  the parts should be evalled in?     NIL ␈↓ 	!9%
␈↓ β ␈↓¬␈↓↓COMPLEXITY-VECTOR␈↓¬  How costly are you?␈↓ 	⊃92%
␈↓ β ␈↓¬␈↓ ∧⊂(.5 .5 .9 .5 .1)
␈↓ β ␈↓A␈α∞vector␈α∞of␈α
utility␈α∞measures.␈α∞ The␈α
≡rst␈α∞component␈α∞says␈α
that
␈↓ β ␈↓␈↓¬INFO-OBTAINER␈↓␈α∩is␈α∩of␈α∪average␈α∩di≠culty␈α∩to␈α∩call.␈α∪Next,␈α∩there
␈↓ β ␈↓exists␈α⊃a␈α⊃.5␈α⊃chance␈α⊃that␈α⊂some␈α⊃descendant␈α⊃will␈α⊃call␈α⊃it␈α⊂again.
␈↓ β ␈↓Next:␈α∩this␈α∩activity␈α∩almost␈α∩always␈α∩succeeds.␈α∩The␈α⊃time/space
␈↓ β ␈↓used␈αin␈α
allowing␈αthis␈αBeing␈α
to␈αtry␈αis␈α
typical.␈αFinally,␈α
there␈αis
␈↓ β ␈↓no␈α∀good␈α∀reason␈α∀for␈α∃inhibiting␈α∀it␈α∀ever.␈α∀ In␈α∃general,␈α∀each
␈↓ β ␈↓component␈α⊃can␈α∩be␈α⊃a␈α⊃␈↓βprogram␈↓,␈α∩not␈α⊃just␈α⊃a␈α∩constant.␈α⊃ These
␈↓ β ␈↓weights,␈α∂like␈α∂the␈α∂contents␈α∂of␈α∂all␈α∂the␈α∂parts␈α∂of␈α∂all␈α∂the␈α∞Beings
␈↓ β ␈↓initially␈α∃in␈α∃the␈α∀experimental␈α∃PUP6␈α∃system,␈α∃were␈α∀decided
␈↓ β ␈↓upon and inserted by hand.
␈↓ β ␈↓␈↓¬Lenat␈↓ ¬bJuly 2, 1975␈↓ λ\page 15␈↓