BRIAN CANTWELL SMITH

THE SEMANTICS OF CLOCKS

The inexorable ticking of the clock may have
had more to do with the weakening of God's
supremacy than all the treatises produced by
the philosophers of the Enlightenment . .. .
Perhaps Moses should have included another
Commandment: Thou shalt not make mechan
leal representations of time.
- Neil Postman [1985, pp.11-12]
I. INTRODUCTION
Clocks?
Yes, because they participate in their subject matter, and participation
- at least so I will argue - is an important semantical phenomenon.
To start with, clocks are about time; they represent it.
l
Not only that,
clocks themselves are temporal, as anyone knows who, wondering
whether a watch is still working, has paused for a second or two, to see
whether the second hand moves. In some sense everything is temporal,
from the price of gold to the most passive rock, manifesting such
properties as fluctuating wildly or being inert. But the temporal nature
of clocks is essential to their semantic interpretation, more than for
other time representations, such as calendars. The point is just the
obvious one. As time goes by, we require a certain strict coordination.
The time that a clock represents, at any given moment, is supposed to
be the time that it is, at that moment. A clock should indicate 12
o'clock just in case it is 12 o'clock.
But that's not all. The time that a clock represents, at a given
moment, is also a function of that moment, the very moment it is meant
to represent. I.e., suppose that a clock does indicate 12 o'clock at noon.
The time that it indicates a moment later will differ by an amount that
is not only proportional to, but also dependent on, the intervening
passage of time. It doesn't take God or angels to keep the clock
coordinated; it does it on its own. This is where participation takes
hold.
3
James H. Fetzer (ed.), Aspects ofArtijiciallntelligence, 3-31.
1988 by Kluwer Academic Publishers.
4 BRIAN CANTWELL SMITH THE SEMANTICS OF CLOCKS
5
As well as representing the current time, clocks have to identify its
"location" in the complex but familiar cycle of hours, minutes, etc. They
have to measure it, that is, in terms of a predetermined set of temporal
units, and they measure it by participating in it. And yet the connection
between their participation and their content isn't absolute - clocks,
after all, can be wrong. How it is that clocks can participate and still be
wrong is something we will have to explain.
For clocks, participation involves being dynamic: constantly changing
state, in virtue of internal temporal properties, in order to maintain the
right semantic stance. This dynamic aspect is a substantial, additional,
constraint. A passive disk inscribed with 'NOW' would have both
temporal properties mentioned above (being about time, and having the
time of interpretation relevant to content) and would even maintain
perfect coordination. A rendering of this word in blinking lights,
mounted on an chrome pedestal, might even deserve a place on
California's Venice Boardwalk. But even though it would be the first
time piece in history to be absolutely accurate, such a contraption
wouldn't count as a genuine chronometer.
We humans participate in the subject matter of our thoughts, too,
when we think about where to look for our glasses, notice that we're
repeating ourselves, or pause to ask why a conversant is reacting
strangely. Why? What is this participation? It's hard to say exactly,
especially because we can't get outside it, but a sidelong glance suggests
a thick and constant interaction between the contents of our thoughts,
on the one hand, and both prior and subsequent non-representational
activity, on the other, such as walking around, shutting up, or pouring a
drink.
Take the glasses example. Suppose, after first noticing their absence,
I get up and look on my dresser, asking myself "Are they here?" My
asking the question will be a consequence of my wonder, but so will my
(non-representational) standing in front of the dresser. Furthermore,
the two are related; the word 'here' will depend for its interpretation on
where I am standing. And who knows, to drive the example backwards
in time, what caused the initial wonder - eye strain, perhaps, or maybe
an explicit comment. The point is that the representational and non-
representational states of participatory systems are inexorably inter-
twined - they even rest on the same physical substrate. We can put it
even more strongly: the physical states that realise our -thoughts are
caused by non-representational conditions, and engender non-represen-
tational consequences, in ways that must be coordinated with the
contents of the very representational states they realise. Participation is
something like that.
AI and general computational systems also participate - more and
more, in fact, as they emerge from the laboratory and take up residence
with us in life itself: landing airplanes, teaching children, launching
nuclear weapons. Far from being abstract, computers are part of the
world, use energy, affect the social fabric. This participation makes
them quite a lot like us, quite unlike the abstract mathematical expres-
sion types on which familiar semantical techniques have been developed.
My real reason for studying clocks, therefore, can be spelled out as
follows. First, issues of semantics, and of the relationship between
semantics and mechanism, are crucial for AI and cognitive science (this
much I take for granted). Second, it is terrifically important to recoguise
that computational systems participate in the world along with us.
That's why they're useful. Third, as I hope this paper will show,
participation has major consequences for semantical analyses: it forces
us to develop new notions and new vocabulary in terms of which to
understand interpretation and behaviour. Clocks are an extremely
simple case, with very modest participation. Nonetheless, their simplicity
makes them a good foil in terms of which to start the new development.
So they're really not such an unlikely subject malter, after all.
2. INFERENCE AND TIME-KEEPING
Let's start by reviewing the current state of the semantical art. Consider
a familiar, paradigmatic case: a theorem-prover built according to the
dictates of traditional mathematical logic. As suggested in Figure 1, two
relatively independent aspects will be coordinated in such a system.
First, there is activity or behaviour - what the system does - indicated
as 1jJ (for psychology). All systems, from car engines to biological
mechanisms of photosynthesis, of course do something; what dlstm-
guishes theorem provers is the fact that their ' implements (som,e
subset of) the proof-theoretic inference relation (f-). Second, there IS
the denotation or interpretation relation, indicated as <I> (for philoso-
phy), which maps sentences or formulae onto model-theoretic struc-
tures of some sort, in terms of which the truth-values of the formulae
are determined. In a computer system designed to prove theorems in
abstract algebra, for example, the interpretation function would map
6 BRIAN CANTWELL SMITH
p.
THE SEMANTICS OF CLOCKS 7
states of the machine (or states of its language) onto groups, rings, or
numbers ~ the subject matter of the algebraic axioms.
TIME
Fig. 1. Activity and semantics for a theorem prover
Four things about this situation are important. First, although proof
theory's putative formality suggests that W must be definable inde-
pendent of !P, you couldn't claim to have a proof-theoretic or inference
relation except with reference to some underlying notion of semantic
interpretation. Conceptually, at the very least, 'If and !P are inextricably
linked (salesmen for inference systems without semantics should be
reported to the Better Business Bureau). Furthermore, the two relations
are coordinated in the well-known way, using notions of soundness and
completeness: inferences (W) should lead from one set of sentences to
another only if the latter are true just in case the former are true (I-
should honour l=). And truth, as we've already said, is defined as in
terms of!p: the semantic relation to the world.
Second, even though the proof-theoretic derivability relation (I-)
can be modelled as an abstract set-theoretic relation among sentences, I
will view inference itself (W) as fundamentally temporal ~ as an
activity. 'Inference' is a derived noun; 'infer' first and foremost a verb,
with an inherent asymmetry corresponding directly to the asymmetry of
time itself. It might be possible to realise the provability relation
non-temporally, for example by writing consequences of sentences
down on a page, but you could hardly claim that the resulting piece of
paper was doing inference.
Third, when its dynamic nature is recognised, inference is (quite
properly) viewed as a temporal relation between sentences or states
of the machine's memory, not as a function from times onto those
corresponding sentences or states. Mathematically this may not seem
like much of a difference, but conceptually it matters a lot. Thus, taking
(1 to range over interpretable states of the system, and t over times, W
is of type (1 ..... a, not t ..... a. Of course it will be possible to define a
temporal state function of the latter type, which I will call 2:; the point
is that it is W, not 2:, that we call inference. Details will come later, but
the relation between the two is rougWy as follows: if t' is one temporal
unit past t, and 2:(t) a, then 2:(t') = W(a). Inference, that is, has
more to do with changes in state than with states themselves. To study
inference is to study the dynamics of representational systems.
Fourth, of all the relations in Figure 1, only W need be effective;
neither !P nor 2: can be directly implemented or realised, in the strong
sense that there cannot be a procedure that uses these functions' inputs
as a way of producing their outputs (the real reason to distinguish W
and 2:). This claim is obviously true for !P. If I use the name 'Beantown'
to refer to Boston, then the relation between my utterance and the town
itself is established by all sorts of conventional and structural facts
about me, about English, about the situation of my utterance, and so
forth. The town itself, however, isn't the output of any mechanisable
procedure, realised in me, in you, or in anyone else (fortunately ~ as it
would be awfully heavy). It might require inference to understand my
utterance, but that would only put you in some state a with the same
referent as my utterance, or state. You don't compute the referent of an
utterance you hear, that is, in the sense of producing that referent as an
output of a procedure. Nor is the reference relation directly mediated, at
least in any immediate sense, by the physical substrate of the world.
Not even the NSA could fabricate a sensor, to be deployed on route
128, that could detect Boston's participation as a referent in an infer-
ential act.
2
That 2: isn't computed is equally obvious, once you see what it
means. The point is a strong metaphysical one: times themselves
metaphysical moments, slices through the flux quo aren't causally
efficacious constituents of activity; they don't have causal force. If they
were, clocks wouldn't have been so hard to develop.3 As it is, mecha-
nisms, like all physical entities, manifest whatever temporal behaviour
they do in virtue of momentum, forces acting on them, energy expended,
etc., all of which operate in time, but don't convert time, compare it to
anything else, or react with it. The only thing that's available, as a
determiner of how a system is going to be, is how it was a moment
before, plus any forces impinging on it. That, fundamentally, is why
8
BRIAN CANTWELL SMITH
THE SEMANTICS OF CLOCKS 9
inference is of type a - a, not t - a. It could not be otherwise. The
inertness of gold, and the indifference of a neutrino, are nothing as
compared with the imperturbability of a passing moment.
Given these properties of theorem provers, what can we say about
clocks? Well, to start with, their situatiou certainly resembles that of
Figure 1. As in the inference case, a clock's being in some state a
represents (<I it's being noon, or 7 : 15, or whatever; the interpretation
function is what matters. Similarly, clocks, like theorem provers, change
state ('If) in a simple but important way. Not only that; state cbange
is what the clock designer has to work with; no mortal machinist,
unfortunately, could build a device that would directly implement L.
Furthermore, as in the case of the theorem prover, the change in state
of the clock face is important only because of its relation to its content.
Forget the Better Business Bureau; no one would buy a clock without a
clue as to how its states represented time. Once again, systematic
coordination between activity and interpretation is what matters.
But despite these similarities, there is a difference between clocks
and theorem provers - suggested by the fact that many people
(including me) would be reluctant to say that a clock was doing
inference. To get at the difference, note that we haven't yet said what
inference's coordinated pattern of events is for (on the face of it, goiug
from truths to truths sounds a little boring). But the answer isn't hard to
find: given a set of senteuces or axioms that stand in (or enable you to
stand in) a given semantical or informational relation to a subject
matter, proofs or inference lead you to a new informational relation to
the same, unchanged, subject matter. For example, the famous puzzle of
Mr. Sand Mr. p4 focuses your attention on a pair of numbers under a
peculiar description; a considerable amount of inference is required in
order to give you access to those same numbers under a more tradi-
tional description (or give you access to other more familiar properties
of numbers - there are many ways to discharge the ontological facts).
The numbers themselves, however, and their possession of all the
relevant properties, are expected to stay put during the inferential
process. None of this implies, of course, that the subject matter of
inference cannot itself be temporal, as the situation calculus and
temporal logics illustrate. The point is that the temporality of the
inference process and the temporality of the subject domain aren't
expected to interact.
The situation for clocks, on the other hand, is almost exactly the
opposite. What changes, across the time slice mediated by \11, isn't the
stance or attitude or property structure that clocks get at. What
changes, rather, is the subject matter itself. Clocks never have a
moment's rest; no sooner have they achieved the desired relationship to
the current time than time slips out from under their fingers, as if God
were constantly saying "!t's later than you think!" Clocks should perhaps
be viewed as the world's first truth maintenance systems: they do what
they do merely in order to retain the validity of their single semantic
claim. Like any other meter or measuring instrument, they must track
the world.
We can summarise:
Inference, at least as traditionally construed, is a technique
that enables you to change your relation to a fixed subject
matter. Clocks, in contrast, maintain a fixed relationship to a
changing subject matter.
If reconstructing time-pieces were really my subject matter, rather than
simply being a foil, I might stop here. But my real interest is in
developing a single semantical framework so that we can not only
handle both of these cases (mathematical inference and real-time
clocks), but also locate everything in between. So let's spend a minute
to see how clocks fit into the general case.
3. SEMANTICALLY COHERENT ACTIVITY
I will use the term 'representational system' to coyer anything whose
behaviour fits within the broad space of semantically constrained
activity. To be a representational system, in other words, is to be an
element of the natural order that acts in a semantically coherent way.
Of all possible kinds of representational activity, inference will be
analysed as a particular type. The representational space is large, of
course, and certainly includes all of computation (more about that in a
moment), but it's still a substantive notion: not everything is in it.
Planets, for example, are excluded, because planets don't represent
their orbits; they just have them. Clocks, on the other hand, do
represent the time, just as I can represent to myself how the sunrise
looked this morning, as I drove down from the mountains.
Clocks do however fall outside most traditional models of computa-
tion, including the "formal symbol manipulation" model so familiar in
10 BRIAN CANTWELL SMITH
THE SEMANTICS OF CLOCKS
11
cognitive science.
5
First, clocks (their faces, and the clockworks that run
them) are fully concrete, physical objects, part of the natural order;
nothing abstract here. Furthermore, this concreteness is crucial to our
understanding of them; for some purposes one might treat clocks at a
level of description that abstracted away from their physical being,
including their temporal being, but since OUf purpose is to show how
participation in their subject matter influences their design, to do so
would be to miss what matters most. Second, at least some clocks
(especially electrical ones operating on alternating current) are analog,
even though more and more recent ones are digital. Third, to the extent
that clocks have representational ingredients, there is no obvious
decQupling to be made between a set of structures that represent and
an independent process that inspects and manipulates them according
to the shapes it sees. In other words, whereas Fodor's characterisation
of a computer's "standing in relation" to representational ingredients
suggests a modular division between symbols and processor, no such
division is to be found in the chronological case. Fourth, there is
another separation that can't be maintained in the case of clocks: that
between "internal" and "external" properties. Time (rather like neu-
trinos) permeates everything equally, being as much an influence on
internal workings as on surrounding context. And of course it is one
and the same time, inside and out - clock design depends on this.
Fifth, clocks, especially analog clocks, aren't usually "programmed" in
any sense; they are designed, but they aren't universal computers
specialised by physical encodings of time-keeping instructions. Like so
many other properties of clocks, this is important, and leads to the sixth
salient difference. Even on the view that Turing machines are concrete,
physical objects (of which abstract mathematical quadruples are merely
set-theoretic models), there is still no guarantee, given a particular
universal one, that any set of instructions could make it be, or even
simulate, an accurate time keeper - because there need be no
consistency or regularity as to how long its state changes take. Turing
machines, qua Turing machines, don't really participate.
I have come to believe, however, that not one of these six properties
- being abstract, being digital, exhibiting a process/structure dichot-
omy, having a clear boundary between inside and outside, being
programmable, or being necessarily equivalent to any Turing machine
- is essential to the notion of computation on which the economy of
Silicon Valley is based, or to the notion that underlies AI's hunch that
the mind is computational. Quite the contrary. In (Smith, forthcoming) I
argue for a much stronger conclusion: that the only regularity essential
to computation has to do with computation being a physically embodied
representational process - an active system or process whose behav-
iour represents some part or aspect of the embedding world in which it
participates. This has the consequence, needless to say, of defIling
computation squarely in terms of undischarged semantIcal
My position on theoretical cartography is therefore the mverse of
Newell's (1980): whereas he dUnks that computer science has answered
the question of what it is to be a symbol, I believe in contrast that the
integrity of computation as a notion rests full-square on semantIcs. So
we have lots of homework, but it's homework for another day.
In the meantime, clocks are a good test case for comprehensive
semantical frameworks. They lack many important properties of more
general computers: they don't act, for example, or have sensors. But
since every semantical property they do exhibit is one that computers
can exhibit too - including participation - they are a useful design
study.
4. THREE POINTS ON TWO FACTORS
In the previous section I distinguished two aspects or factors of
representational system: its behaviour, activity, or causal connectlOn
with the world (which I'll call the first factor) and its interpretation,
content or relation to its subject matter (the second factor). I have
previou'sly used this two-factor framework to reconstruct the semanti.cs
of Lisp, the programming lingua franca of AI, and argued for. Its
general utility in analysing knowledge representatlOn systems (Snuth,
1982,1984, 1986). And I will use it here, to analyse clocks. But three
points must be made clear. .
First, the ordering of the two factors may seem odd. There IS no
doubt that having interpretation or content -standing in semantic
relation to a subject matter - is what particularly distinguishes the
systems we are interested in. Given this pride of place, it might seem
that content should be called first. But this is a mistake. We theore-
ticians typically treat semantics as primary when we analyse both
natural and artifactual languages (such as the predicate calculus). We
typically define semantics over rather abstract entities -
types, for example - and then understandably define the other dunen-
12 BRIAN CANTWELL SMITH THE SEMANTICS OF CLOCKS 13
sian (proof theory, inference) over the same domain. But this overall
sIrategy, especially in conjunction with the formal-symbol manipulation
view of computation, gives a very abstract feel to inference, leading
such people as Searle to wonder how, or even whether, such a system
could ever possess genuine semantical powers. In contrast, by calling
activity the first factor I want to recognise that computational systems
.- are, first and foremost, systems in the world. Everything has what I am
calling a first factor; that's what gives a system the ability to participate.
The second factor of representation or content, which enables a system
(a thinker, a clock) to stand in relation to what isn't immediately
accessible or discriminable, is a subsequent, more sophisticated capac-
ity. It is the second factor, furthermore, that distinguishes the repre-
sentational or interpretable systems from other natural systems, but it
distinguishes them as a sub-type, not as a distinct class. First factor
participation in the world ("being there", roughly) is always available -
which is fortunate, since it is only with respect to the first factor that
second factor content can ever be grounded. In sum, recognising the
metaphysical primacy of the first factor is an important ingredient in
the defense of naturalism.
Second, there is a natural (almost algebraic) tendency to think that,
in accepting a two-factor stance, one is committed to thinking that the
two factors, in any given system, will in some important sense be
independent. This tendency is amplified by the fact that in standard
first-order logic an almost total independence of factors is achieved -
this is one of the many meanings of the ambiguous claim that first-order
logic is formal. Truth, content, and interpretation in logic are thought to
be relatively independent of proof-theoretic role, and provability or
inferential manipulation analogously independent of content or inter-
pretation. In fact it is only because of this conceptual independence
that ,proofs of soundness and completeness, even the very notions of
soundness and completeness, are conceptually coherent. In computer
systems, however ~ and minds, and clocks ~ there is no reason to
expect this total degree of disconnection or independence. We should
expect something more like the relationship between the mass and
velocity of a physical object, on the one hand, and the center of gravity
or resonance of the system of which it is a part, on the other: a web of
constraints and conditions tying the two factors together, piece-wise,
incrementally, thereby giving rise to a comprehensive whole. The
situation of a complete proof system defined on an abstract set of
mathematical expression types is extreme: a global but locally un-
mediated coherence, with no part of the proof or inferential system
touching the semantic interpretation or content, except in the final
analysis, when an outside theorist's proof grandly ties the whole thing
together. For computers, and for us, it seems much more plausible to
take a step or two apart from our subject matter, and then check in with
it, to stay in "sync" - by taking a look, for example, or (following
AT&T's recommendation) by reaching out and touching it. Participa-
tion is a resource, not a complication.
Third, as both the first two points make clear, it's a little hard to
justify calling the two factors semantical, especially when the first is
shared with every other participant in the natural order. It's not just that
the first should be viewed as syntax, the second as semantics (as
application of this more general framework to the predicate calculus
would suggest). Rather, it's not clear what, if anything, the terms
"syntax" and "semantics" should mean in a context where the coupling
between factors is so much richer and more complex than in the
traditional idealised case - if indeed they mean anything at all.
Clockworks are mechanisms that enable first-factor behaviour - that
much seems innocuous enough; calling the momentum of a clock's
pendulum semantic is more difficult. First and second factors aren't
distinct objects that somehow cooperate in engendering semantical
activity; rather, one and the same causal constituents of a semantic
system play both first and second factor roles.
This whole question is complicated by the use of the word 'semantics'
(especially in AI) to describe inferential and structural relations among
ingredients within a computational system. In (Smith, 1986) 1 attempt
to resolve some of these issues, but instead of reconstructing that
argument here I'll simply use the two-factor terminology without preju-
dice as to what does and doesn't have legitimate claim to the overloaded
term.
5. THEORETIC MACHINERY AND ASSUMPTIONS
Let's look, then, at how clocks represent time, starting with some basic
assumptions. As suggested in Figure 2, qua theorists we need accounts
of four things:
1. States of the clock itself, including the face (a);
14
BRJAN CANTWELL SMITH THE SEMANTICS OF CLOCKS 15
2. The typology of clock semantics

<p (represents)
Fig. 3. The model-theoretic approach
a length of 3). The general character and complexity of the model-
clock relation M(J-a, therefore, is the same as that between the clock
and the time it represents (a-i). It is therefore very hard to know
whether what is crucial about a-T will be revealed or hidden in its
Ma-M. form. For example, using simple numbers to represent the
orientations of hands presumes an absolute accuracy on the clock face,
counter to fact. When studying something like natural language, which
makes use of a much more complex representation relation than a
model, the problems of indiscriminate theoretic modelling may be
minor, or (more likely) go unnoticed. In our case, however, the
representation relation we are studying, between clock faces and
periodic times, is essentially an isomorphism. In this situation indis-
criminate modelling would be much more theoretically distracting.
This direct semantical stance will have consequences, of two main
sorts. First, we will need some machinery for talking precisely about the
world without modelling it; for this I will use an informal "pocket
situation theory", based unapologetically on Barwise and Perry (Barwise
and Perry, 1983; Barwise, 1986a). Second, in the analog case it will
be tempting to use some elementary calculus, which is problematic
because a situation-theoretic reconstruction of continuity hasn't been
developed yet On the other hand, since the continuities underlying the
integrity of the calculus presumably derive, ultimately, from the funda-
mental continuity of the phYSical phenomena that the mathematics was
2.
3.
4.
The time or passage of time that the clock represents (r),
The first factor movement or state change between' clock
states (P); and
The second factor representation relation (CP) between clock
states and times.
All four of these are shared with standard semantical analysis; the first
two would be the syntactic and semantic domains; the third, inference
or proof theory; the fourth, semantics or interpretation.
Theorists'
represent-
tations;
I will adopt what I will call a direct rather than model-theoretic
approach to these analytic tasks. Typically, When doing semantics,
instead of talking directly about clock faces, orientations of hands, etc.,
you model them. For example, the state of a three-hand analog clock
mtght be modelled as a triple, consisting of the orientations of the
hour-hand, minute-hand, and second-hand, respectively, measured
clock-wise from the vertical, in degrees. Thus the clock face shown in
Figure 2 would be modelled as follows:
(Sl) M,,: (128.31666 ... ,99.8,228)
The problem with this technique, however, as suggested in Figure 3, is
that a model M of a situation S is itself a representation of S, since
modelling is a particular species of representation (M", for example,
represents the clock face; it isn't the clock face, since for example it has
16 BRIAN CANTWELL SMITH THE SEMANTICS OF CLOCKS 17
Objects and Properties
PrimaryTheoreticFunctions
As opposed to times themselves, I will assume that times are located
on the periodic cycle by what I will call the o'clock properties, such as
that of "being 4: 01: 23", "being midnight", etc. The idea is not so
much to license a continuum of distinct properties, but rather to assume
that they arise out of a continuous relation between times and the
abstract locations on the periodic time cycle to which they correspond
Given these preliminaries, we should set out the ontological type
structure, as summarised in Figure 4. Variables ranging over objects
will be spelled with lower-case italic letters; over properties and
relations, in lower-case Greek; over functions, in upper-case Greek.
Thus c and c' will range over clocks; t, I', etc., over full-blooded times;
which are taken to be instantaneous slices through the metaphysical
flux. Times are meant to include the time Kennedy was shot, the
referent of "now", the point when the ship passed out of sight behind
the island ~ that sort of thing. Intervals - intuitively, temporal
durations between times - will be indicated by Ill, Ilt', etc. I will
extend the use of '+' to allow adding intervals to times (overloading '+',
as computer scientists say); thus I + III will be of type I.
clocks
times (instantaneous moments: slices through the flux quo)
temporal intervals
o'clockproperties:beingmidnight, being4 : 01 : 23,.
T/- the o'clockproperty that holdsof time t
states ofclock faces (both hands pointing upwards, ...)
0c, ,-thestate ofclock c at time t
times plus intervals are times
o'clockproperties plus intervals are o'clock properties
Fig.4. Theoretictypestructure
clockworks (fromclock states and intervals onto clock states)
statefunction (fromclocks and timesonto clockstates)
content function (fromclock states onto o'clock properties)
0,0', ..
T, T',.
c,c',.
t, t', . ..
D..t,D..t',.
t+6.t: t
T+D..t: T
-qr: OJ 6.t ...... a
L :c,t ......
[...1' 0- T
Overloaded Addition
developed to describe, and since exactly those continuous phenomena
will be our subject matter here, I'll take the liberty of applying its
insights anyway. We're not really going to do any mathematics, so we
won't get into trouble.
The direct semantical stance also highlights a question: how as
theorists are we going to describe or register the phenomena we 3fe
--.- -going to study - i.e., in terms of what concepts, categories, and
constraints afe we going to explicate its regularity? When giving
semantical analyses of linguistic or syntactic obj eets (sentences, expres-
sion types, etc.), tradition provides standard registrations in terms
of constituent terms, predicate letters, etc. Similarly, purely abstract
objects are typically categorised in advance in terms of a defining set of
properties or relations. Clocks, on the other hand, afe neither tradi-
tional nor abstract, so the question remains.
. My metaphysical bias is to treat the world as infinitely rich, not only
ill the sense that there is more to everything than anything we can say,
but also in that there is both more uniformity and structure, and more
heterogeneity and individual difference, than theory or language can
ever encompass. So I will say that clock faces, being actual, have
enough structure so that one can be wrong about them, but still don't
come labelled in advance by God, like plant slips at a nursery identified
with a white plastic tag. Since every clock face, furthermore, exemplifies
an infinite number of properties and relations (such as the property of
being the subject matter of this paragraph), even after a basic registra-
tion scheme has been settled on, we have considerable latitude in
making our choice.
None of this is intended to be problematic, or new; it's worth
mentioning only because we need to make room for there being a
difference between how we theorists do it, and how clocks do it, for
themselves or (more likely, in the case of clocks) for their users. The
problem is particularly acute for time itself, especially the periodic cycle
of hours, minutes, and seconds that I keep referring to without explana-
tion. If this were a paper on the semantics of time, not just of clocks,
that explanation would have to be given, which would raise the inces-
tuous fact that clocks themselves are probably largely responsible for
the temporal registration (hours, mintues, seconds, etc.) of the times
they represent, as argued for example by Lewis Mumford (1934). In
thIS paper, however, I will merely adopt the periodic cycle without
analysis, taking its explanation as a debt that should ultimately be paid.
18 BRIAN CANTWELL SMITH THE SEMANTICS OF CLOCKS 19
6. TEMPORAL REPRESENTATION: THE SECOND FACTOR
Given these premises and caveats, let's look at how times are repre-
sented. Intuitively, we are aiming for something like the following:
To do this, we start with (J:l, of type U -> T from (representing) states
of clock faces onto (represented) states of times i.e., onto o'clock
properties. Instead of the name '(J:l', however, I will use so-called
semantic brackets C[ . ..],), in the following way: [... (fe, I ] will be the
o'clock property signified by the state' ... U
C

1
.', assuming that 0c, I is
the state 0 of clock c at time t. For example, the sentence (UC,I.](I)
the property of being 4: 16
(S2)
physical plausibility. Even if quantum physics would theoretically sup-
port there being a fact of the matter as to where a hand points within
10-
50
degrees, say (which it won't), there are also pragmatic realities
of producing a macroscopically observable clock subject to the forces
of gravity, anomalies of manufacture, etc. Furthermore, if the hour-hand
were anything like this accurate, then at least for theoretical purposes
the minute and second hands would be redundant: a perfect observer
could gaze at a clock and read off a time of, say, 4: 15 : 38 : 17.
7
One
might object, of course, that human users wouldn't be able to register
the hour-hand more accurately than, say, 1 or 2, and therefore,
even with internal calculation, wouldn't be able to determine the time
on a single-handed clock more accurately than to within about 5
minutes, no matter how much more accurately than that the time was
actually signified. In fact casual observation suggests that hour handS on
modern analog clocks are much more accurately positioned than
necessary merely to determine which hour the minute hand signifies
time with respect to.
These issues again raise the question of the relation between how we
as theorists register clock faces and the times they represent, and how
clock faces themselves register those represented times.
s
But I won't
answer this question here, since we will primarily be dealing with
semantic constraints on clock and time registrations, rather than with
individual registrations themselves.
("4: 00", etc.). Various possible explanations of this relation are possi-
ble, but since the intent of this paper is not to present an independently
justified metaphysical account of time, but only to relate clocks to such
a thing, I will employ a notation that simply picks up o'clock properties,
whatever they are, from times that have them. Thus I will use 't to refer
to the particular o'clock property that actually holds of time t. Also, I'll
take differences between o'clock properties to be intervals (e.g., the
difference between 5: 00 and 3: 00 will be two hours). Thus the
sentence 'I(t') says of time {' that it has o'clock property 7:( i.e., that it
has whatever o'clock property t has. The term T
1
- TI' denotes an
interval, of type 6.t.
6
In an analogous way, G, G', etc. will range over a continuous (in the
analog case) set of states of clock faces. For traditional circular analog
clocks, a (f representing 4.30 might be "having the hour hand at 135,
the minute hand at 180, and the second hand at 0, all measured
clockwise from the 'XII'."
Given this framework we can type the various semantical functions
already encountered. As suggested in the previous section, L will be a
(non-computed!) function of type t -> G, from times onto clock states;
lV, a function of type U j 6.t -> U, from clock states and temporal
intervals onto clock states; and (J:l j a function of type G -> r, from
clock states onto o'clock properties. The important typological point
for general semantic analysis is that both factors (lV and (J:l) are defined
as functions between states objects can be in, not between objects that
are in them. This is as you would expect for scientific laws.
Two more theoretical points, before we take up the analysis itself.
First, as just mentioned, I claimed in Section 2 that times t weren't
causal agents - that they couldn't be in the domain of a strongly
effective realisable function. It is probably more important to the life of
clock designers that the o'clock properties (.) are equally impotent.
Even if it's 4: 00 all around you, there's nothing that it's being 4 : 00
can cause to happen - like serving tea and crumpets. With respect
to engendering behaviour, a moment's being midnight is more like
Boston's being a referent than it is like being sticky: it just
isn't the sort of thing that a sensor could detect. So functions of the
form.. -> x are as unrealisable (in the strong sense discussed earlier)
as those of type f -> x, for arbitrary x. Such is life.
Second, I mentioned earlier that using numbers to represent the
orientations of the hands of clocks presumes an accuracy that outstrips
20 BRIAN CANTWELL SMITH THE SEMANTICS OF CLOCKS 21
(54)
claims of time I that it has the o'clock property that clock c indicates at
time 1'; 1'1'(a",., lil)I(I) claims of time I that is has the o'clock property
that clock c would indicate lil later than time 1', since '1'(0,.,., lil) is
the state it would then be in.
Using this terminology, we can say that clock c is chronologically
correct at time t just in case t is of the type that the clock then
indicates:
(53) Correct(c, I) =dr[o,,,J(I)
So far, of course, this is a constraint on possible interpretation functions
[...1, since we haven't defined any specific instances. Longer-term
notions of correctness (over extended intervals, for example) could be
defined by quantifying over times; similarly, approximate degrees of
correctness could be characterised in terms of the difference between
what time it actually was and what time was indicated.
7, CLOCKWORKS, THE FIRST FACTOR
With respect to operation, the basic point is this: if at time t a clock is
so-and-so (0), then at some point lil later it will be such-and-such (0'),
where 0' is '1'(0, lil). The function '1', which takes a clock into the
future in this way, must be realised by the underlying physical machine
- must be implemented, that is, by the clockworks. The important
constraint on this relation, which I will call the realisability constraint, is
that '1'(0, lil) can depend on 0 and on lil, but not on the time I that is
"happening" when the clock is in state o.
In symbol manipulation or semantical contexts, where time and
symbols are both digital, we often view \)J as a state-transition function
(such as for a Turing machine controller). In such cases lil drops out,
being assumed to be a single time "click". For example, suppose is a
(discrete) function from states to states (0 - 0). The equation for a
single state change, of the sort one would expect in a digital world,
would be something like 0' ~ (a), or generalised to lit's of n tick's
duration, a' = n( a). In the continuous world of physical mechanics,
on the other hand, '1' is merely "what the world does", explained in
terms of velocities, accelerations, etc. From this perspective, the calculus
can be viewed as a theoretical vehicle with which to explain first factor
futures for continuous systems, where the state a of some system in an
amount of time I::1t after it is in a starting state ao, assumed to depend
on the continuity of the underlying phenomena, can be expressed in the
familiar equation:
da 1 d'a
o ~ a . +-lit+- --liI
2
+.
o dt 2 dl,
My aim isn't to contrast the discrete and continuous cases (I want to
develop results applicable to both analog and digital clocks), but rather
to highlight the common focus on state change, represented computa-
tionally by state transition functions, and physically by temporal deriva-
tives. There is, however, this apparent difference: the theoretic notions
employed in physics (force, acceleration, etc.) are essentially relative;
they describe how the new state will differ from the old one. The real
identity of the new state - what state the system will actually arrive in
- is obtained, as if it were conceptually subsidiary, by altering the
previous state in the prescribed manner. State transition tables, in
contrast, are typically absolute. They still describe state change, of
course - they aren't temporal state functions like L. The point is that
the new state is specific de novO, so to speak, not as a modification of
the old one, though of course the extent to which the new state differs
from the old can be calculated as a difference between the two.
This difference in theoretic stance, however, is superficial, since in
actual use (in describing programs, operations on memory, etc.) state
transition functions are defined with explicit reference to how the new
state differs from the old. In giving environment transition functions, for
example, showing the consequence of binding a variable, the requisite
function from total environments onto total environments is defined as
modifying the value of the given variable in question, and otherwise
being just like the prior one. Practice suggests, in other words, that
in the computational case, as in the physical case, state change is
conceptually prior, new total state conceptually dependent. Thus there
is general support for our specific focus on q.r.
Intuitively, a proper '1' for a clock will specify that it runs at the right
speed. It is easy enough to calculate, in the case of circular analog
clocks, that this amounts to having the hour hand, minute hand, and
second hand rotate at 0.008333'/5ec, 0.1'/sec, and 6'/sec, respectively.
But to characterise correctness this way is exactly like characterising the
correctness of a proof procedure by pointing to the syntactic inference
rules. It may indeed be true that, if this condition is met, the clock will
be running at the correct speed, but that doesn't mean that this
condition expresses what it is to be running correctly. Rather, we want
22 BRIAN CANTWELL SMITH THE SEMANTICS OF CLOCKS
23
(59)
(SI3) Correct(c, I) "" [L(c, 1)1(1)
it foIlows that the constant would be 0 for a correctly set clock, as
expected.
We can summarise these results as follows:
identically 1, yielding:
(SID) 1 limit (la",+",] - la" ,I)
D.I --> 0
The right hand side, however, is merely the derivative, with respect to
time, of the interpretation of the state, We can't differentiate a directly,
its not being a function of time (in fact it's not a function at all), but we
can rewrite (S10) in terms of L:
I
lmu' 't (IL(c, I + M)j - IL(c, 1)1)
(SIl)
6.1 - 0
This enables ns to take the limit (L is continuous by assumption), since
the right hand side is the derivative of a function that is essentially the
composition of the second and first factors (I" ,1
0
L)9 I wiIl abbre-
viate this [Ll, giving us:
d
(SI2) Right-speed"'log(C, I) "df dt IL] I
If the derivative (with respect to time) of a function is unity, of course,
it follows that the function is of the form Al ' I + k - or rather, In our
case, At . i
l
+ k, as dictated by our type constraints - where k is a
constant of type tlt. This is exactly what we would expect: the constant
represents the error in the clock's setting - the difference between.the
actual and indicated times, The equation, predictably, says that If a
clock is running at the right speed the error will (instantaneously)
remain constant. Furthermore, since (S3) implies that
to say that if at time I (say, 12: 00) a clock designates o'clock property
r( (say, 3: 11), then at time 1+ /',.1 (12: 01, for a one minnte /',.t) it
should indicate the property of being /',.t later, i.e., T(+,< (3: 12). We
can do this as follows:
(S5) Right-speed( c, I, /',.t) "df [a" 'H'] [a" ,I + /',.t
which has the consequence, given the definition of 'II, that
(S6) ['(a"" /',.1)1 [a",]+/',.I
Properly, we should state something stronger: that a clock runs cor-
rectly throughout the interval from I to 1 + /',.1 if and only if it
advances at the right speed for the whole time (note that the following
is neutral as to whether this is a continuous or discrete interval - Le.,
as to whether 'rj is a discrete or continuous quantifier):
(S7) Right-speed(c, I, /',.1) "df MI' 10 < /',.1' < /',.1
Ia", +",,1 Ia,,,j +/',.t'
again directly yielding
(S8) "1M I0 < M < /',.t ['(a"" M)j Ia,,,! + M
These equations involve a property identity, but I defer any questions
on that issue to situation theory. Note also that in each version the two
instances of '+' afe of different types: the first takes a time and an
interval onto a time, the second an o'clock property and an interval
onto an o'clock property, No problem,
Given (S3) and (S7), the temporal analogues of soundness and
completeness can be proved: if a clock is correct at time t, and runs at
the right speed during the interval from I to 1', then it will be correct
during that interval, and conversely if it is correct throughout the
interval it must be running at the right speed. But it is more fun to do
this in the continuous case, so let's turn to that.
Very simply, we want to talk of an analog clock's running at the right
speed instantaneously, which means, intuitively, that we should differen-
tiate the temporal state function L - or, what is equivalent, take the
limit of Las /',.1 approaches 0, in the standard way:
limit ((I a,,,! + M) -la",1l limit (Ia,.,'H,J - [a,,1)
6..1 0 6../ 0 At
Since, as we've already said, differences between o'clock properties are
intervals, the left side of this reduces to limit" _ 0 (/',.1//',.1), which is
(SI4)
(SI5)
Correct(c, I) IL(c, 1)1(1)
Right-speed(c, I, I') "df "1/',.110 < /',.1 < (I' - I)
la",H,1 [a",] +M
implyinglhal "1/',.110 < /',.1 < (1'-1)
I'(a" " /',.1)1 Ia" ,I +M
implying Ihal "1M 1 0 < M < (I' - t)
[L(c, I IL(c, 1)1 +/',.1
24
BRIAN CANTWELL SMITH
THE SEMANTICS OF CLOCKS 25
(S16) t) ""'df IL]
dt
and in. their terms define what it is for a clock to be "working" properly
from tIme t to t + f..t:
(S17) Working(c, t, f..t) =df Correct(c, t) A t, M)
(S18) Workinganalog(c, t) Correct(c, t) 1\ t)
For .either versi0.n, the constraint can be shown to be satisfied (over
the or mstantaneously, depending) in exactly the following
condItIOn:
(519) [L(C,t)]=Jt. L/
Given the abbreviation adopted above, we can state this even more
simply:
(S20) iL] = At . T,
I. be the first to admit that (S20) is obvious at least
ill the sense ?nce stated, it is hard to imagine thinking
anything else. In English, It says that the state function and the
pretation function should be proportional inverses; given a clock that
(so to maps onto some sort of complex motion, the
mterpretatlOn function is merely that function that maps
that. motion back onto the ?'cIock properties of the linear progression
of tfile was With. So the putative clock of Figure 5, for
example, .wIth a pendulum and a 24 hour period, would
have .a pomter posItIon (a) proportional to sin(t), and an interpretation
functIOn analogously proportional to sin-I(0').10
5
6 7
Fig. 5. The million mile clock
Still, (S20) isn't trivial, for a reason that shows exactly why clocks
were hard to build. It says that working clocks map all times onto their
o'clock properties. The problem for dockmakers is that L isn't directly
computable, since, to repeat, neither times nor o'clock properties enter
into causally efficacious behaviour. What can be implemented is qJ, not
L, and qJ is essentially the temporal derivative of 2:.
In sum, we have determined the function of clockworks: to integrate
the derivative of time. When you set the hands on the clock's face, you
are supplying the integration constant.
8. MORALS AND CONCLUSIONS
What have we learned? Four things, other than some fun facts to tell
our friends.
The first has to do with the interaction among notions of
tion, realisation, and formality. Clocks' participation in their clubject
matter (being temporal, as a way of measuring time), which depends on
their physical realisation, might seem to violate the formality constraint
that is claimed to hold of computational systems more generally. In fact,
however, clocks' temporality doesn't relieve them of much of the
structure that characterises more traditional systems: separable qJ and
<P, the possibility of being wrong, etc. This similarity of clocks to
symbol manipulation systems arises from the fact that the particular
aspect of times that clocks represent the o'clock properties aren't
within immediate causal reach of a clockwork mechanism (or of much
else, for that matter). In (Smith, forthcoming) I argue that this is a
manifestation of a deep truth: the limitations of causal reach are the real
constraints on representational systems. Formality, as a notion, is
merely a cloudy and approximate projection of these limitations into a
particular construal of the symbolic realm.
The second moral has to do with the impact, for theoretical analysis,
of the relat\on between qJ and <P. The function qJ, realised in
work, is the engineers must implement; without an analysis of it,
effective clocks couldn't be designed. But theories of clocks must go
much further. Our characterisation of what it was for a clock to work
properly, for example, had to reach beyond the immediate or causally
accessible aspects of the underlying clockwork mechanism. Whatever
one might think about more complex cases, methodological solipsism
doesn't work in this particular instance.
Fig. 6. C5: Coordinated constraints on content and causal connection.
--....... causal relations
===):? content relations
27
ltM
d
THE SEMANTICS OF CLOCKS
'C b j'.
Both kinds, in general, will be complex - much more so than we have
seen in the case of clocks. Two aspects of content we haven't encoun-
tered, for example, are its "situational" dependence on surrounding
circumstances, as discussed for example in (Barwise, 1986b; Perry,
1986), and the three-way semantic interactions among language, mind,
and world that arise in cases of communication. Causal connections are
similarly complex, and can be broken down into three main groups:
1. Internal activity or behaviour: the relation between a system at some
time and the same system shortly thereafter. This is what we've
called \(f.
2. External connection: actions the system takes that affect the world,
and effects on the system of the world around it - the results, that
is, of senSOrs and effectors. (Clocks have none of this, but other
systems are clearly not so limited.)
3. Background dynamics: the progress or flow of the surrounding
situation. The passage of time would be counted as one instance, as
would one's conversant's behaviour, or the passing visual scene.
In the traditional case of pure mathematical inference, there is no
connection (action or sensation), and the background situation, as we
saw, is presumed to stay fixed. Barwise's construal of "formal inference"
(the "non-situated" reading), (Barwise, 1986b, p. 331) strengthens this
constraint by assuming that the content relation is also independent of
surrounding situation. The clock example gives us a different point in
the space: again no connection, an essentially unchanging (and relatively
situation-independent) content relation, but an evolving background
situation, mirrored in the internal activity or behaviour. Finally, seman-
tic theories of action, involving everything from intentionally eating
supper to making a promise, must deal with cases where the connection
aspect makes a contribution. They must therefore deal with cases where
the surrounding situation is affected not only by its own background
dynamics, but as a result of internal activity on the part of the
representational agent. But simpler systems will require an analysis of
external connection, as well: computerised (ABS) brakes on late model
cars, for example, are directly connected (even vulnerable) to the
content of their representations, in a way that seems to free them from
the need to have their representational states externally interpreted.
In the end, however, the similarity among these systems is far more
important than the variance. We can put it this way. Causal participation
e'1
... the unJolding ofthe world ...
C::=======:JI TIME CI====:>
BRIAN CANTWELL SMITH
Content:
Background:
26
Internal
activity:
External
connection:
Third, the similarity between the state transition functions of com-
puter science and the temporal derivatives of mechanics, both of which
focus not on time itself but on temporal change, suggest the possibility
of a more unified treatment of representational dynamics in general. So
far most of what we have to say deals with specific cases. So, for
example, in Section 2 we characterised inference as a particular species
of representational activity, having to do with changing content relations
to a fixed SUbject matter. Inference was contrasted with clock's main-
tenance of a fixed content relation to a changing subject matter.
Remembering what is perceived, to take quite a third sort of repre-
sentational behaviour, is a form of retaining a fixed relation to a fixed
subject matter in ways that make it immune to changes in the agent's
circumstances. It doesn't seem impossible that a common framework
could be uncovered.
Fourth, and finally, by occupying a place very different from that of
~ i t h e ~ Turing machines or traditional theorem provers, clocks help
lllummate the fundamental constraints governing computers and repre-
sentational systems in generaL As Figure 6 suggests, there are two basic
kinds of constraint causal relations and content relations - that a
representational system must coordinate as it moves through the world.
NOTES
I Clocks represent time for us, as it happens, not for themselves, but that will count, at
least here. I'm sympathetic to the distinction between original and derivative semantics
(in fact I'm interested in participation for just such reasons), but I am very much against
relativising representation to an observer at the outset, especially to a human observer
in the world is ultimately a two-edged sword. On the one hand, it is
absolutely enabling. Not only could a system not exist without it, but in
a certain sense it's total: everything the system is and does arises out of
its causally supported existence. There are no angels. On the other
hand, causal connection on its own - unless further structured - limits
a system's total participation in the world to those things within
immediate causal reach.
Representation, on this view, is a mechanism that honours the limits
of causal participation, but at the same time stands a system in a
content relation to aspects of the world beyond its causal reach. The
trick that the system must solve is to live within the limits (and exploit
the freedoms!) of the causal laws in just such a way as to preserve its
representational stance to what is distal. This much is in common
between an inference system and a clock.
29 THE SEMANTICS OF CLOCKS
! don't know (he /lumbers.
! knewyou didn't. Neither do 1.
Now! do.
Now! do too.
Mr.P.
Mr.S.
Mr. P.
Mr. S.
What are the numbers?
The earliest publication of this problem I am aware of is by H. Freudenthal in the
Dutch periodical Nieuw Archie! Voor Wiskunde, series 3, 17, 1969, p. 152 (a solution
by 1. Boersma appears in the same series, 18, 1970, pp. 102-106). It was subsequently
submitted by David J. Sprows to Mathematics Magazine 49(2), March 1976, p. 96
(solution in 50(5), Nov. 1977, p. 268). Perhaps the most widely read version appears in
Martin Gardner's 'Mathematical Garnes' column in Scientific American 241(6), Dec.
]979, pp. 22-30, with subsequent discussions and slight variations in 1980: 242(3),
March, p. 38; 242(5), May, pp. 24-28; and 242(6), June, p. 32.
(Winograd and Flores, 1986). To do that would be to abandon any hope of explain.ing
how the human mind might itself be representational, my ultimate goal. See (Smith,
forthcoming).
2 In computer science the claim that reference isn't computed is viewed suspiciously,
for a very interesting reason. To see it, consider why the claim is true. Suppose in a
room of 100 people some person A is the average height. Then suppose a new person
enters the room. Suddenly, and without any computation, a different person B will be
the average height. No work needs to be done to lift the property from A and settle it
on B; no energy expended, no symbols massaged. The new state just comes to be,
automatically, in virtue of the maze of conditions and constraints that hold. Reference, I
take it is something like that' conditions and constraints hold so that, when a word is
uttered or a thought some object becomes the referent. (Nor is it possible
to reply "Well, the room computed it"; on that recourse everything that happens would
be computed, which would make the word 'compute' vacuous.)
How could computer scientists object to this? FaT the following reason. Note that
the way that B becomes the person of average height is by participating in the situation
at hand: he enters the room. Participation, in other words, is what enables relationship
to exist Computers, on the other hand, are traditionally viewed in purely abstract
terms, and abstractions, whatever they are, presumably don't participate. The closest an
abstraction comes to the property of average height - or indeed to anything at all - is
by designating it. And so, because of this abstract conception of computers, one gets
lulled into thinking that everything has to come into being in this disconnected,
putatively "computational" way.
Needless to say, I don't believe the abstract conception of computers is right. More
strongly, I am arguing that participation - the opposite of abstraction - is exactly what
allows you to connect to the world in other ways than through explicit symbol
manipulation. See Section 8, and (Smith, forthcoming).
J For accurately measuring distances on roads, one attaches a "fifth wheel" to a car and
reads off the passing miles. Maybe, if time had been causally efficacious, we could have
built clocks the same way, running a wheel against time and reading off the passing
seconds.
4 There are two numbers between 1 and 100. Mr. P knows their product, and Mr. S
their sum. They have the following conversation:
BRIAN CANTWELL SMITH
28
ACKNOWLEDGEMENTS
This paper grew out of a bet made with Richard Weyhrauch during a
discussion late one night in a bar in Alghero, Sardinia about what was
involved in reading one's watch. Specifically, I promised to develop a
semantical analysis of the familiar behaviour cited in the third para-
graph of the paper: waiting a second to see whether a watch moves
before reading the time. This paper is part one of the answer; inter-
preting a clock will come later. My thanks to him and other members of
the Cost 13 Workshop on Reflection and Meta-Level Architectures,
especially including Jim des Rivieres and John Batali. Thanks also to
Jon Barwise, John Etchemendy, David Israel, John Perry, Susan Stucky,
and the other members of the situation theory and situation semantics
(STASS) group at CSLI, to Pat Hayes for discussions of measurement,
and to John Lamping for his help on celestial mechanics. The research
was supported by Xerox Corporation and the System Development
Foundation, through their mutual support of the Center for the Study
of Language and Information.
30 BRIAN CANTWELL SMITH THE SEMANTICS OF CLOCKS
31
5 The two other primary models, conceptually distinct from the formal symbol manipu-
lation idea, are the automata-theoretic notion of a digital or discrete system and the
related idea of a machine whose behaviour is equivalent to that of some Turing
machine. Although the formal symbol manipulation view seems to go virtually unchal-
lenged in cognitive science, the other two have much more currency in modern com-
puter science. See (Smith, forthcoming).
6 A more detached theoretic viewpoint should point out that o'clock properties Or are
in fact two-place relations between times and places (a time that is midnight in
London will be 7: aD p.m. in New York). More generally, whereas I assume throughout
that activity (tV) and interpretation (4)) are functions, they should properly be viewed
as more complex relations between agents and their embedding circumstances.
7 "third, n.... 5. The sixtieth part of a second of time or arc." - Webster's New
International Dictionary, Second Edition. New York: G. & C. Merriam, Co. 1934.
g Clock faces, and representations in general, don't need to register themselves, in
order to represent.
9 Strictly speaking this isn't quite accurate, since both [...J and should depend on c
and t: the function we are differentiating should really be AC, t. t)]. But being
strict would add only complexity, not insight.
10 This clock would be even harder to build than you might suppose. At first blush, it
might seem as if the equation of motion for a pendulum would imply that a very large
bob, swinging in an arc at the surface of the earth (an arc, say, 100 feet in length),
whose mass completely dominated the mass of a long string by which it was suspended
from a geosynchronous point 1150000 miles above the surface of the earth, would
have a period of 24 hours. Unfortunately, however, such a device would have a period
of slightly less than an hour and a half. Why this is so, and how to modify the design
appropriately are left as an exercise for the reader (hint: the result would be difficult to
read).
REFERENCES
Barwise, Jon, and Perry, John: 1983, Situations and Attitudes, Bradford Books,
Cambridge, Iv1A
Barwise, Jon: 19863, The Situation in Logic - III: Situations, Sets and the Axiom of
Foundation', in Alex Wilkie (ed.) Logic Colloquium 84, North Holland, Amsterdam.
AJso available as CSLI Technical Report from the Center for the Study
of Language and Information, Stanford University, 1985.
Barwise, Jon: 1986b, 'Information and Circumstance', Notre Dame Journal of Formal
Logic, 27(3) 324-338.
Brachman, Ronald J., and Levesque, Hector J. (eds.): 1985, Readings in Knowledge
Representation, Morgan Kaufmann, Los Altos, CA.
Fodor, Jerry: 1975, The Language of Thought, Thomas Y. Crowell Co., New York.
Paperback version, Harvard University Press Cambridge, MA, 1979.
Fodor, Jerry: 1980, 'Methodological Solipsism Considered as a Research Strategy in
Cognitive Psychology', The Behavioral and Brain Sciences, 3(1), pp. 63-73.
Reprinted in Fodor, Jerry, Representations, Bradford, Cambridge, MA, 1981.
Mumford, Lewis: 1934, Technics and Civilization, Harcourt, Brace & Co., New York.
Reprinted 1943.
Newell, Allen: 1980, 'Physical Symbol Systems', Cognitive Science 4, 135-183.
Perry, John: 1986, 'Circumstantial Attitudes and Benevolent Cognition', in J. Butterfield
(ed.), Language, Mind and Logic, pp. 123-134, Cambridge University Press
Cambridge.
Postman, Neil: 1985, Amusing Ourselves to Death: Public Discourse in the Age of Show
Business, Penguin Books, New York.
Smith, Brian c.: 1982, Reflection and Semantics in a Procedural Language, Technical
Report MITILCSITR-272, M.LT., Cambridge, MA, 495 pp. Prologue reprinted in
(Brachman and Levesque, 1985), pp. 31-39.
Smith, Brian c.: 1984, 'Reflection and Semantics in Lisp', Conference Record of 11th
Principles of Programming Languages Conference, pp. 23-35, Salt Lake City,
Utah. Also available as Xerox PARC Intelligent Systems Laboratory Technical
Report ISL-5, Palo Alto, California, 1984.
Smith, Brian c.: 1986, 'The Correspondence Continuum', appeared with the Proceed-
ings of the Sixth Canadian AI Conference, Montreal, Canada, May 21-23. Avail-
able as CSLl Technical Report CSLI-87-71 from the Center for the Study of
Language and Information, Stanford University, 1987. To appear in Artificial
Intelligence (forthcoming).
Smith, Brian c.: Is Computation Formal? MIT Press/A Bradford Book Cambridge,
MA. (Forthcoming.)
Winograd, Terry, and Flores, Fernando: 1986, Understanding Computers and Cogni-
tion: A New Foundation for Design, Ablex, Norwood, New Jersey.
Intelligent Systems Laboratory, and
Center for the Study ofLanguage and Information
Xerox Palo Alto Research Center
Palo Alto, CA 94304, u.s.A.
STUDIES IN COGNITIVE SYSTEMS
James H. Fetzer
University ofMinnesota, Duluth
Editor
ADVISORY EDITORIAL BOARD
Fred Dretske
University of Wisconsin, Madison
Ellery Eells
University of Wisconsin, Madison
Alick Elithorn
Royal Free Hospital, London
Jerry Fodor
Ciry University ofNew York
Alvin Goldman
University ofArizona
Jaakko Hintikka
Florida State University
Frank Keil
Cornell University
William Rapaport
State University ofNew York at Buffalo
Barry Richards
University ofEdinburgh
Stephen Stich
University a/California at San Diego
Lucia Vaina
Boston University
Terry Winograd
Stanford University
ASPECTS OF
ARTIFICIAL
INTELLIGENCE
Edited by
JAMES H ~ FETZER
Dept. ojPhilosophy and Humanities, University ojMinnesota,
Duluth, Us.A.
KLUWER ACADEMIC PUBLISHERS
DORDRECHT/BOSTON/LANCASTER/TOKYO