University of California
Los Angeles CA 90095-1530
The Power of Simulation: What Virtual Creatures Can Teach Us
Yearning for the light, the creatures struggle after it. In water
they grow tails and learn to undulate like snakes. On land they clump along,
relegated by fate and biology to rectangular shapes joined together with
moveable hinges. They show extraordinary ingenuity in making the most of
these limitations, crawling, hopping, jumping, always toward the light.
Then their creator gives them a new goal, a colored cube reminiscent of
a squared-off hockey puck. Put into competition with one another, the creatures
learn to jostle and shove their opponents, to encircle the cube, to knock
it out of the way so their opponents can't reach it. When they meet a new
opponent, they develop counter-strategies to meet these challenges. I marvel
at their adaptability, cleverness, and determination.
This passage describes my reactions while watching the videotape of Karl Sims' evolutionary simulation, "Evolved Virtual Creatures." Judging from audiences I have seen viewing the tape, my responses are typical. Invariably viewers attribute to these simulated creatures motives, intentions, goals, and strategies. Even people who know perfectly well they are seeing visualizations of computer programs still inscript the creatures into narratives of defeat and victory, cheering the winners, urging on the losers, laughing at the schlemiels. Much more is going on here than simple anthropomorphic projection. "Evolved Virtual Creatures" is a laboratory not only in evolution (its intended purpose), but also in the impact of distributed cognitive systems on traditional modes of description, analysis, and understanding. Emerging from this laboratory are resources to re-think the much-ballyhooed divide between scientists and cultural critics, especially the conflict between seeing the body as an externally-existing object with a more or less constant physical reality (the view of most biologists), and as a discursive construction varying widely across cultures and historical periods. Part of what virtual creatures can teach us is that this divide is itself historically contingent, a result of the on-going transition from the traditional liberal self to the contemporary posthuman subject.
But I am getting ahead of my story. Let us return to the virtual creatures to explore their construction and dynamics. Compared to the world in which we live, the environment of "Evolved Virtual Creatures" is extremely simple, so simple that it can be described almost completely. How to define the boundaries of this world is a centrally important issue to which I will return. For now, let us consider the world to be the computer programs, the hardware on which the programs run, and the visualization routines that render these programs as pixelated images of embodied creatures. Even this simple world is complex enough to require three different modes of interrogation: what it is (the material); what it does (the operational); and what it means (the symbolic). Feedback loops connect the material, operational, and symbolic into an integrated, recursively structured hierarchy. At the bottom of the hierarchy are electronic polarities, where the material and operational join to create bits, the semiotic markers of one and zero. Logic gates structure signals into bits; compiler languages are built out of bit patterns; programming languages are created from compiler languages; and functions are defined using programming languages such as Lisp. By the time we arrive at functions, the level at which Karl Sims discusses his design for "Evolved Virtual Creatures," we have reached a point where the patterns created by the programmer become explicit. Instantiated in these patterns are the programmer's purposes in creating this particular hierarchy of materio-semiotic codes.
Sims's design follows John Koza's proposal that evolutionary programs should take advantage of modular structures which can be repeated over and over to create more complex structures. The strategy appears often in nature; a fern, for example, displays a growth algorithm that uses the same basic shape for stems, branches, and leaves. Like the fern, Sims's creatures are built using functions that are repeated with variations to create self-similar morphologies. One function specifies blocks that are multiplied and attached at various positions on a central rectangle to create a "trunk" with several "limbs." Another function specifies the kind of articulation, or joint, between blocks; still another, the degrees of freedom through which a joint can move. Recursive loops within a function multiply the effects of that function to create more of the same, for example, more limbs of the same shape. Recursive loops between functions allow different parts of the creature to evolve together, so that the "brain" or central control circuits co-adaptively change with the morphology. The advantages of these modular structures, achieved by using programs called directed graphs, are two-fold. In addition to economy of description (because the same module can be used repeatedly with minor variations), the modules also ensure that some structure will persist in the midst of mutation and variation. If all of the programming elements were subject to mutation as independent entities, the resulting complexity would quickly become too chaotic to track effectively. When the elements are grouped and mutated as modules, the spectrum of possible variations is reduced to a manageable level.
The next step moves from the design of individual creatures to a population of creatures. With this step, the symbolic aspects of the program become apparent. The idea is to evolve creatures by introducing diversity into the population and defining fitness criteria that determine which creatures get to reproduce. Diversity is accomplished through sexual reproduction that, following various schemes, combines portions of one creature's genotype with another's. Additional diversity is introduced through mutation. Behaviors take place within an environment governed by an artifactual physics, which includes friction, inertia, momentum, gravity, light, three-dimensional space, and time. Fitness values are determined according to how successful the creatures are in reaching various goals-- following a light, moving through fluids and across terrains, cornering the puck while keeping an opponent away from it. To facilitate adaptation to these goals, the creatures are given photosensors which can neurologically evolve to respond to a beacon, the presence of the puck, and positions of competitors, each represented by a differently colored light source.
The designer's intentions, implicit in the fitness criteria he specifies and the values he assigns to these criteria, become explicit when he intervenes to encourage "`interesting'" evolutions and prohibit "inelegant" ones ("3-D Morphology," pp. 31, 29). For example, in some runs creatures evolved who achieved locomotion by exploiting a bug in the way conservation of momentum was defined in the world's artifactual physics: they developed appendages like paddles and moved by hitting themselves with their own paddles. "It is important that the physical simulation be reasonably accurate when optimizing for creatures that can move within it," Sims writes. "Any bugs that allow energy leaks from non-conservation, or even round-off errors, will inevitably be discovered and exploited by the evolving creatures," ("Evolving Virtual Creatures," p. 18). In the competitions, other creatures evolved to exceptionally tall statures and controlled the cube by simply falling over on it before their opponents could reach it ("3-D Morphology," p. 29.) To compensate, Sims used a formula that took into account the creature's height when determining its starting point in the competition; the taller the creature, the further back it had to start. Such adjustments clearly show that the meaning of the simulation emerges from a dynamic interaction between the creator, the virtual world (and the real world on which its physics is modeled), the creatures, the computer running the programs, and in the case of visualizations, the viewer watching the creatures cavort. In much the same way that the recursive loops between program modules allow a creature's morphology and brain to co-evolve together, so recursive loops between these different components allow the designer's intent, the creatures, the virtual world, and the visualizations to co-evolve together into a narrative that viewers find humanly meaningful.
An adequate account of the simulation, then, requires expanding the boundaries of the system beyond the programs and computer to include the virtual world, the creator, and the viewer. The evolutionary dynamics of this larger world functions as a distributed cognitive system composed of human and nonhuman actors, each of which also acts as an independent cognizer. As Michael Dyer has noted in another context, with distributed cognitive systems there is no free lunch: the fact that all the parts interrelate means that if one part of the system can only function as a relatively low-level cognizer, the slack has to be taken up somewhere else by making another part smarter. Compared to artificial intelligence, artificial life simulations typically front-load less intelligence in the creatures and build more intelligence into the dynamic process of co-adapting to well-defined environmental constraints. When the environment fails to provide the appropriate constraints to stimulate development, the creator steps in, using his human intelligence to supply additional adaptive constraints, for example when Sims put a limit on how tall the creatures can get. But it would be a mistake to see the creator as the court of last resort. The point of such simulations is that the creator does not always need to be as smart as his creatures, for he is counting on their ability to come up with solutions that have not occurred to him. "When a genetic language allows virtual entities to evolve with increasing complexity," Sims observes, "it is common for the resulting system to be difficult to understand in detail. In many cases it would also be difficult to design a similar system using traditional methods. Techniques such as these have the potential of surpassing those limits that are often imposed when human understanding and design is important. The examples presented here suggest that it might be easier to evolve virtual entities exhibiting intelligent behavior than it would be for humans to design and build them" ("3-D Morphology," p. 38).
Since distributed cognitive systems co-evolve together, the functioning
of any one actor can be fully understood only in relation to that actor's
interactions with all the other actors. It is in this context we should
think about the narratives human viewers create for themselves when they
watch "Evolved Virtual Creatures." Spliced into a distributed cognitive
system, we create these narratives not by ourselves alone, but as part
of a dynamic evolutionary process in which we are co-adapting to other
actors in the system, including pixelated images on a CRT screen and electronic
polarities flickering beyond the scale of human perception.
When we attribute to Sims' virtual creatures motives and intentions, we interpolate their behaviors into narratives in which events are causally related to one another and beings respond to their environment in purposeful ways. As Alex Argyros among others has suggested, the creation of narrative may itself be an evolutionary adaptation of remarkable importance. Narratives, with their emphasis on causality, meaningful temporal sequence, and interrelation between behavior and environment, allow us to construct models of how others may be feeling and acting, models which co-evolve with our on-going interior monologues describing and interpreting to ourselves our own feelings and behaviors. When narratives for some reason cannot be constructed, the result is likely to be a world without order, a world of inexplicable occurrences and bewildering turns of events. Simon Baron-Cohen describes such a world in Mindblindness, suggesting it is characteristic of how autistic people perceive their environment. As Baron-Cohen points out, autism is associated with an inability to construct narratives that will make sense of the behaviors of others. Autistic people have no model in their minds for how others act, and consequently, most actions are to them inexplicable and frightening. Another graphic description of a state of extreme disconnection akin to autism is rendered in Joan Didion in TheWhite Album.9 To illustrate, she describes a New York street scene as a terrifying cacophony of isolated and therefore unpredictable events in which nothing relates to anything else. What these accounts make clear is that narrative has an explanatory force which literally makes the world make sense. It is easy to see why the creation of narratives would confer evolutionary advantages on creatures who construct them. Without the presuppositions embedded in narratives, most of the accomplishments of Homo sapiens could not have happened.
When we construct narratives about virtual creatures, we use an evolved behavior to understand the evolved behaviors acted out in the simulation. It is no accident that in this scenario a feedback loop appears whose recursive structure resembles the recursive structures of the programs generating Sims's virtual creatures. Across a wide variety of research programs, from Stuart Kauffman's claims for the evolution of life at the edge of chaos to John Koza's work with artificial life simulations and Humberto Maturana's theories of autopoiesis, recursive loops are associated with the emergence of complexity and consequently with life, consciousness, and intelligent behavior. Maturana, for example, suggests that consciousness consists of the ability to make representations of representations (of representations of representations . . . ). Luc Steels has named this phenomenon of spiraling recursions second-order (and higher) emergence and underscored its importance for artificial life simulations. (First-order emergence, of course, is any behavior or property which cannot be found in a system's individual components or their additive properties, but which arises, often unpredictably, from the interaction of a system's components. Emergent properties appear on the global level of the system, not the local level of the system's parts.) Second-order emergence arises when a system develops a behavior that enhances its ability to develop adaptive behaviors--that is, when it evolves the capacity to evolve.12 The goal of most artificial life simulations is precisely to achieve such second-order (and higher) emergence, for then the simulation really takes off. Given the importance of recursive structures to complex systems, it is possible that autistic "mindblindness" is associated with the highly recursive structures of the central nervous system. Cut off from the recursive spiral that folds explanations for the behavior of others back into the construction of self through interior monologues, the autistic person would be unable to create the narratives we use to make sense of the virtual creatures.
"Mindblindness" is an exquisitely appropriate term, for in addition to recursive structures, another important element in the creation of narrative is the ability to "see" a scene, either literally or metaphorically in the mind's eye. With training and experience, humans are able to translate a large variety of inputs into these imagined scenarios. No doubt an experienced programmer such as Karl Sims can look at a program's functions and "see" the morphologies and behaviors of his creatures with no more difficulty than an experienced novel reader can "see" Isabel Archer in Henry James' tellingly entitled novel, The Portrait of a Lady. These translation processes draw upon and extend capabilities developed in evolutionary history. Our sophisticated perceptual-cognitive visual processing evolved co-adaptively with our movement through three-dimensional spaces, so it is no surprise that the creation of narrative is deeply tied up with imagining scenes in which actions can take place. When Sims chooses some of his creatures for visual rendering, he taps into this evolutionary history by creating pixelated images that, through culture and training as well as biologically-determined capacities, we recognize as representations of three-dimensional spaces. Articulated in this lingua franca of Western cultural perception, the images allow narrative to kick in with maximum force, for the action is "seen" in terms we can easily relate to our on-going narrativizing of the world. ( I recently came across an advertisement for an academic job encouraging "visible minorities" to apply, a phrase I had not heard before and that immediately evoked thoughts of Ralph Ellison's Invisible Man. Leaving aside the complex cultural history embedded in this phrase, it serves the purpose here of highlighting the importance of making bodies visible if they are to enter into the canonical narratives of a culture, a requirement that seems both strange and familiar when applied to visualizations of virtual creatures.)
Let us turn now from the structural preconditions for the creation of narratives to content. As Jerome Bruner has pointed out, one of the principle purposes narrative serves is to create a sense of why things happen. The narratives we create typically inscript actions into a set of more or less canonical stories that invest actions with meaning. When Joan Lucariello studied which stories stimulated the most vigorous creation of narratives by young children, she discovered that non-expected actions gave rise to the most stories in response, for example, a description that has Mary crying when she sees her birthday cake and dumping a glass of water all over the candles. To make sense of these strange actions, the children invented a wide variety of stories that had the effect of suturing the actions back into a predictable and expected range of behaviors. In one small child's account, Mary was upset because her mother wouldn't let her wear the dress she wanted, and that's why she cried and ruined her cake. Presented with non-canonical actions, the children sometimes employed another narrative strategy of marking the behavior as unusual or deviant, which again allowed the social fabric of expectations to be maintained by bracketing this behavior as an exception. It is surely no accident that in his evolutionary simulations Sims designs programs which can be "seen" as creatures striving after a goal and winning against competitors, for these are among the most canonical narratives in traditional accounts of evolutionary history (not to mention in Western capitalist society). The banality of the narrative content suggests that what needs to be sutured here is not so much deviant action as the deviant actor. When we "see" the virtual creatures engaging in these activities, we have models in our minds for what these behaviors mean and so the creatures, despite their odd shapes and digital insides, seem familiar and understandable.
At this point, some readers may object that however functional narrative may be for everyday social intercourse, it leads to serious mistakes when we use it to understand virtual creatures. Not only do these creatures have nothing in their heads; in a literal sense, they have no heads (because they are virtual, and because their morphology is a series of blocks, the uppermost of which we "see" as the head). Attributing desires to these clumps of blocks may seem as ridiculous as thinking electrons have motives. Well, yes and no. Certainly the creatures are merely computer programs which have evolved certain behaviors (and therefore attributes we interpret as embodied action toward a goal) as a result of the fitness criteria used to select which genotypes will be allowed to reproduce. Or more accurately, which coding arrangements will be replicated with what variations, since "genotype" and "reproduce" are themselves metaphors designed to reinforce the analogy with biological life forms. On the other hand, these programs are designed to simulate biological evolution and visually rendered so that narrative inscripting will take place. There is a sense in which we respond correctly, not mistakenly, when we attribute desires to these virtual creatures, for everything about them has been crafted to ensure that such interpretations will occur.
One way to think about this situation is to note that distributed cognition also implies distributed causality. The creatures may not have motives and intentions, but the programmer does (at least in the conventional understanding of human actions). Remember that what we "see" in the visualization is the global result of present and past interactions between all the actors in this recursively structured complex adaptive system. When Sims decides which fitness criteria to use, which programs to eliminate, and which to render visually, he injects doses of his human intelligence into the system, along with the attributes we conventionally assign to humans, including desires and intentions. Moreover, his articles clearly show how his intentions affected virtually every aspect of the design, so it is not possible to bracket out his intentions by saying we should consider only the programs in themselves, not the global system.
Human intentionality, then, infects the creatures, marking them with a trace that cannot be eradicated. Recall that in this recursively structured complex adaptive systems, all of the actors are involved in and therefore affected by the interactions. Is it also the case that the blind operations of the programs infect the humans, marking them with a trace that cannot be eradicated? To entertain this hypothesis is to suppose that the human tendency to anthropomorphize the creatures has as its necessary and unavoidable supplement a counter tendency to "see" human behavior as a computer program carrying out instructions. We may think we have desires and intentions (just as we think the creatures do), but our behaviors can be explained materially and operationally in terms similar to Sims's programs. This argument that has been made by several researchers in artificial intelligence and artificial life, including Rodney Brooks and Marvin Minsky. In their view, human behavior is the result of many semi-autonomous agents running simple programs. To illustrate, Minsky suggests that "love" is a combination of one agent running an "attraction" program and another agent running a program that shuts off the "critical" agent. Such proposals indicate that anthropomorphizing the creatures is accompanied by what I might call, for lack of a better term, computationalizing the humans. According to the logic of this relation, blind programs engaging in human-like behaviors make plausible the interpretation of human behaviors as blind programs. We humanize the virtual creatures, they virtualize us, and the recursive loops cycling through the system bind both behaviors together in a network of complex co-adaptations.
Which leaves us with an interesting question: what happens now
to narrative and its function of making human sense of the world?
Computing the Human: Analogue and Digital Subjects
Following the work of Michel Foucault on the death of the author, Mark Poster has expanded on Foucault's fourth and final stage of the author's disappearance to suggest that digital technologies and culture are bringing about a significant reconfiguration of contemporary subjectivity. To illuminate this shift, Poster posits two different kinds of subjects, which he calls analogue and digital, respectively. The analogue subject is based on relations of resemblance. Although Poster does not use this example, the mind-heart conjunction illustrates the concept. Consciousness is taken to resemble the soul or the heart, so that what is at the forefront of mind is also imagined to be deep inside. Similarly in the English Renaissance, a period governed as Foucault has shown by cultural relations based on analogy, human sperm was thought to contain a homunculus resembling the man who would grow from the sperm. Walnuts were considered to be "brain food" because walnut meat resembles the human cortex. Analogical relations require that the integrity of the units taken to resemble one another be preserved; otherwise, the correspondence is lost and the relation broken. If one tosses a handful of walnuts into a blender and turns it on, the walnuts are pulverized and no longer resemble a cortex. If walnuts were only available in this form, it seems unlikely they would have been considered good food for thought. Attributes of the analogue subject include, then, a depth model of subjectivity in which the most meaningful part of the self is seen to reside deep inside the body, and units which have a natural integrity of form and scale which must be preserved if the subject is to be maintained intact.
Drawing on Mark Rose's work on copyright, Poster focuses his discussion of subjectivity on the "cultural figure of the modern author," a figure that emerged in the eighteenth century in a "confluence of print technology, a book market, a legal status, and an ideology of individual as creator" (p. 6). In Poster's view, analogue subjectivity is deeply bound up with the dominance of print culture. At the same time alphabetic writing breaks the pictorial analogical resemblance that connects an ideogram to the object represented, it forges a new connection between a sound and a mark. This connection differs from pictorial writing in that the association of sound with mark is entirely conventional, and the resulting arbitrariness makes alphabetic writing much more economical than ideograms (thousands of ideograms versus some thirty letters of the Greek alphabet). There is also another shift, for now the resemblance is not between word and thing but between "a written symbol and its utterance, between two forms of language, writing and speech. The relation between the word and thing becomes conventional, arbitrary, whereas the relation within language between trace and voice is stronger, more direct" (p. 22). Thus to the extent that print can be considered an analogue medium, it connects voice to mark, and thus author as speaker to the page.
Reinforcing the sense that print texts are "voiced" by an individualistic creator is the uniformity, stability, and durability of print. "The reader could return time and again to the page and re-examine the words it contained," Poster writers. "A readerly imaginary evolved which paid homage to this wonderful author who was always there in his or her words. . . the world of analogue authors was leisurely, comforting, reassuring to the cognitive function and expanding through continuous exercise of the visual function" (p. 26). Although literary history is largely outside the scope of Poster's analysis, the role of the novel in reinforcing the depth model of interiority and the stability and individuality of the analogue subject has long been recognized in literary studies. The legal fight to insure copyright, the cult of the author, print technology, and print culture worked hand in glove to create a depth model of subjectivity in which analogue resemblances guaranteed that the surface of the page was matched by an imagined interior within the author, which evoked and was also produced by a similarly imagined interior in the reader.
In contrast to this dynamic are the correspondences producing the digital subject. Digital technologies employ hierarchical program structures similar to those we saw at work in "Evolved Virtual Creatures." Unlike the depth model of meaningful interiority in the analogue subject, the further down into the coding levels the programmer goes, the less intuitive is the code and the more obscure the meaning. Speaking as someone who has programmed in machine language, I can testify how murderously difficult it is to translate thoughts into binary code. Programming in C++ is a breeze by comparison. Moreover, with genetic algorithms and programs the important developments are emergent properties which appear at the global level of the system once the programs are set running. The mantra for such programs is "simple rules, complex behaviors," which implies that the further down into the system one goes, the less interesting it is.
Although the digital subject has depth, the structures governing the relation of surface to interior differ dramatically for the analogue subject. The digital subject--say, one of Sims's virtual creatures--instantiates hierarchical coding levels operating through a dynamic of fragmentation and recombination. Unlike analogue subjectivity, where morphological resemblance imposes constraints on how much the relevant units can be broken up, the digital subject allows for and indeed demands more drastic fragmentation. This difference can easily be seen in the analogue aspects of print media compared to the fragmentation of digital technologies. Each letter of the alphabet must be treated as a distinct unit for writing to be legible, and the corresponding phoneme also acts as an intact unit. In contrast are digital sampling techniques, where sound waves may be sampled some 40,000 times a second, digitally manipulated, and then recombined to produce the perception of smooth analogue speech. In fact emergence depends on such fragmentation, for it is only when the programs are broken into small pieces and recombined that unexpected adaptive behaviors can arise. Instead of a depth model of meaningful interiority, the digital subject manifests global behaviors which cannot be predicted by looking at the most basic levels of code when the program begins running. Complexity becomes visible first on the surface, not deep inside. Moreover, the complex surface bears no analogical resemblance to the mind-numbing simplicity of ones and zeros.
To summarize: the analogue subject implies a depth model of interiority, relations of resemblance between the interior and surface that guarantee the meaning of what is deep inside, and the kind of mind-soul correspondence instantiated by and envisioned within the analogue technologies of print culture. The digital subject implies a surface complexity which is related through hierarchical coding levels to simple underlying rules, a dynamic of fragmentation and recombination that gives rise to emergent properties, and a disjunction between surface and interior instantiated by and envisioned within the digital technologies of computational culture.
What happens when we become part of a complex adaptive system by "seeing" the virtual creatures? I suggested earlier that two processes are at work simultaneously: on the one hand humans anthropomorphize the virtual creatures, and on the other hand the virtual creatures computationalize the humans. The narratives we construct as we watch the virtual creatures inscript their behaviors into an analogue world, but observant viewers will notice details that cannot be explained by supposing that the complex surfaces are matched analogically with equally complex interiors. One creature, for example, moves in a way that indicates it samples the position of the puck and opponent once at the beginning of the competition and thereafter ignores all cues about position. Clearly, here is an instance of a relatively simple program creating an impression of surface complexity that contrasts with the simplicity of the underlying rules. Another example is provided by a small mobile robot made by Lego that, on the surface, appears to be capable of following a white line on a black ground. A viewer might suppose that inside the robot is an intelligent program that has an internal representation of a line and can match this representation up with what it sees so accurately that it can distinguish many different kinds of lines, including ones that are curved and even looped. Underlying the surface complexity, however, are three simple rules: if from white to black, turn right; if from black to white, turn left; if no change, continue straight. Although the robot globally follows the line, this is an emergent behavior. In fact, it simply swerves left when it first comes across a white line on a black background and then, as it begins to veer off the line, immediately turns again, so its "line-following" behavior consists of a series of small swerves that a viewer may interpret as corrections the robot initiates to make sure it follows the line. Simple rules, complex behavior.
On the global level, our narratives about the virtual creatures
can be considered as devices that suture the analogue subjects we still
are as we move in the three-dimensional spaces in which our biological
ancestors evolved together with the digital subjects we are becoming as
we interact with virtual environments and digital technologies. In fact,
this essay can be read as a narrative designed to accomplish just such
a suturing. Hence my insistence on using the plural first person, despite
the risk of indulging in oppressive universalisms, for I want to insist
that my readers, like me, participate every day of our lives in the distributed
cognitive complex adaptive systems created by digital technologies in conjunction
with global capitalism. So pervasive have these technologies become that
it would be difficult to find anyone who remains completely outside their
reach. Certainly, here in the U.S. their presence is ubiquitous. In this
sense, we do not need to slot Sims' videotape into the VCR to watch virtual
creatures. We see them all the time, all around us, including when we look
into the mirror.
Reconfiguring the Material/Discursive Divide
Let me return to the suggestion that the material/discursive divide is a historically contingent formulation characteristic of a moment that may already be passing. To develop this idea, I will find it useful to review acccounts that scientific realists and cultural critics give in their respective projects. For the realist, the flow of structuring information about physical reality moves from the material (say, a field of morning glories of varied colors) through the operational (experiments in breeding that operate upon the plants and plant genomes to isolate colors from one another) to the symbolic (graphs and charts showing how the colors migrate back to an equilibrium distribution after being separated). The closer the researcher is to the embodied reality of the plants, the fuzzier the picture is likely to be as various sources of "noise" and "contamination" complicate the regularities presumed to be revealed by such inscriptions as graphs and charts. The idea is to remove the noise or, failing that, compensate for it as much as possible in the experimental design and subsequent analysis, so the form of the underlying regularities become sharp and well-defined.
In the movement from embodied reality to inscription, much is gained and some things are lost. The most important gains, of course, are the regularities revealed through the inscriptions, a point to which we will return. Also important is the implication that once these regularities are durably inscribed, they can circulate through different media without affecting their meaning. If I xerox the chart showing morning glory color distribution and discuss it with my research seminar, everyone assumes we are seeing the same graph that appeared in the scientific journal, even though the method of producing the image and the materials comprising it (toner ink and copier paper) differ from the original. Similarly, if the researcher illustrates a lecture on her work with slides, these count the same as the graphs printed on the journal pages. The case would be otherwise if we examined morning glory plants. Say I buy morning glory plants at Home Depot and take them in to my seminar. Since they are obviously not the same plants the researcher examined in her test fields some months earlier, questions would inevitably arise about material differences that may exist between our plants and hers. Material embodiments do not circulate effortlessly because they are always instantiated, specific, and located in a certain time and place. By contrast, inscriptions can circulate because cultural conventions distinguish between the forms expressed by the inscriptions and their instantiations in particular media such as print, xerox, and photographic negative.
Normally one says inscriptions are transportable or transmissible, but perhaps a more appropriate term to describe their circulation is transmigration. Just as the soul, conceived as a disembodied entity, is thought to move from one corporeal body to another in transmigration, so the abstract form of the inscription is counted as moving from one incorporation to another, despite differences between material instantiations. (A partial exception to this convention is the signature, which is presumed to embody the signer's material presence and so not to be transmigratable from one medium to another. A photocopy of a will does not count the same as the original signed document. This presumption of embodiment appears to be giving way with the spread of new communication technologies; faxes, for example, are increasingly accepted as legally binding documents. Even here, however, there continues to be some whiff of embodiment, for a fax occupies a different legal position than e-mail, which has no signatures that can be linked with embodiment).
Inscription, then, is crucially important to the transformation of embodied reality into abstract forms. It is worth noting that many, perhaps most, scientific instruments produce inscriptions through morphological proportionality to physical properties. Sound waves hit a membrane and the vibrations capture an analogue resemblance, which is conveyed through a linking mechanism to a pen tracing a line on a graph paper, which in turn bears an analogue resemblance to the vibrations. Even though scientific instrumentation increasingly uses digital technologies for analysis and imaging, there usually remain some portions of the chain that employ analogue representation, typically at the beginning and end of the process. Analogue relations appear at the beginning when data are initially gathered and recorded, because sensing instruments generally require proportionality to transform physical signals into data. Analogue representations appear at the end because humans can best interpret complex phenomena using their sophisticated visual-cognitive skills. I call this analogue-digital-analogue structure the oreo, for like the two black biscuits sandwiching a white filling between them, the initial and final analogue representations connected with embodied materialities sandwich between them a digital middle where fragmentations and recombinations take place. An example of an oreo structure is positron emission tomography, or PET images. The process begins with the ingestion of radioactive substances by the patient. An instrument senses the decay products using analogue proportionality, and the results are inscribed as an array of numerical data. These data are then digitally analyzed and manipulated to create life-like analogical resemblances that humans interpret as metabolic processes occurring with the cortex. These images are often read as showing "thinking in action," but they may more accurately be understood as showing the oreo effect in action.
Let me return now to inquire about the status of the forms transmigrating through inscriptions, an issue that goes to the heart of the differences between realist and constructivist viewpoints. From a realist point of view, the forms are always already instantiated in the embodied reality and the inscriptions merely reveal their true nature. From a constructivist point of view such as that articulated by Bruno Latour in Science in Action, the forms do not precede the inscriptions but are produced by them. In making this argument, constructivists point toward the contingencies and local conditions that always accompany embodied reality: the air pump cannot produce the same results in Holland as in England; two equivalent scientific instruments cannot be calibrated to produce the same results unless someone who knows how to calibrate the first instrument physically travels to the second one. Few doubt that regularities exist in nature, but the problem comes when these regularities are seen as "laws" that can be abstracted from embodied contexts and expressed as the transmigrating forms of scientific inscriptions. As Evelyn Fox Keller wittily puts it, every scientist knows what hard work it is to get nature to obey the laws of nature. Does nature count as the abstracted form or the embodied materiality, which is always more complex than the form allows? Moreover, how is it determined which abstracted form will count as an adequate signifier for the reality, an issue that leads directly to the social processes involved in establishing a claim as a scientific fact.
More extreme versions of the constructivist position are found in cultural studies, where critics talk about scientific facts as "truth effects" produced by discursive regimes. While it seems clear that such interpretations explain older cultural notions now discredited--the idea, for example, that inside a sperm is a homunculus that analogically resembles a man --the debate becomes more heated when the topic under consideration comes from current scientific research. Hence the famous "symmetry principle" in the social studies of science, which says that one cannot explain bad or incorrect science using one set of methods and good or correct science using another set of criteria. All this is well known and needs no further comment here. My point in rehearsing it is to point out the mirror relationship that obtains between these versions of scientific realism and cultural constructivism. While the realist assumes the flow of structuring information moves from material through the operational to the symbolic, the constructivist assumes it goes from symbolic (Enlightenment ideas about clarity of vision as an enactment of rationality) through the operational (Bentham's plans for a model prison) to the material (the construction of the Panopticon). Notice that both the realist and constructivist rely on analogical relations to make their cases: the realist by the analogical proportionality that ties the material to the operational, the embodied reality to the inscription; the constructivist through the horizontal analogical resemblances that tie one discursive site to another (prison reform to handwriting practices to medical theories). To break open the hegemony of scientific realism, it was no doubt helpful to take the strong counter position that Latour, among others, articulates in Science in Action . But such strategies are limited in their options by the very assumptions they resist. Defining himself by what he revolts against, the revolutionary ends up looking like his opponent reflected in a mirror.
The last decade or so of work in the cultural and social studies of science has been marked by various strategies to escape the limitations imposed by these symmetry relationships. The Latour of Science in Action differs significantly, it seems to me, from the Latour of We Have Never Been Modern, who insists that the objects of scientific research are at once discursively constructed, socially produced, and materially real. I want to put Sims' virtual creatures into conversation with these on-going debates, for I think they have something important to contribute. In my view, they suggest other ways to skew the symmetry relations of the materialist/discursive divide and to re-think the transmigrations of forms through inscriptions that have the effect of leaving embodied reality behind.
Unlike experiments in the natural world that which must abstract the forms out of embodied materiality, often with great effort and ingenuity, the forms underlying the virtual creatures are easily accessible and open to view. At the bottom of the hierarchy are the ones and zeros of binary code. Some see the emergence of complexity out of these simple elements as confirming the Platonic nature of reality. In this interpretation, eloquently articulated by Christopher Langton, a prominent researcher in artificial life, conventional science abstracts the forms through analytical procedures that start with complex phenomena and break them down into simpler components, while artificial life starts from the opposite end and uses synthetic methods to build complex phenomena out of simple components. In the view of Langton and others, these symmetrical relationships between the analytical and synthetic approaches confirm that the ultimate nature of reality is mathematical in form and computational in process. Edward Fredkin, for example, has an on-going research project dedicated to demonstrating that underlying quantum mechanics and particle physics are cellular automata, ultimately simple units governed by a small set of simple rules. Reality in this view is a program run by a universal computer, and computational code is the true language of nature. These interpretations do not contest the received view that forms transmigrate through inscriptions. Rather, they extend transmigration to computational processes and enlarge its domain to include biological and artificial life.
There are other ways to look at the success of artificial life simulations in creating virtual creatures. Instead of placing the emphasis on the simplicity of the underlying forms, some researchers point to the importance of recursive structures in generating complexity, the input of the observer in attributing life to the creatures, and the novelty and unpredictability of emergent phenomena. In these viewpoints, it matters that complexity disappears at the beginning of traditional scientific modeling and appears at the end of artificial life simulations. Whereas the complexity of the real world--which is to say, the messiness of embodied materiality --is left behind in the ( necessary and useful) analytical division of an environment into discrete components and the abstraction of form out of these components, complexity is precisely what is produced by the recursively structured adaptive systems of artificial life. Although this complexity is generated from simple elements, it is not reducible to their combined properties, nor it is predictable from them, for it emerges dynamically from their interactions. Which implies that if you have only the simple rules and transmigrating forms, the most interesting part of reality may have slipped through your fingers. When the emphasis falls on complexity, the effect is not to create symmetrically interlocking accounts in which the synthetic procedures of artificial life confirm the analytical insights of traditional science. Rather, such emphasis reminds us what is left out of account when embodied materiality is reduced to inscription.
Further skewing the mirror reflections locking realist and constructivist accounts into an oppositional dichotomy is the dynamic producing the virtual creatures. Recall that both realist and constructivist accounts use analogue relations that operate vertically and horizontally, respectively. In the computer, by contrast, the more complex oreo effect is at work. The embodied materiality of electronic polarities is analogically related to the ones and zeros of binary code, for the bits preserve the morphological proportionality of the material medium. Once the bits are produced, however, the dynamic changes from analogue representation to the scaled coding relations between different levels in a hierarchy of programs. As we have seen, an important effect of this arrangement is to permit much richer fragmentation and recombination than is possible with analogue relations. Because of this flexibility, it is possible to change the appearance of something--the font and color in which this text appears, for example--with a keystroke, for the coding relationships between hierarchical levels act like semiotic levers, leveraging the effects of a single command to produce large surface differences. Analogue proportionalities, by virtue of being proportionalities, are more limited in their leveraging possibilities. In scientific experiments, these limitations are typically overcome by abstracting the analogue proportionalities producing the inscriptions into abstract forms that migrate through inscriptions (the fluctuating membrane generates a trace on a graph, which then circulates as an inscription). With digital technologies, by contrast, transmigration is replaced by the oreo effect. After the analogue processes creating the bits comes the digital center, followed by analogue surfaces at the end of the coding hierarchies, for example when the functions defining morphology and behavior are visualized as virtual creatures. The embodied complexity which drops from the scene in the circulation of scientific inscriptions is not recovered in transmigration, thus facilitating the perception that the forms underlie physical reality and, in the idealist view, are reality. By contrast, the complex analogue surfaces that appear as the top layer in artificial life simulations (and other digital simulations such as PET images) evoke for the viewer the embodied complexities of action in a three-dimensional world. As I hope this discussion has made clear, analogue relations are unlikely to be displaced by digital fragmentation and recombination. Analogue proportionality and resemblance, whether at the bottom or top of the oreo, remain indispensable links with embodied materiality. As humans, we are embodied life forms who perceive the world in analogue terms, and our intuitions about the world, richly developed through eons of interactions with an immensely complex three-dimensional environment, operate most powerfully and acutely with analogue surfaces. It is no accident that digital technologies frequently culminate in analogue surfaces, for it is here that we are at our best as cognizers participating in these distributed cognitive systems.
Nevertheless, the digital center of the oreo--the dynamics of fragmentation
and recombination where digital cognizers are at their best--makes the
structure of these distributed cognitive systems operate in distinctively
different ways than traditional transmigration. These differences provide
a way to break out of the materialist/discursive dichotomy, for they complicate
the symmetries and suggest new ways in which to think about the relation
between embodied materiality and the abstract forms transmigrating through
inscriptions. Instead of materialist/discursive, they suggest such pairings
as analogue/digital, resemblance/code, linear/recursive, isolated/co-adaptive,
visualization/program, abstraction/emergence, complexity at material level/complexity
at symbolic level, behavior/function, forms that transmigrate/forms that
compute. In my view, it is crucially important that we understand these
pairings not as dichotomies in which the terms confront one another in
mutual exclusivity, but as dynamic dialectics in which the two terms interact
co-adaptively with one another. Life is not a choice. It is a continually
evolving complex adaptive system with many interacting cognizers, which
these days includes intelligent machines, intelligent programs, intelligent
environments, and intelligent humans. That's what virtual creatures can
 My thanks to Karl Sims for making the videotape available to Nicholas Gessler, and to Nicholas Gessler for allowing me to view it. I am also indebted to Nicholas Gessler for many of the ideas in this essay. He conveyed them to me over several years of on-going discussions, and by now they are so woven into my own thoughts that I am no longer sure which started as his suggestions. Let me simply say, then, that this essay could not have been written without his help.
 The transition from liberal humanist subject to the contemporary posthuman subject is described in N. Katherine Hayles, How We Became Posthuman: Virtual Bodies in Cybernetics, Literature, and Informatics (Chicago: University of Chicago Press, 1998).
 The world of "Evolved Virtual Creatures" is described in Karl Sims, "Evolving Virtual Creatures," ACM-0-89791-667-0/94/007/0015, presented at SIGGRAPH, Orlando FL, July 24-29, 1994. See also "Evolving 3-D Morphology and Behavior by Competition," unpublished manuscript.
 See John R. Koza, Genetic Programming: On the Programming of Computers by Means of Natural Selection (Cambridge: MIT Press, 1992) and Genetic Programming II: Automatic Discovery of Reusable Programs (Cambridge: MIT Press, 1994), where Koza shows how genetic programming can itself dynamically evolve the functions it needs to evolve further. Koza calls these Automatically Defined Functions (AFTs) and remarks, "Hierarchical organization and reuse [of ATFs] seem to be required if automatic programming is ever to be scaled up from small problems to large problems" (p. 4).
 Benoit Mandelbrot in The Fractal Beauty of Nature (San Francisco: W. H. Freeman, 1982) discusses how relatively simple computer algorithms can generate complex plant shapes.
 Michael G. Dyer, "Toward Synthesizing Artificial Neural Networks that Exhibit Cooperative Intelligent Behavior: Some Open Issues in Artificial Life," Artificial Life 1, no. 1/2 (Fall 1993/Winter 1994): 111-135. I am also indebted to Michael Dyer for conversations clarifying these issues for me.
 Alex Argyros, A Blessed Rage for Order: Deconstruction, Evolution, and Chaos (Ann Arbor: University of Michigan Press, 1991).
 Simon Baron-Cohen, Mindblindness: An Essay on Autism and Theory of Mind (Cambridge: MIT Press, 1995).
9 Joan Didion, The White Album (New York: Simon and Schuster, 1979).
 Stuart A. Kauffman, The Origins of Order: Self-Organization and Selection in Evolution (Oxford: Oxford University Press, 1993); John Koza, Genetic Programming and Genetic Programming II; Humberto Maturana and Francisco Varela, Autopoiesis and Cognition: The Realization of the Living, Boston Studies in the Philosophy of Science, vol. 42 (Dordrecht: D. Reidel, 1980).
 Luc Steels, "The Artificial life Roots of Artificial Intelligence," Artificial Life 1, no. 1/2 (Fall 1993/Winter 1994): 75-110.
12 A point made by Richard Dawkins in The Blind Watchmaker (New York: Norton, 1986).
 I use "see" in quotations marks to imply not only the physical act of visual perception but also the culturally-conditioned cognitive processes by which we invest what we see with meaning.
 Jerome Bruner, Acts of Meaning (Cambridge: Harvard University Press, 1990).
 Joan Lucariello is cited in Bruner, pp. 81-82 and in note 26, p. 157, where the source is identified as "private communication."
 Bruner, pp. 82-83.
 Marvin Minsky, The Society of Mind (New York: Simon and Schuster, 1986). Rodney A. Brooks, "Intelligence Without Representation," Artificial Intelligence 47 (1991): 139-59. See also Luc Steels and Rodney A. Brooks, editors, The Artificial Life Route to Artificial Intelligence: Building Embodied Situated Agents (Hillsdale: Erlbaum Associates, 1995).
 Marvin Minsky, "Why Computer Science Is the Most Important Thing That Has Happened to the Humanities in 5,000 Years" (public lecture, Nara, Japan, May 15, 1996). I am grateful to Nicholas Gessler for providing me with a transcript of this lecture.
 Michel Foucault, "What is an Author?" in Paul Rabinow, editor, The Foucault Reader, translated by Josué Harari (New York: Pantheon, 1984), p. 119.
 Mark Poster, "What's the Matter with the Internet," unpublished manuscript supplied courtesy of the author. I am grateful to Mark Poster for discussions clarifying his ideas in this manuscript.
 Poster's discussion can be clarified by noting that analogue technologies do not necessarily rely on resemblance, only on morphological proportionality. A phonograph record, for example, does not look like the sound waves it captures and reproduces, but there is a morphological proportionality between the sound waves and the spacings of the record grooves. Analogue relations which depend on resemblance are more properly called analogies (or analogizing relations). Analogies are a subset of analogue relations, which also contain the relations, typical of analogue technologies, that depend on morphological proportionality.
 Michel Foucault, The Order of Things: An Archeology of the Human Sciences (New York: Vintage Books, 1970).
 Mark Rose, Authors and Owners: The Invention of Copyright (Cambridge: Harvard University Press, 1993).
 This example comes from Poster, p. 20.
 This behavior was pointed out to me by Nicholas Gessler.
 The robot was displayed in an art show featuring images, simulations, and installations created by artificial life techniques, entitled "The Art and Aesthetics of Artificial Life," exhibited at the UCLA Center for Digital Arts June 22-July 23, curated by Nicholas Gessler.
 Bruno Latour, Science in Action: How to Follow Scientists and Engineers Through Society (Cambridge: Harvard University Press, 1987).
 Steven Shapin and Simon Schaffer, Leviathan and the Air-Pump: Hobbes, Boyle, and the Experimental Life (Princeton: Princeton University Press, 1985), "Replicating the Pump: London and Holland," pp. 235-244.
 H. M. Collins, Changing Order: Replication and Induction in Scientific Practice (London: Sage Publications, 1985).
 Bruno Latour, We Have Never Been Modern, translated by Catherine Porter (Cambridge: Harvard University Press, 1993).
 Christopher Langton, "Artificial Life" in Artificial Life, edited by Christopher Langton (Redwood City: Addison-Wesley, 1989), p. 1.
 Stefan Helmreich, a cultural anthropologist who spent several months at the Santa Fe Institute, in "Anthropology Inside and Outside the Looking-Glass Worlds of Artificial Life" (unpublished manuscript, 1994, supplied courtesy of the author) records the views of many artificial life researchers then working at the Santa Fe Institute. He discusses particularly those researchers who feel that computational techniques are in some deep sense natural because they reflect the mathematical nature of reality.
 Edward Fredkin, "Digital Mechanics: An Information Process Based on Reversible Cellular Automata," Physica D 45 (1990): 254-70.
 Helmreich also notes that some researchers focused on the observer. One respondent declared, "` [The life] is in the eye of the beholder. It's not the system, it's the observer,'" cited in Helmreich, p. 11.
 Nancy Cartwright in How the Laws of Physics Lie (Oxford: Oxford University Press, 1993) clearly analyzes the kinds of complexities that are left behind when a holistic interactive environment is divided into separate components.
 The idea of semiotic leverage through computer codes is developed in N. Katherine Hayles, "Virtual Bodies and Flickering Signifiers," October 66 (1993): 66-93.