Confirmation that a voice is the
machines
own can be confirmed using two methods.
Constructing a Replacement for the Soul - Bourbon
Making a choice (a kind of
thinking) involves constructing a world, a frame, or a domain of relations organized according to a particular grammar (a set of rules and values).
Chimps make choices. Do they then make conceptual worlds? In an experiment a female chimp was trained to add and count. Part ofthe training procedure required the chimp to play a game where two piles of candies were counted out. Whichever pile the chimp first point to was given to another chimp sitting in front of her in a cage. The object of the game was to get the most candies. Whenever the candies were counted out the chimp invariably pointed at the larger pile, thus losing. When numbers written on cards were substituted for the candies, however, the chimp was always able to point to the smaller number, thus insuring that she would be rewarded with the larger number of treats.
One could speculate. When the chimp saw the food an instinctual grab/eat rule (or agent)determinedherchoice. Sheplayedthecandygameusingherinstinctualrules. She needed instead to nest the candy game within the her instinctual game, by replacing the instinctual mechanism for achieving her desire for the most food with the mechanism
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
584
defined by the game. By encoding the "amounts" o f candy within a numerical domain the relationship between the two piles can be determined without invoicing the grab/eat rule. Thisinformationisnolongerstoredasanimagewithinthegrab/eatrule. Thenumerical domain indicates the value of each number in relation to the other, not in relation to the chimp or her desire to eat. The failure of any abstract system to define this kind of personal relation requires that its values be accessed and assessed from another domain and according to another grammar. In this case, the chimp can then apply the rules o f the candy game to these "amounts" in order to determine the winning choice.
Any abstract system sets up a play between at least two different systems or domains. This recursive interaction creates a kind o f consciousness, because it creates the illusion o f free choice. This choice is not in any sense free, it is simply the mediation o f one domain by another. The chimp evaluates how to get food [an algorithm defined by CHOOSE THE MOST DESIRABLE and THE MOST DESIRABLE IS THE MOST FOOD] through the algorithm of the candy game [CHOOSE THE LEAST FIRST], The transcription of candies into number allows both of these algorithms to function together.
The three systems of relations in the previous experiment are all defined as systems o f value. These systems o f value are really systems o f meaning because they are sets o f relations. Semantic meaning within any discourse is primarily the articulation of a particularsetofrelations. Weassumethissetofrelationsisnotarbitrary. Thismeans they are constructed according to a particular logic. This logic is a logic of difference.
The statement "x means y" requires at some level, although not necessarily at the conceptual level Saussure suggests, that the notions "x" and "y" be understood as "not z",
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
585
"not w". Even to say this, however, requires that the notion "x" be crystallized in such a way that such a comparison is possible. How do we know that 'x' is not the same as ly'? We know this because they sound different, or because their shapes are different. Shape and sound serve as common denominators. Thus to speak o f difference requires that "x" function within a realm of common terms or qualities, which means a realm of pre- established possible relations and configurations. Thus what I first called a logic of difference is a difference mediated by a commonality. This common context is not given, but is constructed. (On a fundamental level, however, the initial contexts would be created primarily out o f the biological structures used to interpret our interactions with our environment).
Value is determined through differences expressed within a common context. A common context means in relation to a binary descriptive system: between two similar phonemes, between more and less, and so on. Values generated out ofthese metaphoric matrices are re-coded into more complex systems o f relations. We can easily see how this works in the chimp experiment. How can the chimp nest the rules o f the candy game within its instinctual grab/eat game? Encoding the amount o f candy in numbers created a domain of pure value organized around more and less. This information was embedded in the world o f the candy game where the output o f the algorithm CHOOSE THE LESS selects the lesser value. This value is eliminated and the larger value is transferred into the
domainofthechimpsgrab/eatrule. Ineachcasevalueisdeterminedbyhowtheeventis encoded within a system ofvalues. The situation facing the chimp is successively translated into different conceptual worlds, where the meaning o f that information is
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
586
determined by the values (in this case simple algorithms) organizing the relations within that world.
Thus, thinking is a process ofworld-making. A Dutch psychologist, de Groot, discovered that chess masters through a process o f implicit pruning do not see bad moves. They act within or through a grammar or a domain defined exclusively by good moves. They construct a conceptual world of good moves. This world defines what they see: it defines the moves the chess master recognizes as possible.
How does someone determine which of the possible moves or relations is correct, best, or relevant? This is too large a question to answer here. I can suggest, however, how we determine the relevant from the irrelevant within a frame or grammar in a general way. In an obvious way relevance is a function of understanding how something functions within a frame or domain. Frames must be extremely flexible to be able to handle the high degree of uncertainty and complexity of our experience. If we re-define determining relevance as making a choice we can make some progress. Choice arises through a kind of recursive nesting of domains and grammars. The interaction of these domains and grammars selects relationships from each that are integrated within a new domain and grammar, thus constructing an alternate world. This world defines what is possible. One assumes that at some level or in some world one is left with only a single possibility. This choice (really an algorithm) is transported to higher levels o f organization where it appears like we have made a conscious choice.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
587
14. 4 A wink in time means two
Before we can build a belief in the future, we must build awareness. We can build a basic time detector by using difference detectors, or what Minsky calls time-blinking:
Any agent that is sensitive to changes in time can also be used to detect differences. For whenever we expose such an agent, first to a situation A and then to a situation B, any output from that agent will signify some difference between A and B. (240)
While time-blinking partially removes the problem of comparing two inputs from different agents without requiring synchronous and identical frames (a trick it does largely by having a temporary memory), it does not explain how difference is translated into an ordered succession o f moments. One could, however, easily input the recognition o f a perceptual difference to a kind o f tape, as if marking a date or a second, which is then advanced. Suchaprocedurewouldrecordsuccessivedifferences,whichcouldfunction like a clock for human beings. It could never generate something like temporal experience for a machine. By what process is temporal order perceived and organized such that consciousness understands itself as the present? Do all the inputs that make up consciousness run through a single agent (an agent dangerously close to a homunculus) in rapidsuccessioninorderthatwecanperceivedifferenceandcallthatdifferencetime? If not, how is the information that our brain receives, processes and constructs, all at different rates, constructed into a now? The formation of defined moments is a process of constructing meaningful information. A moment, therefore, defines those "differences that makeadifference. " Timeisaninformationstructurewithinwhichwefindourselves.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
588
Thus our machine must generate representations o f its world in succession in which it finds itself. It has to turn itselfinside out.
14. 5 "Be! Verb umprincipiant through the trancitive spaces! "
My machine lives in a world o f simple objects. As a time machine it does not recognize objects as objects, but objecthood, as that which can be noticed as something, which it perceives in what it considers a partial instant o f time. It does not recognize the world or the environment, but only something that triggers its optical apparatus as a somethingtobenoticed. Itsopticalapparatushasalreadysimplifieditsphenomenalfield into this categorical perception, which we can call the set o f objects. This information is sent to the Optical (O) input agent which registers this information within a formally defined phenomenal field. The machine can rcognize separate example of objecthood within a range from 1 to 5. This phenomenal field is defined by a set o f switches (5) that representanobjectbybeingonandrepresentnoobjectbybeingoff. Thusthemachine optical apparatus can immediately register up to 5 objects. An initial input, N l, once it has set these switches (once the optical information is represented as an internal state), is transferred to an Optical analyzer agent where it again sets a similar set of switches. If the next input (N2) to the Optical input agent resets the switches, it is registered as a different moment and is thus sent on to O difference detector. The difference detector records that adifferencehasoccured. IfitsinformationisidenticaltoNl,nochangeofstatetakes place. Because our machine has only one sensory input at this point, the failure to register
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
589
difference is understood as non time; the first moment "continues" because the first state continues.
Once Nl has reset the difference detector, indicating a different moment registered as a change o f state, it enters a feedback loop. The detector's switches are then immediately reset by N2 (if there is an N2). N2 is similarly sent on a feedback loop but at twice the rate. Immediately after N2 resets the difference detector, N l resets it again.
This difference is converted into a value indicating the difference between these two states (degree o f change). The second cycle resets the detector to N2, before the next signal from O input is received.
Although it is not necessary to separate the process of detecting change from that of determining the degree of change, their functional difference allows us to split the presentintotwophenomenalelementsorinteractingmodels: anexperientialmodelandan evaluatory process. The difference detector, as the second element, serves as a kind of short-term memory within the phenomenal experience. It has, however, changed the nature o f the information perceived by the machine. Both kinds o f information determine the nature o f the present. It is the detector's separation from the initial model o f change
that allows the machine (to the degree that it is defined by its internal processes) to be aware of change. This means that the difference detector makes change, as defined by the 0 input agent, meaningful. The moment is actually not defined by any single state but by two states defined in relation to each other, embedded within the degree o f difference determined by the difference detector (so that Nl and N2 will be stored together as PI and P2). The difference detector, however, is inadequate for any short term memory. We
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
590
need a memory that encodes the consciousness o f the machine, and that requires that the information from both agents be stored. Both Nl and N2 are sent into a single short term memory unit (they are connected by a simple K-line). This unit is organized within short term memory (STM) according to two parameters: the ordered received (representing temporal succession) and degree of difference. . Short term memory (STM) stores both our experience o f change (suitably simplified) and a particular fact about these experiences. Thesearenotidenticalbitsofinformation. Theanalyzerdeterminesthe relevant facts for the time machine. Given the formally defined character of the phenomenal field, this information can be used to actually structure STM. (There are 10
possible values 1 through 5 and -1 through -5 defined in relation to the second state (N2): there is no O, o f course)
Optical Difference Machir (N2 Input State)
Difference Valu
(fact)
N1 = P1
Optical Inupt Agent (Change detector)
Optical Difference Detector
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Experiential Order
Short Term Memory
591
This is, ofcourse, not a model ofhuman perception. We have eliminated some ofthe criticalproblemsinconstructingamodelofoursenseoftime. Theworldandthe informationfromthatworldisimmeasurablysimplified. Inmymachine,theformal continuity between moments is built into the form in which the sensory input is represented (a defined set o f switches). With only one kind o f sensory information (number of objects), we do not have the problem of synchronizing and integrating the information from different senses (and their processing agents within the brain). The initial formation o f the moment is defined by the limit inherent within the speed by which the optical apparatus can convert its sensory input into a set o f objects. Changes that take placeata ratefasterthanthemachine'sphysicalabilitytorecognizeasetofobjectsare simply not registered.
14. 6 The logic of short-term prediction
If my machine can recognize patterns, such that the number of objects defined in one moment is always followed by a defined number in another moment, it can make short-term predictions. Once a pattern has been established, it can be recalled and run through a memory recall agent (another difference detector) at a rate faster than the information processed by the Optical agent. I will not spend much time constructing the mechanics of such a process. What is important in this problem is less the way in which this information is stored, than in the syntactical logic it necessarily generates. The logic here is simple: if x (a certain number of objects) is registered it will be followd by y
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
592
(another number ofobjects). Such predictions can, ofcourse, only happen ifthere is a pattern to be recognized.
Initially Short Term Memory (STM) is evaluated to determine a match with the initial optical input (Nl). If a match is found (through difference detectors attached to each memory slot), then the entire temporal unit, containing two states (PI, P2; remember PI) is transferred to the Memory Recall Agent. Nl and PI are processed as identical. The P2 state will not have an initial correlate state (it is run before N2). Because a prediction is based on understanding P2 as if it were N2, P2 must be understood as an N,
butnotastherealN2. Consequently,anysuchpredictionsrequireasymboliceconomy allowing for the conversion of Px into *Nx (where * indicates this input has an ontological claim equal in value but different in kind). The creation of an Nx state represents the creation of a temporal syntax. In order for the correlation of P2 with Nx to function as an expectation of N2, the machine must be able to act (or change its state as with Pavlov's dogs) on the basis of this correlation.
In Pavlov's bell experiment a similar temporal logic is manipulated. The equation o f food = salivate is re-wired to bell = salivate and is therefore based on an associated equivalence between food and bell. This equivalence, however, does not take place within a synchronic plane. It works because the sound of the bell recalls an established pattern recorded in memory. The P2 state of food following the PI state of the bell is understood asthesameasanNstate. Asymbolicinterchangetakesplacebetweenthepresentandthe future (patterned on a previously established relationship). A structural relationship has
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
593
been established between these two diachronic states, which allows the content o f one state (the bell) to stand for the content of the other (food).
Predictions based on short term memory are stimulus dependent. They disappear as soon as the environmental stimulus that suggested them disappears. They define only a temporaryandunstablefuture. Amachinewithonlythiskindofmemorycouldnot engage in any long term planning. Such plans require an ability to invoke the possibility of the future on command, as well as much more complex reasoning and creative modeling.
14. 7 The continuous future: hearing of and speaidng in the new world order
The Time Machine's world is now a rectangular plane, on which, at some unspecified but unreachable distance, a number o f objects flash in and out o f existence. This plane is divided into a series of parallel districts (wide strips). The number of objects ineachdistrictchangeindependentlyofthenumberofobjectsinotherdistricts. No machine, o f which now their are many, can see the objects (the number o f objects) in a district other than the one it is in. The machines can move to other districts. To prevent any machine from becoming a part ofthe object field ofanother machine, all machines will move along the same line and can move around each other if one machine is in the way of another. (The eyes move to the side o f the machines head. The machines bump into other machines and then move around them; I will not include this perception and response withinmymodelinghereforavarietyofreasons. Thinkofthisbumpingandmovmentas a function o f the machine autonomic nervous system)(see Figure A).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
594
Figure A
Object Field
OO
Object Field
Object Field
Districts
For our machine to survive in this world it will need, in addition to the optical system (0 input; O diffemce detector; memory, etc. ), a vocal generator (V) an auditory sytem (A) and more complex memory and recognition processing. A picture o f these systems, while confusing in its details, at least gives a sense ofthe increase in complexity that we need to make this simple world meaningful in even the smallest degree (Figure B):
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
achin
595
14. 8 Another blueprint of time Figure B
input Self
Pattern Recognizer
A STM
(no match)
A/O lvalue tQL
Difference DetectorN
STM (match)
Input
O Diffe rence Detec or
0 STI
lemo jecalL
A/O LTM
14. 9 Mental imperialism: to make the world the mind
The goal o f the machines in this world is to find and cohabitant with another machine within its district through the recognition of its vocalization of its perception of thenumberofobjectswithinthedistrictitfindsitself. Thismeansamachineistryingto recognize that another machine 'exists' within the first machine's 'time-world. ' A machine will recognize another machine as an external expression o f its own temporal perception o f its time-world. Its time-world becomes its awn time-world by being
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
0 LTM (Patterns)
596
externalized (recognized) as an external limit (the other machine) o f its own internal perceptions (in other,words it orients itself toward the world as if that world is both its world (defined by succession) and not its world (marked by its recognition o f another machine as another machine within its world).
A machine will now have a vocalizing agent (V) through which it will express the degree of change between two moments. This machine talks time. This vocalizing agent will receive this information from the O Difference Detector ater every five machine moments.
In order to hear these vocalizations, each machine will also have an auditory agent (A). Topreventconfusingthemachine'sownvoicewithanother's,thevocalizingagent will signal the auditory agent that it is speaking.
Confirmation that a voice is the machines own can be confirmed using two methods. Because the distance between the vocalizing agent and the auditory agent is always the same, the rate at which the A agent receives an input after it receives the signal from V will always be the same. In addition the signal from V will also initialize the auditory agent at the specific value to be vocalized and thus the auditory agent can function as a difference machine where the difference between input and initial state should be 0. This information is shunted into its own special memory.
14. 10 Time and the other
The auditory agent will have a different causal history, or in terms of my temporal
machines, a different time-scale than the Optical agent. Organic brains receive information from their sensory inputs at different rates depending on the distance from the brain to the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
597
sensory receptor, the distance between processing agents in the brain, and the different rates at which these receptors and agents work. The more complex the sensory model of the world, the more complex the temporal model generated from each sense. Such complexityrequiresmediation. Inmymachine,sensoryconflictarisesfromthedifferent rates and content of the information processed by the O and A Agents. The mediation between these agents is accomplished by constructing a common structure in which the differences in their content can be expressed independently o f each other.
The auditory agent in the machine will construct a short-term memory using a difference evaluator, similar to the O Difference Detector. Because the information received by A is already mediated by the optical agent, it cannot be used to directly model the number of objects in the external world. This auditory information describes the internal state of other machines. The Auditory Agent, therefore, will evaluate differences between its auditory inputs (between machines) only in order to develop a time scale for its auditory perception. (In a more advanced machine it would be possible to begin to determine which inputs originate from the same machine). The machine has two sensory inputs. Twoidenticalauditoryinputswillnotbemistakenforasinglemomentunlesssuch an identity occurs "simultaneously" at O and A. These two different forms o f sensory input require the machine to construct its own internal representation o f time in order to mediate the differences, conflicts, and confusions between the auditory and optical time- scales. Because the auditory input communicates the degree of change perceived by what we know is another machine, this information must be evaluated in relation to the degree
o f optical change perceived by the listening machine. Unlike the O Agent, the A Agent is
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
supported by another more fundamental agent that batches together two auditory states into a single auditory moment. The A Agent we are constructing interprets these momentsasmeaningfulpackagesofsound. Whymeaningful? Meaningisusedhereonly to indicate that these inputs do not refer to the same thing as a machine's initial optical input, but must be related to the information generated from its own Optical Difference Detector.
The difference values generated from its Auditory agent can have little sense to our machine. If we first imagine the machine's auditory sense without reference to its alreadyestablishedopticalsense,webegintounderstandtheconfusion. Itcanhear information communicated from beyond its own district. It has no idea how many machines it can hear nor can it tell the difference between one machine or another (all machine voices are identical and it senses are not refined enough to distinguish between thedistanceofvoiceexceptwhentheyareagoodnumberofdistrictsaway). This auditory information can only be used to define a series o f random differences marking the changes that they themselves cause in the auditory perception apparatus of our machine: an auditory time-scale or map.
If, however, these auditory inputs can be compared with the vocalizations of our machine, or even better with the values generated from the Optical Difference Detector, this auditory information no longer simply serves to define a time-scale in which it is the difference between inputs, and not their content, which carries information. . The machine will test all o f its O short term memory slots (with priority given to the most recent). Once it finds a match (again through a difference test), it will wait for 5 moments in order
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
599
to retest the match, which will have to match not only in value but also in the O memory slot it is found. This means that a test will take at least as long as 5 moments, slow but necessary if a district is as wide as the distance it takes sound to travel for 5 machine moments(itmightbefilledwithsomedensegasorevenwater). Thisretestingwillalso reduce the problem o f false matches (where a degree o f change value o f +2 could be obtained by a 5 moment followed by a 3 moment and a 3 moment followed by a 1 moment). It is, o f course, not even necessary that a real match, from our perspective, take place. Itisenoughthatitseemslikeamatchtothemachines(althoughanyfalsematch willdissolvefairlyquickly. Itisunlikelygiventhenumberofpossiblecombinationsinour 5 object system that the difference ratio between moments would remain the same in two
different districts for more than 3 moments). The machine edits out A inputs different from its own internal state as "noise" once a match has been discovered.
14. 11 Looking for Mr. Goodmachine
If a machine fails to find an auditory match with its own O Difference Detector
after 3 vocalizations ( in order to give machines which have not yet vocalized but are within the same district time to communicate with this other machine before it moves), it moves the distance possible in 3 moments, and then trumpets its degree o f change value. (The machine has eyes on the sides o f its head so that it can see, and thus function within theprimaltime-scalethatdeterminesitsmomentsanditsrateofvocalization). Ifit processes a match in transit, the machine will stop and retest the match. If it finds a match then it doesn't move. Both machines are in the same district and, therefore, happiness!
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
600
Once a positive match has been made three things will happen. First, a pattern detector will correlate a signal from the Optical Difference Detector sent every Optical moment with its matched auditory input. It will discover that the match happens every 5 moments. Second, the correlated auditory and optical information is stored in Auditory/Optical Long-Term Memory (A/O LTM). Third, a signal is sent from the
A/O mediating agent to V in order to re-initialize the timing for vocalizations to match the vocalizationsofthemachinethatiswithinthedistrict. (Iamignoringthecaseofmore than one machine in the district, although the districts are large enough to accommodate more than one machine). At the Vocalizing Agent a difference detector is installed in order to detect the change from the previous vocalizing timing. This information is then sent to A/O LTM. The A/O LTM, therefore, is organized as a series of bifurcated units representing an Auditory value and an equal Optical value within the domain marked by the change in vocalization timing.
14. 12 Meta-temporal identities: the modeling of others as syntax
The A/O Evaluator has constructed a meta-temporal identity within LTM, an
identity (not yet an entity) that organizes temporal information (two levels of abstraction above the initial perception of change), instead of representing particular temporal states. The A/O Evaluator has integrated the O agent and A agent, and in so doing has created a more abstract and mediated information structure. The information in the A/O LTM, because it is distinguished from the machine's own vocalizations, represents the external world. This is an external world, however, that includes representations of what the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
601
machine recognizes as internal states. By matching A input with its own phenomenal perception the machine models itself within itself by modeling another machine (although the machine understands this as simply a sensory input). As a model for humans, this describes consciousness as a means ofconstructing a model ofthe minds ofother humans in order to predict their actions and manipulate them.
It is not quite so simple, however, to ascribe this A input to another machine when itisimpossibleforonemachinetoidentifyanotherasthesourceofthisinput. The machine must first recognize its internal state, the information generated by A, V, and O, as its own internal state. In my machine self-realization emerges from the synchronization o f the machine's vocalization with the vocalization represented by the matched auditory input. This synchronization is a kind of internal state change, caused, however, by an external stimulus. Because the machine is able to separate out its own vocalization, it can distinguish between I and not-I in relation to this vocalization and its recognition as a vocalization generated from its optical system. By matching the auditory input with its O Evaluator input it constructs an informational object that represents the not-I as the I, and the I as a not-I. This information, the meta-temporal identity, can have not meaningful significance, however, before it is used in some control function.
The short-term memories generated by those auditory inputs that do not match the machine's own degree o f change value have little significance, until the matched inputs cease. Once this happens, there is no longer an internal representation o f the temporal history o f the A agent. It is critical for our objective o f constructing a temporal syntax o f thefuturethatthisnothappen. Whilesuchacontinuoushistorydoesnottakeplaceinour
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
602
brains, an analogous history is written in our external representations of time. The earliest o f these are the notations and sticks and bones, identified by Marshack, marking the lunar cycle. There are curious parallels between these 35,000 year old markings and the internal symbolic economy ofthe machine. These markings are a record ofexperimental observations. One cannot imagine a use for them unless their exists a cultural logic, supported by long-term prediction and planning, into which the notations fit. Once could imagine something like, "Every other lunar cycle, at the first quarter moon and the final new moon, we will do X. " These markings become signals both of time and correlated human activity. The need to determine events through this kind of record keeping suggests a complicated social organization that requires some objective form of organizing its activities, and which supports a wide variety of cultural activities (where questions of correct timing might cause conflict without an "objective" time map).
The markings, however, are non-arithmetic. This means the lunar cycle is not defined by a number o f days (nor are any o f the phases), but by grouping markings in a general pattern. Thus the number of days may vary between the full and half moon in any month (depending on the observations of the scribe), but their basic temporal relationship remainsthesame. ThiskindofgroupingissimilartothecreationofdomainswithinA/O LTM. Thus the regularity of the moon changes is analogous to the less regular, but still periodic, vocal timing changes. In the human mind the creation of organizing syntactical domains is largely a social phenomena, and thus needs to be externalized. Because my machines only form very simple social groups, these structures cannot be supported by culture.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
The ability to perform this kind o f representational abstraction requires the same kind o f ability to derive form from particular content that I am trying to build into the Time Machine. For early humans, a single mark stands for the idea, the concept of a unit relative to the lunar cylce, and not its particular content (except to the degree that an event or action is meaningful in relation to the logic of the notational series as a whole).
Similarly the grouping o f these units, because they lose the particularity o f their content, can be used to "mean" a particular lunar period. These kinds of representational systems are symbolic languages, where markings, like words, are generalized in their meanings (word less so than the marks for days) so that they can be exchanged and organized within ameta-structurenotoverwhelmedbydetail. Symbolizationisaprocessofsimplification that allows an event to be reduced to a particular representation that can then be used to form other meaningful statements. All units within these lunar calendars must be relatively identical, in the way that words define events in relation to pre-established categories of meaning.
For our machine a continuous representation of time will be used, like it was in early human bone markings, to define a continuous temporal structure in which all machine activity can take place. If the A input ceases to match the O Evaluator values, the A unmatched STM is shunted and nested within the A/O LTM. By nesting I mean the informational structure generated in the A unmatched STM is translated into the language, the form, of the A/O LTM. The A unmatched STM provides a temporal link betweentheendingofonemachinematchtothebeginningofanother. Theinformationit stores in A/O LTM must fit within the parameters previously set up to accommodate the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
604
matched information. The unmatched information (e. g. a difference o f -3 between the machines internal Optical difference and its auditory input) has no meaning in relation to thematcheddifferencevaluesof0,unlesssimplytoindicatematchorunmatched. This kind o f simplification, or abstraction, is necessary if the unmatched STM is to function in relationtoA/OLTM. Theeffectofsuchanesting,therefore,istoignorethecontentof the Auditory (A) unmatched STM, and simply use this information to mark the primary temporal units constructed in the Optical Agent. The structure of A unmatched STM is used to augment and continue the structure ofAJO LTM. . Once this is accomplished, the A/O LTM can be understood as a continuous, simplified representation, through external surrogates,oftheinternal temporalstructureofthemachine.
Because the A input and O input values are correlated with each other, their content is in effect erased. Because each memory slot in A/O LTM encodes so little information, it becomes in effect a series o f marks which are themselves marked by the change in vocal timing. Sensory input is organized within a structure generated in relation to an internal state change. The A/O LTM models both external and internal time within asinglestructure. Evenmoresignificantlythelevelofabstractionoftheinformationused by the A/O Evaluator has allowed the creation of a model of the structure of time divorced from its content (defined in relation to a unit o f difference, not an instant in time). Our problem now is to figure out how to use this structure.
14. 13 Negative entropy
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
605
The A/O LTM gives our machine the ability to make a long range prediction. We must assume that at least one machine can outlive a number of its matches. If so, the only pattern a memory analyzer o f A/O LTM could identify would be that every vocal synchronization change will be followed by another change after a period o f time. Because the content of the A/O LTM is so poor, if a change in vocal synchrony occurs and recalls the memory of past changes (that is, the pattern of such changes), it will recall the temporal structure embodied in the LTM. Thus, whenever a new match occurs a long-range prediction would be possible: this vocal change will be followed by a series of differences (temporal units) which will themselves end in a new vocal timing change.
The structure supporting this prediction (the moments represented by the O Difference
Detector) will be in place until this expectation (the next vocal timing change) has been met, at which point it will begin again. Attached to the A/O Evaluator will be a memory recall agent that will work in parallel with the Evaluator, but will also organize both A and O input within the temporal structure imported from the A/O LTM. We have a future.
14. 14 Re-building ourselves in the other
This leaves us with two problems however. How can we integrate the other
perceptual, evaluatory, and memory agents into this structure? And what is the significance o f the two states (match and unmatched) within the temporal series in the A/O LTM?
The notion of the future, dependent as it is on a society of machines (identical in everything but life-span), requires a society of different sensory agents. Is it necessary,
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
606
however, to integrate the O LTM, with its patterns and short-term predictions, within the long-term future of A/O agent? If our machine receives optical information that triggers a short-term prediction, we might want this to signal the A/O Evaluator. But the A/O agent could not understand this kind of pattern (between numbers of objects). Such a pattern would not be reflected in its degree of change information. The only patterns it could recognize, other than the change in vocal synchronization, is the shift from its shunting of matched to unmatched information to LTM and a correlation between a matched auditory inputandeveryfivesignalsfromtheOpticalDifferenceDetector. IftheAagentcould activate a pattern recognition agent to evaluate the unmatched A STM that would be sensitive to more abstract patterns o f difference (every degree o f change 2 is followed by a degree o f change 4), and if we pretend that these patterns are present in some districts, then our two memory structures could be integrated easily by running them both through
the same memory recall agent. The existence o f such patterns in the machine world would indeed be convenient, but then so is the bipolar structure of water.
We have three sub-agents, constituting three internal time scales, not fully exploited in modeling time: the (1)V Difference Detector indicating a change in vocal timing, the (2) A/O Evaluator's awareness that every match happens after 5 optical moments, and the (3) A/O Evaluator's (one o f its internal agents) ability to detect the shift from matched to unmatched outputs to memory. (1) and (3) constitute long-term regularities, while (2) is a short-term prediction that can act as the structural content organizing the temporal structure enabling long-term predictions.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
607
The machine's change in (1) vocal timing finds its analogue in the death o f the external matched signal: the first registers as an internal change; the second as an external change. We still do not know if our machine understands this as the death o f its machine mate. An existential understanding ofthe loss ofthe matching voice would require (ifwe ignore the lack o f a mammalian brain, or some might say any brain) an analogic equivalence between the A/O Evaluator shift to unmatched memory output and the death o f itself. It seems unlikely that a machine without a more developed symbolic life could
understand this loss. The predictive power of this kind of system, however, remains stimulus dependent. When the stimulus vanishes, the future disappears.
14. 15 Thepursuitofdeath
If we could make a higher level analyzing agent to determine and chart the larger patterns of change in A/O LTM [(3): matched to unmatched], we might be able to stimulate the machine into a weak awareness of the subjectivity of another machine encoded in the A/O LTMs meta-temporal identity. The A/O Evaluator sends a signal to
this agent when the shift from matched to unmatched occurs, and this signal causes this Pattern Recognizer to recall previous shifts and their defining vocal timing domains. The Pattern Recognizer could construct an information structure that charts the internal state changes against the representation o f catastrophic loss o f the external matching voice.
The information such a chart could encode is rather simple, and the prediction that follows even simpler: every vocal timing change will be followed not only by another, but also by the loss ofthe not-I representation. Because the representation ofthe not-I is used to
with permission of the copyright owner. Further reproduction prohibited without permission.
determine the actions o f the machine (it stays and observers), its loss results in a change in action (the machine now moves). Before moving can be a form o f mourning, it must also representalossofbeingforthemachine. Ifweremovethemachine's eyesonthesideof its head, it will lose its primary temporal portal. It will still have its auditory input (and STM) by which it can chart a different temporal order, but the loss of this primary information will be recorded with the shift from matched to unmatched in the A/O LTM and will be noted by the A/O LTM Evaluator. This sensory loss, and the time/ memory scale it constructs, can stand for, can symbolize, the loss o f the external voice and the change in vocal timing. The prediction that a loss o f the external voice will necessarily follow a change in the machine's own vocal timing will also be tied to another internal
effect: thecommandtomoveanditssubsequenttemporarylossofsightandthetime- scale that sight generates. This causal structure (Cause [external loss] and Effect [internal loss]) mirrors the vocal timing change and is nested within the basic structure generated from that timing change. Our sense o f the future is growing more complex.
14. 16 Learning from the future
The bifurcation o f the memory and predicative domain defined and signaled by the
vocal timing change sets up a meta-time scale at which level internal symbolic exchanges begin to make some kind o f sense. The O agent wants to be able to predict when it will fail, and when it's short-term predictions will be irrelevant, while the A/O wants to be able to predict when it will have to shift its memory output from matched to unmatched. Our machine could not discover such detailed information, nor could these agents as they now
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
609
stand actually want anything. They would first have to need this information as a means to act. But what is important here is the conceptual problem this presents. For the Optical Agent the loss of the external voice could stand for its loss o f sight, and because it happens prior to its effect, it could serve as a short-term prediction (generated out of a long-term prediction structure). This represents an internal symbolic communication, where one loss can stand for another. The signal itselfwould go from the A/O Evaluator to the O Agent. This signal is not, however, caused by an algorithm: if shift to unmatched output, send a signal to O that loss of sight is imminent. The meaning ofthe shifttounmatchedoutputisdeterminedbytheA/OLTMPattern. Itconcludesfrom previous patterns that the shift means a loss o f sight. This is a conclusion based on an internal analysis (abstraction) from a lower level abstraction (A/O LTM) which is in turn
generated from lower level sensory modeling. This is not the most elegant of control structures (but neither is much of our learning compared to "lower" animals: human weavingcomparedwithspiderweaving). Ithastheadvantage,however,ofmakingall predictions a function ofmemory (they are learned), and not of pre-programming.
14. 17 Inside a Chinese box: when structure becomes content within larger structures
A matched memory state (any input of value 0) within the A/O LTM, re- interpreted in terms o f the time scale generated from the Optical Agent, stands for 5 Optical moments. If a match has been found, then every five O moments will result in a matched input to A/O LTM. The Pattern Recognizer attached to A/O Evaluator
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
610
functions as a symbolic catalyst linking the two time scales.
thinking) involves constructing a world, a frame, or a domain of relations organized according to a particular grammar (a set of rules and values).
Chimps make choices. Do they then make conceptual worlds? In an experiment a female chimp was trained to add and count. Part ofthe training procedure required the chimp to play a game where two piles of candies were counted out. Whichever pile the chimp first point to was given to another chimp sitting in front of her in a cage. The object of the game was to get the most candies. Whenever the candies were counted out the chimp invariably pointed at the larger pile, thus losing. When numbers written on cards were substituted for the candies, however, the chimp was always able to point to the smaller number, thus insuring that she would be rewarded with the larger number of treats.
One could speculate. When the chimp saw the food an instinctual grab/eat rule (or agent)determinedherchoice. Sheplayedthecandygameusingherinstinctualrules. She needed instead to nest the candy game within the her instinctual game, by replacing the instinctual mechanism for achieving her desire for the most food with the mechanism
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
584
defined by the game. By encoding the "amounts" o f candy within a numerical domain the relationship between the two piles can be determined without invoicing the grab/eat rule. Thisinformationisnolongerstoredasanimagewithinthegrab/eatrule. Thenumerical domain indicates the value of each number in relation to the other, not in relation to the chimp or her desire to eat. The failure of any abstract system to define this kind of personal relation requires that its values be accessed and assessed from another domain and according to another grammar. In this case, the chimp can then apply the rules o f the candy game to these "amounts" in order to determine the winning choice.
Any abstract system sets up a play between at least two different systems or domains. This recursive interaction creates a kind o f consciousness, because it creates the illusion o f free choice. This choice is not in any sense free, it is simply the mediation o f one domain by another. The chimp evaluates how to get food [an algorithm defined by CHOOSE THE MOST DESIRABLE and THE MOST DESIRABLE IS THE MOST FOOD] through the algorithm of the candy game [CHOOSE THE LEAST FIRST], The transcription of candies into number allows both of these algorithms to function together.
The three systems of relations in the previous experiment are all defined as systems o f value. These systems o f value are really systems o f meaning because they are sets o f relations. Semantic meaning within any discourse is primarily the articulation of a particularsetofrelations. Weassumethissetofrelationsisnotarbitrary. Thismeans they are constructed according to a particular logic. This logic is a logic of difference.
The statement "x means y" requires at some level, although not necessarily at the conceptual level Saussure suggests, that the notions "x" and "y" be understood as "not z",
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
585
"not w". Even to say this, however, requires that the notion "x" be crystallized in such a way that such a comparison is possible. How do we know that 'x' is not the same as ly'? We know this because they sound different, or because their shapes are different. Shape and sound serve as common denominators. Thus to speak o f difference requires that "x" function within a realm of common terms or qualities, which means a realm of pre- established possible relations and configurations. Thus what I first called a logic of difference is a difference mediated by a commonality. This common context is not given, but is constructed. (On a fundamental level, however, the initial contexts would be created primarily out o f the biological structures used to interpret our interactions with our environment).
Value is determined through differences expressed within a common context. A common context means in relation to a binary descriptive system: between two similar phonemes, between more and less, and so on. Values generated out ofthese metaphoric matrices are re-coded into more complex systems o f relations. We can easily see how this works in the chimp experiment. How can the chimp nest the rules o f the candy game within its instinctual grab/eat game? Encoding the amount o f candy in numbers created a domain of pure value organized around more and less. This information was embedded in the world o f the candy game where the output o f the algorithm CHOOSE THE LESS selects the lesser value. This value is eliminated and the larger value is transferred into the
domainofthechimpsgrab/eatrule. Ineachcasevalueisdeterminedbyhowtheeventis encoded within a system ofvalues. The situation facing the chimp is successively translated into different conceptual worlds, where the meaning o f that information is
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
586
determined by the values (in this case simple algorithms) organizing the relations within that world.
Thus, thinking is a process ofworld-making. A Dutch psychologist, de Groot, discovered that chess masters through a process o f implicit pruning do not see bad moves. They act within or through a grammar or a domain defined exclusively by good moves. They construct a conceptual world of good moves. This world defines what they see: it defines the moves the chess master recognizes as possible.
How does someone determine which of the possible moves or relations is correct, best, or relevant? This is too large a question to answer here. I can suggest, however, how we determine the relevant from the irrelevant within a frame or grammar in a general way. In an obvious way relevance is a function of understanding how something functions within a frame or domain. Frames must be extremely flexible to be able to handle the high degree of uncertainty and complexity of our experience. If we re-define determining relevance as making a choice we can make some progress. Choice arises through a kind of recursive nesting of domains and grammars. The interaction of these domains and grammars selects relationships from each that are integrated within a new domain and grammar, thus constructing an alternate world. This world defines what is possible. One assumes that at some level or in some world one is left with only a single possibility. This choice (really an algorithm) is transported to higher levels o f organization where it appears like we have made a conscious choice.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
587
14. 4 A wink in time means two
Before we can build a belief in the future, we must build awareness. We can build a basic time detector by using difference detectors, or what Minsky calls time-blinking:
Any agent that is sensitive to changes in time can also be used to detect differences. For whenever we expose such an agent, first to a situation A and then to a situation B, any output from that agent will signify some difference between A and B. (240)
While time-blinking partially removes the problem of comparing two inputs from different agents without requiring synchronous and identical frames (a trick it does largely by having a temporary memory), it does not explain how difference is translated into an ordered succession o f moments. One could, however, easily input the recognition o f a perceptual difference to a kind o f tape, as if marking a date or a second, which is then advanced. Suchaprocedurewouldrecordsuccessivedifferences,whichcouldfunction like a clock for human beings. It could never generate something like temporal experience for a machine. By what process is temporal order perceived and organized such that consciousness understands itself as the present? Do all the inputs that make up consciousness run through a single agent (an agent dangerously close to a homunculus) in rapidsuccessioninorderthatwecanperceivedifferenceandcallthatdifferencetime? If not, how is the information that our brain receives, processes and constructs, all at different rates, constructed into a now? The formation of defined moments is a process of constructing meaningful information. A moment, therefore, defines those "differences that makeadifference. " Timeisaninformationstructurewithinwhichwefindourselves.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
588
Thus our machine must generate representations o f its world in succession in which it finds itself. It has to turn itselfinside out.
14. 5 "Be! Verb umprincipiant through the trancitive spaces! "
My machine lives in a world o f simple objects. As a time machine it does not recognize objects as objects, but objecthood, as that which can be noticed as something, which it perceives in what it considers a partial instant o f time. It does not recognize the world or the environment, but only something that triggers its optical apparatus as a somethingtobenoticed. Itsopticalapparatushasalreadysimplifieditsphenomenalfield into this categorical perception, which we can call the set o f objects. This information is sent to the Optical (O) input agent which registers this information within a formally defined phenomenal field. The machine can rcognize separate example of objecthood within a range from 1 to 5. This phenomenal field is defined by a set o f switches (5) that representanobjectbybeingonandrepresentnoobjectbybeingoff. Thusthemachine optical apparatus can immediately register up to 5 objects. An initial input, N l, once it has set these switches (once the optical information is represented as an internal state), is transferred to an Optical analyzer agent where it again sets a similar set of switches. If the next input (N2) to the Optical input agent resets the switches, it is registered as a different moment and is thus sent on to O difference detector. The difference detector records that adifferencehasoccured. IfitsinformationisidenticaltoNl,nochangeofstatetakes place. Because our machine has only one sensory input at this point, the failure to register
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
589
difference is understood as non time; the first moment "continues" because the first state continues.
Once Nl has reset the difference detector, indicating a different moment registered as a change o f state, it enters a feedback loop. The detector's switches are then immediately reset by N2 (if there is an N2). N2 is similarly sent on a feedback loop but at twice the rate. Immediately after N2 resets the difference detector, N l resets it again.
This difference is converted into a value indicating the difference between these two states (degree o f change). The second cycle resets the detector to N2, before the next signal from O input is received.
Although it is not necessary to separate the process of detecting change from that of determining the degree of change, their functional difference allows us to split the presentintotwophenomenalelementsorinteractingmodels: anexperientialmodelandan evaluatory process. The difference detector, as the second element, serves as a kind of short-term memory within the phenomenal experience. It has, however, changed the nature o f the information perceived by the machine. Both kinds o f information determine the nature o f the present. It is the detector's separation from the initial model o f change
that allows the machine (to the degree that it is defined by its internal processes) to be aware of change. This means that the difference detector makes change, as defined by the 0 input agent, meaningful. The moment is actually not defined by any single state but by two states defined in relation to each other, embedded within the degree o f difference determined by the difference detector (so that Nl and N2 will be stored together as PI and P2). The difference detector, however, is inadequate for any short term memory. We
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
590
need a memory that encodes the consciousness o f the machine, and that requires that the information from both agents be stored. Both Nl and N2 are sent into a single short term memory unit (they are connected by a simple K-line). This unit is organized within short term memory (STM) according to two parameters: the ordered received (representing temporal succession) and degree of difference. . Short term memory (STM) stores both our experience o f change (suitably simplified) and a particular fact about these experiences. Thesearenotidenticalbitsofinformation. Theanalyzerdeterminesthe relevant facts for the time machine. Given the formally defined character of the phenomenal field, this information can be used to actually structure STM. (There are 10
possible values 1 through 5 and -1 through -5 defined in relation to the second state (N2): there is no O, o f course)
Optical Difference Machir (N2 Input State)
Difference Valu
(fact)
N1 = P1
Optical Inupt Agent (Change detector)
Optical Difference Detector
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Experiential Order
Short Term Memory
591
This is, ofcourse, not a model ofhuman perception. We have eliminated some ofthe criticalproblemsinconstructingamodelofoursenseoftime. Theworldandthe informationfromthatworldisimmeasurablysimplified. Inmymachine,theformal continuity between moments is built into the form in which the sensory input is represented (a defined set o f switches). With only one kind o f sensory information (number of objects), we do not have the problem of synchronizing and integrating the information from different senses (and their processing agents within the brain). The initial formation o f the moment is defined by the limit inherent within the speed by which the optical apparatus can convert its sensory input into a set o f objects. Changes that take placeata ratefasterthanthemachine'sphysicalabilitytorecognizeasetofobjectsare simply not registered.
14. 6 The logic of short-term prediction
If my machine can recognize patterns, such that the number of objects defined in one moment is always followed by a defined number in another moment, it can make short-term predictions. Once a pattern has been established, it can be recalled and run through a memory recall agent (another difference detector) at a rate faster than the information processed by the Optical agent. I will not spend much time constructing the mechanics of such a process. What is important in this problem is less the way in which this information is stored, than in the syntactical logic it necessarily generates. The logic here is simple: if x (a certain number of objects) is registered it will be followd by y
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
592
(another number ofobjects). Such predictions can, ofcourse, only happen ifthere is a pattern to be recognized.
Initially Short Term Memory (STM) is evaluated to determine a match with the initial optical input (Nl). If a match is found (through difference detectors attached to each memory slot), then the entire temporal unit, containing two states (PI, P2; remember PI) is transferred to the Memory Recall Agent. Nl and PI are processed as identical. The P2 state will not have an initial correlate state (it is run before N2). Because a prediction is based on understanding P2 as if it were N2, P2 must be understood as an N,
butnotastherealN2. Consequently,anysuchpredictionsrequireasymboliceconomy allowing for the conversion of Px into *Nx (where * indicates this input has an ontological claim equal in value but different in kind). The creation of an Nx state represents the creation of a temporal syntax. In order for the correlation of P2 with Nx to function as an expectation of N2, the machine must be able to act (or change its state as with Pavlov's dogs) on the basis of this correlation.
In Pavlov's bell experiment a similar temporal logic is manipulated. The equation o f food = salivate is re-wired to bell = salivate and is therefore based on an associated equivalence between food and bell. This equivalence, however, does not take place within a synchronic plane. It works because the sound of the bell recalls an established pattern recorded in memory. The P2 state of food following the PI state of the bell is understood asthesameasanNstate. Asymbolicinterchangetakesplacebetweenthepresentandthe future (patterned on a previously established relationship). A structural relationship has
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
593
been established between these two diachronic states, which allows the content o f one state (the bell) to stand for the content of the other (food).
Predictions based on short term memory are stimulus dependent. They disappear as soon as the environmental stimulus that suggested them disappears. They define only a temporaryandunstablefuture. Amachinewithonlythiskindofmemorycouldnot engage in any long term planning. Such plans require an ability to invoke the possibility of the future on command, as well as much more complex reasoning and creative modeling.
14. 7 The continuous future: hearing of and speaidng in the new world order
The Time Machine's world is now a rectangular plane, on which, at some unspecified but unreachable distance, a number o f objects flash in and out o f existence. This plane is divided into a series of parallel districts (wide strips). The number of objects ineachdistrictchangeindependentlyofthenumberofobjectsinotherdistricts. No machine, o f which now their are many, can see the objects (the number o f objects) in a district other than the one it is in. The machines can move to other districts. To prevent any machine from becoming a part ofthe object field ofanother machine, all machines will move along the same line and can move around each other if one machine is in the way of another. (The eyes move to the side o f the machines head. The machines bump into other machines and then move around them; I will not include this perception and response withinmymodelinghereforavarietyofreasons. Thinkofthisbumpingandmovmentas a function o f the machine autonomic nervous system)(see Figure A).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
594
Figure A
Object Field
OO
Object Field
Object Field
Districts
For our machine to survive in this world it will need, in addition to the optical system (0 input; O diffemce detector; memory, etc. ), a vocal generator (V) an auditory sytem (A) and more complex memory and recognition processing. A picture o f these systems, while confusing in its details, at least gives a sense ofthe increase in complexity that we need to make this simple world meaningful in even the smallest degree (Figure B):
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
achin
595
14. 8 Another blueprint of time Figure B
input Self
Pattern Recognizer
A STM
(no match)
A/O lvalue tQL
Difference DetectorN
STM (match)
Input
O Diffe rence Detec or
0 STI
lemo jecalL
A/O LTM
14. 9 Mental imperialism: to make the world the mind
The goal o f the machines in this world is to find and cohabitant with another machine within its district through the recognition of its vocalization of its perception of thenumberofobjectswithinthedistrictitfindsitself. Thismeansamachineistryingto recognize that another machine 'exists' within the first machine's 'time-world. ' A machine will recognize another machine as an external expression o f its own temporal perception o f its time-world. Its time-world becomes its awn time-world by being
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
0 LTM (Patterns)
596
externalized (recognized) as an external limit (the other machine) o f its own internal perceptions (in other,words it orients itself toward the world as if that world is both its world (defined by succession) and not its world (marked by its recognition o f another machine as another machine within its world).
A machine will now have a vocalizing agent (V) through which it will express the degree of change between two moments. This machine talks time. This vocalizing agent will receive this information from the O Difference Detector ater every five machine moments.
In order to hear these vocalizations, each machine will also have an auditory agent (A). Topreventconfusingthemachine'sownvoicewithanother's,thevocalizingagent will signal the auditory agent that it is speaking.
Confirmation that a voice is the machines own can be confirmed using two methods. Because the distance between the vocalizing agent and the auditory agent is always the same, the rate at which the A agent receives an input after it receives the signal from V will always be the same. In addition the signal from V will also initialize the auditory agent at the specific value to be vocalized and thus the auditory agent can function as a difference machine where the difference between input and initial state should be 0. This information is shunted into its own special memory.
14. 10 Time and the other
The auditory agent will have a different causal history, or in terms of my temporal
machines, a different time-scale than the Optical agent. Organic brains receive information from their sensory inputs at different rates depending on the distance from the brain to the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
597
sensory receptor, the distance between processing agents in the brain, and the different rates at which these receptors and agents work. The more complex the sensory model of the world, the more complex the temporal model generated from each sense. Such complexityrequiresmediation. Inmymachine,sensoryconflictarisesfromthedifferent rates and content of the information processed by the O and A Agents. The mediation between these agents is accomplished by constructing a common structure in which the differences in their content can be expressed independently o f each other.
The auditory agent in the machine will construct a short-term memory using a difference evaluator, similar to the O Difference Detector. Because the information received by A is already mediated by the optical agent, it cannot be used to directly model the number of objects in the external world. This auditory information describes the internal state of other machines. The Auditory Agent, therefore, will evaluate differences between its auditory inputs (between machines) only in order to develop a time scale for its auditory perception. (In a more advanced machine it would be possible to begin to determine which inputs originate from the same machine). The machine has two sensory inputs. Twoidenticalauditoryinputswillnotbemistakenforasinglemomentunlesssuch an identity occurs "simultaneously" at O and A. These two different forms o f sensory input require the machine to construct its own internal representation o f time in order to mediate the differences, conflicts, and confusions between the auditory and optical time- scales. Because the auditory input communicates the degree of change perceived by what we know is another machine, this information must be evaluated in relation to the degree
o f optical change perceived by the listening machine. Unlike the O Agent, the A Agent is
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
supported by another more fundamental agent that batches together two auditory states into a single auditory moment. The A Agent we are constructing interprets these momentsasmeaningfulpackagesofsound. Whymeaningful? Meaningisusedhereonly to indicate that these inputs do not refer to the same thing as a machine's initial optical input, but must be related to the information generated from its own Optical Difference Detector.
The difference values generated from its Auditory agent can have little sense to our machine. If we first imagine the machine's auditory sense without reference to its alreadyestablishedopticalsense,webegintounderstandtheconfusion. Itcanhear information communicated from beyond its own district. It has no idea how many machines it can hear nor can it tell the difference between one machine or another (all machine voices are identical and it senses are not refined enough to distinguish between thedistanceofvoiceexceptwhentheyareagoodnumberofdistrictsaway). This auditory information can only be used to define a series o f random differences marking the changes that they themselves cause in the auditory perception apparatus of our machine: an auditory time-scale or map.
If, however, these auditory inputs can be compared with the vocalizations of our machine, or even better with the values generated from the Optical Difference Detector, this auditory information no longer simply serves to define a time-scale in which it is the difference between inputs, and not their content, which carries information. . The machine will test all o f its O short term memory slots (with priority given to the most recent). Once it finds a match (again through a difference test), it will wait for 5 moments in order
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
599
to retest the match, which will have to match not only in value but also in the O memory slot it is found. This means that a test will take at least as long as 5 moments, slow but necessary if a district is as wide as the distance it takes sound to travel for 5 machine moments(itmightbefilledwithsomedensegasorevenwater). Thisretestingwillalso reduce the problem o f false matches (where a degree o f change value o f +2 could be obtained by a 5 moment followed by a 3 moment and a 3 moment followed by a 1 moment). It is, o f course, not even necessary that a real match, from our perspective, take place. Itisenoughthatitseemslikeamatchtothemachines(althoughanyfalsematch willdissolvefairlyquickly. Itisunlikelygiventhenumberofpossiblecombinationsinour 5 object system that the difference ratio between moments would remain the same in two
different districts for more than 3 moments). The machine edits out A inputs different from its own internal state as "noise" once a match has been discovered.
14. 11 Looking for Mr. Goodmachine
If a machine fails to find an auditory match with its own O Difference Detector
after 3 vocalizations ( in order to give machines which have not yet vocalized but are within the same district time to communicate with this other machine before it moves), it moves the distance possible in 3 moments, and then trumpets its degree o f change value. (The machine has eyes on the sides o f its head so that it can see, and thus function within theprimaltime-scalethatdeterminesitsmomentsanditsrateofvocalization). Ifit processes a match in transit, the machine will stop and retest the match. If it finds a match then it doesn't move. Both machines are in the same district and, therefore, happiness!
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
600
Once a positive match has been made three things will happen. First, a pattern detector will correlate a signal from the Optical Difference Detector sent every Optical moment with its matched auditory input. It will discover that the match happens every 5 moments. Second, the correlated auditory and optical information is stored in Auditory/Optical Long-Term Memory (A/O LTM). Third, a signal is sent from the
A/O mediating agent to V in order to re-initialize the timing for vocalizations to match the vocalizationsofthemachinethatiswithinthedistrict. (Iamignoringthecaseofmore than one machine in the district, although the districts are large enough to accommodate more than one machine). At the Vocalizing Agent a difference detector is installed in order to detect the change from the previous vocalizing timing. This information is then sent to A/O LTM. The A/O LTM, therefore, is organized as a series of bifurcated units representing an Auditory value and an equal Optical value within the domain marked by the change in vocalization timing.
14. 12 Meta-temporal identities: the modeling of others as syntax
The A/O Evaluator has constructed a meta-temporal identity within LTM, an
identity (not yet an entity) that organizes temporal information (two levels of abstraction above the initial perception of change), instead of representing particular temporal states. The A/O Evaluator has integrated the O agent and A agent, and in so doing has created a more abstract and mediated information structure. The information in the A/O LTM, because it is distinguished from the machine's own vocalizations, represents the external world. This is an external world, however, that includes representations of what the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
601
machine recognizes as internal states. By matching A input with its own phenomenal perception the machine models itself within itself by modeling another machine (although the machine understands this as simply a sensory input). As a model for humans, this describes consciousness as a means ofconstructing a model ofthe minds ofother humans in order to predict their actions and manipulate them.
It is not quite so simple, however, to ascribe this A input to another machine when itisimpossibleforonemachinetoidentifyanotherasthesourceofthisinput. The machine must first recognize its internal state, the information generated by A, V, and O, as its own internal state. In my machine self-realization emerges from the synchronization o f the machine's vocalization with the vocalization represented by the matched auditory input. This synchronization is a kind of internal state change, caused, however, by an external stimulus. Because the machine is able to separate out its own vocalization, it can distinguish between I and not-I in relation to this vocalization and its recognition as a vocalization generated from its optical system. By matching the auditory input with its O Evaluator input it constructs an informational object that represents the not-I as the I, and the I as a not-I. This information, the meta-temporal identity, can have not meaningful significance, however, before it is used in some control function.
The short-term memories generated by those auditory inputs that do not match the machine's own degree o f change value have little significance, until the matched inputs cease. Once this happens, there is no longer an internal representation o f the temporal history o f the A agent. It is critical for our objective o f constructing a temporal syntax o f thefuturethatthisnothappen. Whilesuchacontinuoushistorydoesnottakeplaceinour
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
602
brains, an analogous history is written in our external representations of time. The earliest o f these are the notations and sticks and bones, identified by Marshack, marking the lunar cycle. There are curious parallels between these 35,000 year old markings and the internal symbolic economy ofthe machine. These markings are a record ofexperimental observations. One cannot imagine a use for them unless their exists a cultural logic, supported by long-term prediction and planning, into which the notations fit. Once could imagine something like, "Every other lunar cycle, at the first quarter moon and the final new moon, we will do X. " These markings become signals both of time and correlated human activity. The need to determine events through this kind of record keeping suggests a complicated social organization that requires some objective form of organizing its activities, and which supports a wide variety of cultural activities (where questions of correct timing might cause conflict without an "objective" time map).
The markings, however, are non-arithmetic. This means the lunar cycle is not defined by a number o f days (nor are any o f the phases), but by grouping markings in a general pattern. Thus the number of days may vary between the full and half moon in any month (depending on the observations of the scribe), but their basic temporal relationship remainsthesame. ThiskindofgroupingissimilartothecreationofdomainswithinA/O LTM. Thus the regularity of the moon changes is analogous to the less regular, but still periodic, vocal timing changes. In the human mind the creation of organizing syntactical domains is largely a social phenomena, and thus needs to be externalized. Because my machines only form very simple social groups, these structures cannot be supported by culture.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
The ability to perform this kind o f representational abstraction requires the same kind o f ability to derive form from particular content that I am trying to build into the Time Machine. For early humans, a single mark stands for the idea, the concept of a unit relative to the lunar cylce, and not its particular content (except to the degree that an event or action is meaningful in relation to the logic of the notational series as a whole).
Similarly the grouping o f these units, because they lose the particularity o f their content, can be used to "mean" a particular lunar period. These kinds of representational systems are symbolic languages, where markings, like words, are generalized in their meanings (word less so than the marks for days) so that they can be exchanged and organized within ameta-structurenotoverwhelmedbydetail. Symbolizationisaprocessofsimplification that allows an event to be reduced to a particular representation that can then be used to form other meaningful statements. All units within these lunar calendars must be relatively identical, in the way that words define events in relation to pre-established categories of meaning.
For our machine a continuous representation of time will be used, like it was in early human bone markings, to define a continuous temporal structure in which all machine activity can take place. If the A input ceases to match the O Evaluator values, the A unmatched STM is shunted and nested within the A/O LTM. By nesting I mean the informational structure generated in the A unmatched STM is translated into the language, the form, of the A/O LTM. The A unmatched STM provides a temporal link betweentheendingofonemachinematchtothebeginningofanother. Theinformationit stores in A/O LTM must fit within the parameters previously set up to accommodate the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
604
matched information. The unmatched information (e. g. a difference o f -3 between the machines internal Optical difference and its auditory input) has no meaning in relation to thematcheddifferencevaluesof0,unlesssimplytoindicatematchorunmatched. This kind o f simplification, or abstraction, is necessary if the unmatched STM is to function in relationtoA/OLTM. Theeffectofsuchanesting,therefore,istoignorethecontentof the Auditory (A) unmatched STM, and simply use this information to mark the primary temporal units constructed in the Optical Agent. The structure of A unmatched STM is used to augment and continue the structure ofAJO LTM. . Once this is accomplished, the A/O LTM can be understood as a continuous, simplified representation, through external surrogates,oftheinternal temporalstructureofthemachine.
Because the A input and O input values are correlated with each other, their content is in effect erased. Because each memory slot in A/O LTM encodes so little information, it becomes in effect a series o f marks which are themselves marked by the change in vocal timing. Sensory input is organized within a structure generated in relation to an internal state change. The A/O LTM models both external and internal time within asinglestructure. Evenmoresignificantlythelevelofabstractionoftheinformationused by the A/O Evaluator has allowed the creation of a model of the structure of time divorced from its content (defined in relation to a unit o f difference, not an instant in time). Our problem now is to figure out how to use this structure.
14. 13 Negative entropy
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
605
The A/O LTM gives our machine the ability to make a long range prediction. We must assume that at least one machine can outlive a number of its matches. If so, the only pattern a memory analyzer o f A/O LTM could identify would be that every vocal synchronization change will be followed by another change after a period o f time. Because the content of the A/O LTM is so poor, if a change in vocal synchrony occurs and recalls the memory of past changes (that is, the pattern of such changes), it will recall the temporal structure embodied in the LTM. Thus, whenever a new match occurs a long-range prediction would be possible: this vocal change will be followed by a series of differences (temporal units) which will themselves end in a new vocal timing change.
The structure supporting this prediction (the moments represented by the O Difference
Detector) will be in place until this expectation (the next vocal timing change) has been met, at which point it will begin again. Attached to the A/O Evaluator will be a memory recall agent that will work in parallel with the Evaluator, but will also organize both A and O input within the temporal structure imported from the A/O LTM. We have a future.
14. 14 Re-building ourselves in the other
This leaves us with two problems however. How can we integrate the other
perceptual, evaluatory, and memory agents into this structure? And what is the significance o f the two states (match and unmatched) within the temporal series in the A/O LTM?
The notion of the future, dependent as it is on a society of machines (identical in everything but life-span), requires a society of different sensory agents. Is it necessary,
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
606
however, to integrate the O LTM, with its patterns and short-term predictions, within the long-term future of A/O agent? If our machine receives optical information that triggers a short-term prediction, we might want this to signal the A/O Evaluator. But the A/O agent could not understand this kind of pattern (between numbers of objects). Such a pattern would not be reflected in its degree of change information. The only patterns it could recognize, other than the change in vocal synchronization, is the shift from its shunting of matched to unmatched information to LTM and a correlation between a matched auditory inputandeveryfivesignalsfromtheOpticalDifferenceDetector. IftheAagentcould activate a pattern recognition agent to evaluate the unmatched A STM that would be sensitive to more abstract patterns o f difference (every degree o f change 2 is followed by a degree o f change 4), and if we pretend that these patterns are present in some districts, then our two memory structures could be integrated easily by running them both through
the same memory recall agent. The existence o f such patterns in the machine world would indeed be convenient, but then so is the bipolar structure of water.
We have three sub-agents, constituting three internal time scales, not fully exploited in modeling time: the (1)V Difference Detector indicating a change in vocal timing, the (2) A/O Evaluator's awareness that every match happens after 5 optical moments, and the (3) A/O Evaluator's (one o f its internal agents) ability to detect the shift from matched to unmatched outputs to memory. (1) and (3) constitute long-term regularities, while (2) is a short-term prediction that can act as the structural content organizing the temporal structure enabling long-term predictions.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
607
The machine's change in (1) vocal timing finds its analogue in the death o f the external matched signal: the first registers as an internal change; the second as an external change. We still do not know if our machine understands this as the death o f its machine mate. An existential understanding ofthe loss ofthe matching voice would require (ifwe ignore the lack o f a mammalian brain, or some might say any brain) an analogic equivalence between the A/O Evaluator shift to unmatched memory output and the death o f itself. It seems unlikely that a machine without a more developed symbolic life could
understand this loss. The predictive power of this kind of system, however, remains stimulus dependent. When the stimulus vanishes, the future disappears.
14. 15 Thepursuitofdeath
If we could make a higher level analyzing agent to determine and chart the larger patterns of change in A/O LTM [(3): matched to unmatched], we might be able to stimulate the machine into a weak awareness of the subjectivity of another machine encoded in the A/O LTMs meta-temporal identity. The A/O Evaluator sends a signal to
this agent when the shift from matched to unmatched occurs, and this signal causes this Pattern Recognizer to recall previous shifts and their defining vocal timing domains. The Pattern Recognizer could construct an information structure that charts the internal state changes against the representation o f catastrophic loss o f the external matching voice.
The information such a chart could encode is rather simple, and the prediction that follows even simpler: every vocal timing change will be followed not only by another, but also by the loss ofthe not-I representation. Because the representation ofthe not-I is used to
with permission of the copyright owner. Further reproduction prohibited without permission.
determine the actions o f the machine (it stays and observers), its loss results in a change in action (the machine now moves). Before moving can be a form o f mourning, it must also representalossofbeingforthemachine. Ifweremovethemachine's eyesonthesideof its head, it will lose its primary temporal portal. It will still have its auditory input (and STM) by which it can chart a different temporal order, but the loss of this primary information will be recorded with the shift from matched to unmatched in the A/O LTM and will be noted by the A/O LTM Evaluator. This sensory loss, and the time/ memory scale it constructs, can stand for, can symbolize, the loss o f the external voice and the change in vocal timing. The prediction that a loss o f the external voice will necessarily follow a change in the machine's own vocal timing will also be tied to another internal
effect: thecommandtomoveanditssubsequenttemporarylossofsightandthetime- scale that sight generates. This causal structure (Cause [external loss] and Effect [internal loss]) mirrors the vocal timing change and is nested within the basic structure generated from that timing change. Our sense o f the future is growing more complex.
14. 16 Learning from the future
The bifurcation o f the memory and predicative domain defined and signaled by the
vocal timing change sets up a meta-time scale at which level internal symbolic exchanges begin to make some kind o f sense. The O agent wants to be able to predict when it will fail, and when it's short-term predictions will be irrelevant, while the A/O wants to be able to predict when it will have to shift its memory output from matched to unmatched. Our machine could not discover such detailed information, nor could these agents as they now
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
609
stand actually want anything. They would first have to need this information as a means to act. But what is important here is the conceptual problem this presents. For the Optical Agent the loss of the external voice could stand for its loss o f sight, and because it happens prior to its effect, it could serve as a short-term prediction (generated out of a long-term prediction structure). This represents an internal symbolic communication, where one loss can stand for another. The signal itselfwould go from the A/O Evaluator to the O Agent. This signal is not, however, caused by an algorithm: if shift to unmatched output, send a signal to O that loss of sight is imminent. The meaning ofthe shifttounmatchedoutputisdeterminedbytheA/OLTMPattern. Itconcludesfrom previous patterns that the shift means a loss o f sight. This is a conclusion based on an internal analysis (abstraction) from a lower level abstraction (A/O LTM) which is in turn
generated from lower level sensory modeling. This is not the most elegant of control structures (but neither is much of our learning compared to "lower" animals: human weavingcomparedwithspiderweaving). Ithastheadvantage,however,ofmakingall predictions a function ofmemory (they are learned), and not of pre-programming.
14. 17 Inside a Chinese box: when structure becomes content within larger structures
A matched memory state (any input of value 0) within the A/O LTM, re- interpreted in terms o f the time scale generated from the Optical Agent, stands for 5 Optical moments. If a match has been found, then every five O moments will result in a matched input to A/O LTM. The Pattern Recognizer attached to A/O Evaluator
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
610
functions as a symbolic catalyst linking the two time scales.
