Do Algorithms Have Fun On Completion, Indeterminacy and Autonomy in Computation

Luciana Parisi/Texts/Essays/Do Algorithms Have Fun On Completion, Indeterminacy and Autonomy in Computation.pdf

P. 1
4 Do Algorithms Have Fun? On Completion, Indeterminacy and Autonomy in Computation Luciana Parisi and M. Beatrice Fazi Algorithms, not androids In 1968 Philip K. Dick wondered whether androids, which at the time epitomized the ultimate mechanization of mental processes, could dream of electric sheep.1 Dick’s provocative suggestion was meant to address the prospect that mechanized thought could develop a human unconscious, and thus become truly autonomous. In Dick’s novel, dreaming stands as the ultimate expression of the human capacity to think beyond logical processing, without the functions of cognitive reflection or recognition. It could be argued that, in Dick’s story, a mechanized being that is able to dream realizes the promise of total autonomy from instructions and pre-programmed finalities. A dreaming android never becomes a human subject, but is nonetheless as close to the human emotional dimension of thinking as it could ever be. Today, nearly 50 years after Dick’s proposition, it is similarly possible to wonder whether algorithms – which are now the quintessence of mechanized thought – could exhibit a dimension of autonomous thinking. While androids had a body but were looking for a soul, in the present day mechanized thought is instantiated in computational processes that are themselves already material and incarnated, insofar as they are always performed and embedded in society, culture and the economy. To be autonomous, these computational systems do not need to simulate the more or less conscious requirements of a ‘psyche’, insofar as they already have a ‘point of view’ (or a sort of subjective dimension, so to speak) that is expressed through their material agency. Within the current configuration of digital technology, it could then be claimed that the mechanization of thought carried out by algorithms has approached the ultimate goal
P. 2
110 Fun and Software of total automation by incorporating thinking as material performativity, thus showing an immediate tendency to break free from the human master that has nonetheless programmed them. Following on these considerations, one could contend that, in computational science and in computational culture alike, the final goal for automated thought is no longer to escape the pre-programmed algorithmic rule and thus to revert into a sort of computational unconscious. Instead, automated thought is epitomized in a new vision of the algorithmic rule itself. Now this rule is always executed, and in this execution it can adjust itself to external changing conditions. To sum up, most algorithms have become interactive, and the mechanization of thought has to be accounted for vis-à-vis the now dominant expansion of such interactive approaches to computing. It is in this respect, one could argue, that today the quest for the autonomy of mechanized thought involves the prospect of mechanized systems changing their own pre-programmed regulations. This interactive vision of software has importantly questioned the self-sufficiency of axiomatics, logic and formalism, and has tried to complement the formal rule with the environmental input. Consequently, in the cultural analysis of these new modes of mechanization of thought, interactive software has been seen as operating affectively; that is, through the expansion of the field of potential activities that disrupt and re-invent meaning.2 Here, digital technology becomes the ground for creativity and for the production of paradoxical propositions that do not aim to reproduce a pre-constituted intelligence, established by a representational conception of thought. Rather, paradoxes and ambiguities strive to expose the sensible dimension of thought against the dominance of computational understandings of cognition.3 From this perspective, freeing the computational unconscious becomes secondary to the possibility of producing, in computation, sense beyond formal logic.4 It is exactly within this context, we believe, that the question of fun in software has acquired a significant momentum. Fun in computing emerges as an affective force that exceeds the formal logic that has nonetheless generated it; fun thus coincides with the disruption, caused by uncertainty, of the decisional power of algorithms operating within changing data structures. We may thus agree that the strength of this proposition is that it offers a radical account of the exceptional condition of instability and malleability of the computational rule – a condition that is vital to the creative drives of programming cultures. To a certain extent, the present chapter builds upon the rich debates that have put
P. 3
Do Algorithms Have Fun? 111 forward this proposition, and at the same time it aims to push this new view of computation in another direction. The question that we propose – ‘Do algorithms have fun?’ – is meant to echo the hypothesis posed by the title of Dick’s novel, Do Androids Dream of Electric Sheep? Like Dick’s provocation, our own hypothesis may also sound absurd; yet such an absurd proposition may nonetheless provide an opportunity to investigate another scenario regarding the significance of fun in software. To put it in other words, our proposition still aims to question computation, logic and formalism as representational approaches to thought. However, in order to do so, it is necessary for us to revisit the possibility that autonomy in computation might not be geared uniquely towards the liberation of the algorithmic procedure from the program or the master programmer. The question, we contend, is less whether a procedure might become free from mechanization, but whether mechanization itself already comes with a form of autonomy, insofar as it can be defined as a series of procedures aiming towards completion. As will become clear further below in the chapter, the notion of completion is used to define the process by which a final determination is achieved. Algorithmic autonomy for us is not about a machine sharing or opposing the human spectrum of cognition and perception in which the possibility of having fun is traditionally inscribed. On the contrary, the question, ‘Do algorithms have fun?’, is predicated upon a speculative move suggesting that fun corresponds to the capacity of algorithms to enjoy themselves while processing data. Enjoyment here is, in turn, mainly defined in terms of the final purpose of the computational process, which is to say, by its functionalist imperative to complete a task. In a counter-intuitive twist of the debate, we therefore take fun not to be the subversion of a rule, or the break from the constriction of a cold and strict procedure, rather we suggest that fun requires an end activity on behalf of the rule itself. Rather than escaping the rule, fun thus points to a new understanding of algorithmic order: one that complies with the question of the fulfilment of procedures, or achievement of a result, in computation. Fun in software is for us neither posthuman (there is no an alleged equivalence between human and machine), nor inhuman (there is no opposition between the two). Fun here deploys another level of mechanized processing altogether – a level that is not exclusively preoccupied with human–machine relations or the lack thereof. In this procedural order, relations are secondary to the finality of the rule’s function. Fun, we contend, is the final achievement of autonomy in mechanical thought.
P. 4
112 Fun and Software Forms of process The questions raised in this chapter are directed towards the speculative re-conceptualization of computational algorithms in terms of actualities, expressed as forms of process. We take the latter notion from the mathematician and philosopher Alfred N. Whitehead, who uses it to define the actual organization of data in the world.5 In our opinion, this Whiteheadian concept is key to the possibility of re-thinking what processes (and, therefore, also computational processes such as algorithms) are. Following the Whiteheadian view, processes can in fact be re-thought not as a continuous movement of variation, but instead in terms of the final achievement of a finite state. Whitehead employs the notion of ‘form of process’ to explain the ‘final mode of unity in virtue of which there exists stability of aim amid the multiple forms of potentiality, and in virtue of which there exists importance beyond the finite importance for the finite actuality’.6 In this respect, Whitehead’s ‘form of process’ needs to be investigated vis-à-vis a key element in the Whiteheadian philosophy: actual occasions. An actual occasion is actual not merely because it is something that is present here in front of us and impresses us (as Hume would have held actuality to be). In Whitehead, actuality is instead a process of concrescence, or ‘a process in which the universe of many things acquires an individual unity’.7 This is, for Whitehead, a process of satisfaction, or of selfenjoyment. In actual occasions, data are organized and unified; when this unity is achieved, the actual occasion is ‘satisfied’ because it has fulfilled its scope by becoming complete. It is our contention that Whitehead’s philosophy of actuality can contribute to questioning the antagonism between, on the one hand, the conception of process (intended as the continuity of variation, forming a whole that is bigger than its parts) and, on the other, the notion of processing (intended instead as a self-contained procedure based on already determined and finite parts aggregating into a whole). In Whitehead’s oeuvre, these two concepts are not opposed; it could be said, in fact, that both process and processing contribute to the determination of the actual occasion. We thus wish to appropriate Whitehead’s philosophy of actuality to go beyond this opposition which seems to pervade the critical debates about computational culture.8 To this end, we suggest that algorithms can be conceived as discrete actual entities, whose aim is to achieve unity. This achievement can be defined, again in Whiteheadian terms, as satisfaction (which is, in Whitehead’s words, ‘the completion of the
P. 5
Do Algorithms Have Fun? 113 actual togetherness of the discrete components’)9 or, from the perspective of this chapter, fun. On the one hand, an algorithm is a procedure that performs a processing, involving the execution of a sequential number of steps that organize data towards a result. This processing, however, does not simply lead to the aggregation of parts into a whole, but rather corresponds to the creation of a new unity that is added upon (and yet is not contained by) the aggregation of parts. Enjoyment, in this respect, involves the processing of data towards a final aim (this aim being the completion of the procedure). On the other hand, however, we want to stress that an algorithm is also a process, inasmuch as it involves an indeterminate capacity for variation. Yet, what should be emphasized here is that this variation is not the result of continuous change, but it is the final product of the process of determination carried out by the algorithmic procedure itself. This process of determination defines an algorithm that attains its completion by realizing its function. Satisfaction, from this perspective, is the exhaustion of the algorithmic function in the process. In computational culture, to sum up, we find evidence that process and processing are defined less by their differences than by a common finality, and that fun (re-worked speculatively via the Whiteheadian notions of enjoyment and satisfaction) might be an instance of such a finality. To carry out such re-articulation of algorithms in terms of forms of process – or, it is worth repeating here, in terms of discrete entities achieving satisfaction through the unification of data – we need to linger a little longer on the proposed parallel between algorithms and actual occasions. As already mentioned, Whitehead’s philosophy of actualities is useful for our investigation of algorithmic computation precisely because his approach is based on an atomistic conception of the universe that might counterintuitively help us to revise the conceptualization of digital procedures. Moreover, it can also help us to re-think what the unity of discrete entities might be from the standpoint of a process of determination. Instead of describing finite entities as already constituted unities, Whitehead argues that actual occasions are the result of a tendency of the multiple to become united. The accomplishment of unity marks for Whitehead the end point of an actual occasion. In this respect, actual occasions are finite and determined, also because they are short-lived. Once they reach unity, their process ends and they perish. One can also add here that completion is the finality of an actual occasion, insofar as its process of unification is its ultimate goal. The atomicity of an actual occasion corresponds to its finality: discreteness is not given, but produced from within the process. An
P. 6
114 Fun and Software actual occasion, therefore, does not come from nowhere. It inherits data from past actualities and is characterized by the tendency to eventually terminate just such a process of selection that ultimately constitutes it as one. Data, in turn, do not correspond to what already exists. For Whitehead, data are transmitted from the past to the present in a manner that is not simply re-used, but re-processed by the actual occasions under new conditions. It is important here to stress that for Whitehead this principle of inheritance cannot explain, on its own, the ‘becoming one’ of the actual occasion. Whitehead’s metaphysical proposition instead gives a major role to what he calls eternal objects. These are a priori yet immanent ideas, which according to Whitehead can be described ‘only in terms of [their] potentiality for “ingression” into the becoming of actual entities’.10 In this sense, an eternal object is pure potentiality from the standpoint of the actualities that select them. Eternal objects, then, cannot be described in terms of one potential plane of continual differentiation (for instance Deleuze and Guattari’s plane of immanence and/ or the virtual) or in terms of a plurality of equivalent ideas (such as the postmodern claim about the relativity of truth). Rather, this other element in Whitehead’s metaphysics offers us a notion of indeterminacy. Eternal objects are in fact non-connected among themselves, and only acquire togetherness when selected by actual occasions. In this togetherness, they concretize indeterminacy by allowing the actual occasion to achieve finality. Eternal objects are maximally unknowable, but nonetheless they inform the determinate state or the atomism of actual occasions. At this point we should add that both eternal objects and existing actual data are not passively received but are instead prehended by the actual occasion. Contra the empiricist perceptual capacity to synthesize the world through sense impressions, Whitehead argues that prehensions are modes of evaluation and selection, measuring and dividing an infinite variety of data each time. Instead of being just sensations, prehensions, for Whitehead, are both physical and conceptual. This is because each actual occasion ‘is essentially bipolar’;11 in other words it has a physical and mental pole, expressing the actual occasion’s capacity to physically and conceptually process data. The physically prehensive nature of an actual occasion means that actual occasions themselves are always informed by what is no longer there, by the data of the past actual occasions that are accumulated and crystallized in any present actuality. Moreover, as we have just discussed, actual occasions are also informed by what is not there at all: the unknown or indeterminate ideas or eternal objects. The prehensions of
P. 7
Do Algorithms Have Fun? 115 eternal objects are, for Whitehead, conceptual prehensions. As we saw earlier, eternal objects define the direction towards the determinateness of the actual occasion, its final unity and discreteness. Here again we find the processual nature of actual occasions, in which the indeterminacy of eternal objects adds a new level of determination to the data selected. A process becomes, however, a processing to the extent that data are compressed and arranged in a finite procedure, oriented by the actual occasion to drive towards its own satisfaction. In conclusion, prehensions, we could argue, are the catalysts for the concretization of the indeterminacy of the eternal object. More interestingly, however, they also reveal the persistent concreteness of indeterminacy itself in actuality. Completion and indeterminacy in algorithmic processing Having introduced these key elements and conceptual manoeuvres of Whitehead’s philosophy, we can now return to this chapter’s initial contention: the possibility of re-conceiving algorithms in terms of Whitehead’s actual occasions by drawing on his notion of ‘form of process’. In our view, to claim that an algorithm is an actual occasion implies arguing that the computational procedure is more than simply processing; i.e. more than a procedure based on pre-established instructions. In this respect, we want to argue that an algorithmic procedure cannot be thought of away from the indeterminacy that it is always tending towards. This passage is complicated, and to develop it we need first to explain what an algorithm in computation is, and what it is supposed to do. In formal sciences, an algorithm is described as a sequence of instructions meant to fulfil a task. An algorithm is a procedural method for the transformation of input into output; its instructions are well defined and executed through a finite number of inferential steps. This is the understanding of the algorithm that entered computing via the seminal work of mathematician Alan Turing. In his famous paper ‘On Computable Numbers, with an Application to the Entscheidungsproblem’ (1936)249 Turing sets out to answer David Hilbert’s question of whether, given a certain input, a universal method for taking a yes-no decision could be ultimately found in mathematics. In that paper, Turing rephrased the problem of decidability from the standpoint of computability. The notion of computability, in turn, was defined by Turing in terms of an effective procedure (i.e. an algorithm) to solve a problem in a finite number of sequential
P. 8
116 Fun and Software steps. In his attempt to answer Hilbert’s challenge, Turing developed a simple thought experiment (the Turing Machine) that was meant to mechanize a valid and determinate method for calculation. This hypothetical device performs computations by moving its head left and right along an infinite tape, that is in turn divided into discrete cells; the head moves one cell at a time, writing a symbol from a finite alphabet for each cell according to some given instructions. By proving that certain functions could not be computed by such a hypothetical machine, Turing demonstrated that there is not a method of ultimate decision of the kind that Hilbert had wished for. In addition to this, however, the Turing Machine also offered a viable formalization of a mechanical procedure. Computational processing, as we know it today, was born. The strength of Turing’s proposition is that, via his formalization of the Turing Machine, the function of computation came to be defined as being much more than just that of crunching numbers. Turing’s computing machines – and indeed the contemporary electronic digital computing machine that has developed from them – can solve problems, take decisions and fulfil tasks, with the only proviso that these problems, decisions and tasks are formalized through symbols and a set of discrete and finite sequential steps. In this respect, Turing’s effort can be rightfully inscribed into a long series of attempts in the history of thought geared towards the mechanization of reason. Turing’s foundational work in the theory of computability, and his conception of a computing machine, were meant to put the intuitive and informal operations of a person actually computing a number into formal terms. Mechanization implied that these formal terms could be automatically repeated because they were inscribed as symbols and discretized as a sequence of steps. At the technical level, we could argue that the Turing Machine epitomizes a mechanism that works insofar as it completes an operation. Moving now to the conceptual level, we can understand this operativity as one tending towards unity. An algorithmic procedure can in fact be seen as an aggregation of parts; when these parts achieve unity (that is, when the task is fulfilled), the mechanical process shuts down and then another procedure starts over again. Here one could object that, in light of what we have just described, the parallelism between an algorithm and an actual occasion becomes a difficult claim to sustain. It could in fact be argued that the Whiteheadian actual occasion does not merely aggregate into a bigger whole; rather, its parts are realized to become another part that is a new unity. In other words, in an actual occasion parts transform themselves in relation to a whole. However, the actual occasion remains,
P. 9
Do Algorithms Have Fun? 117 atomic: its parts never complete the universe of possibilities of what the actual occasion could be, or will be. An actual occasion is always a spatio-temporal relation between elements in process; it is indeed an occasion, its own eventuality determines its own prehensions, and, consequently, its own constitution. Continuing the objection, it could be also claimed that, in Turing’s formulation, an algorithm, conversely, has to do with pre-established functions (indeed, with axioms). In a Turing Machine, everything is pre-set, pre-programmed – in total indifference to content and context. From this standpoint, a Turing Machine merely performs processing: a linear procedure of cause and effect where one step inferentially determines the other. The objection according to which the parallelism between an algorithm and an actual occasion is flawed would thus seem to be justified: while an actual occasion is dynamic, an algorithm is not. Having identified these difficulties, might it still be possible to advance our intended parallel between algorithms and actual occasions? Our answer is yes. The parallel between algorithms and actual occasions is still valid by virtue of an inherent and unavoidable character of computation itself: incomputability. With his 1936 paper, Turing formalized the notion of computability; however, what is most striking about that proposition is that the formal foundations of this notion were laid on the logical discovery that some things cannot be computed. In 1931, Kurt Gödel proved the fundamental incompleteness of formal axiomatic systems.12 In 1936, Turing built on Gödel’s discovery to demonstrate that computation, in its mechanized formalization, is also intrinsically limited, and that these limitations are due to the formal axiomatic nature of computational mechanisms. Algorithms drive towards completion; this is of course evident at the level of their execution, but it can also be suggested that, at the formal level, completion is always entailed within the finality of the procedure. However, in his logico-mathematical formulation of computability, Turing foresaw that, despite the incredible versatility and efficiency of his proposed calculating machines, there are problems that cannot be solved, tasks that cannot be fulfilled and decisions that cannot be taken through them. These are incomputable problems for which there is no algorithmic decision. A Turing Machine can compute anything that can be calculated via algorithmic means. If something cannot be computed, then nor can it be put in algorithmic terms. For some functions, there is not a yes-no (or binary) answer. In this respect, completion in computation cannot always be attained. After this brief but crucial overview of the logico-mathematical origin of computation, the significance of the proposed parallel between algorithms and
P. 10
118 Fun and Software actual occasions might perhaps become more evident, and we can thus begin to justify our understanding of computational procedures in terms of forms of process. We want to argue that, while the unknown dimension of the actual occasion is the eternal object, unknowability in computation is expressed through the notion of the incomputable. We are thus here contending that the incomputable is to the algorithm what the eternal object is to the actual occasion. It is true to say that the algorithmic procedure is pre-established; it is undeniably a processing of data. However, this processing is an actuality that also has to confront the indeterminacy of the incomputable in the same way as the actual occasion is a finite spatio-temporality whose determination results from a process of prehension of the indeterminacy of eternal objects. In other words, eternal objects are ingredients in the constitution of actual occasions, and thus their indeterminacy becomes one with the finite atomistic nature of the universe. Similarly, the unknown condition of an algorithmic occasion is an incomputable condition, which serves the algorithmic actual occasion to speculatively become more than it was. The work of mathematician Gregory Chaitin can now be introduced to further expand on the contention that the incomputable is the eternal object in computational processing. In his algorithmic information theory, Chaitin combines Turing’s computability with Shannon’s information theory to understand computational processing in terms of calculation of probabilities – probabilities that aim to include what in this processing cannot be known in advance. Chaitin understands the threshold of computability from the standpoint of maximally unknowable probabilities.13 In every computational process, he explains, the output is always bigger than the input. For Chaitin, something happens in the computational processing of data, something that challenges the equivalence between input and output and thus the very idea that processing always leads to an already pre-programmed result. This something, according to Chaitin, is algorithmic randomness. The notion of algorithmic randomness implies that information cannot be compressed into a smaller program, insofar as an entropic transformation of data occurs between the input and the output of algorithmic processing, resulting in a tendency of these data to increase in size. From this standpoint, the output of the processing does not correspond to the input instructions; its volume tends, in fact, to become bigger than it was at the start of the computation. Chaitin has explained the discovery of algorithmic randomness in computational processing in terms of the incomputable: increasing yet unknown quantities of data that characterize information processing.
P. 11
Do Algorithms Have Fun? 119 Space constraints prevent us from expanding further on Chaitin’s work; the brief overview above should, however, suffice to sustain the claim that his work is indeed useful to us, inasmuch as it can contribute towards explaining what we mean by the ‘form of process’ in computation. Chaitin’s investigation of the incomputable reveals, in fact, that the linear order of sequential procedures (namely, what constitutes computational processing) shows an entropic tendency to add more parts to the existing aggregation of instructions established at the input. Since now this processing inevitably includes, from the perspective proposed by Chaitin, not only a transformation of existing data, but also the addition of new data on top of what was already pre-established in the computational procedure, we believe it becomes possible to argue for a form of process in computation. In other words, here processing is also a process. From this point of view, computational processing does not guarantee the return to initial conditions, and does not simply correspond to the aggregation or disaggregation of parts that can be pre-programmed. This is because Chaitin’s conception of incomputability no longer perfectly matches with the notion of the limit in computation (i.e. the limit to what is calculable). Chaitin’s incomputable involves, in fact, the addition of new and maximally unknowable parts to the whole; parts, however, that might be bigger than the aggregate whole. For us, such a reworking of the incomputable is striking and speculatively productive, because what was conceived to be the external limit of computation (i.e. the incomputable) has now become internalized in the sequential arrangement of algorithms (randomness works within algorithmic procedures). One can thus even suggest that algorithmic randomness is not ‘outside’, but has become constitutive of the actuality of the procedure. Processing then includes a form of process, because indeterminacy is a fundamental part of information. But if the incomputable involves parts that can be bigger than a whole, how do we explain the completion or the achievement of a task of an algorithmic procedure? Isn’t algorithmic randomness (i.e. the tendency of information to increase in size) always there to threaten any achievement of a function to complete a task? This is where, by Chaitin’s own admission, it is necessary to see algorithmic randomness as a continuation of Turing’s attempt to account for indeterminacy. Whereas for Turing some tasks cannot be achieved, and thus computation stops when the incomputable begins, for Chaitin, all tasks have a margin of incomputability. In Chaitin, incomputability does not break from completion. A closer investigation of his theories helps to reveal that algorithmic randomness has changed the nature of computational processing:
P. 12
120 Fun and Software completion has been expanded beyond its limits to involve the processing of maximally unknown parts in order to accomplish a task. To put it in other terms, completion is now demarcated by an indeterminate condition that subtends all efforts equating computational processing with the calculation of pre-programmed and already known outputs. In Chaitin, the finality of completing a task importantly requires that instructions respond and resolve unknowns each time the indeterminacy of data emerges in processing. Interactive procedures At this stage in our discussion, it may not be too difficult to justify the question of indeterminacy as being a speculative problem within computational processing itself, and thus point to the existence of a ‘philosophy of the algorithm’ that should attempt such speculations. We started our own speculative investigation by proposing a parallelism between Whitehead’s notion of actual occasions and the computational notion of completion. Having discussed incomputability via Turing and Chaitin, it became possible to think that the algorithmic completion of a task cannot be achieved without the participation of indeterminate quantities of data within processing. In particular, we have sought to develop Chaitin’s mathematical work on algorithmic randomness to prove that indeterminacy is always part and parcel of a determinate actual occasion/algorithm as it strives towards completion. From this perspective, we believe that completion in computation corresponds to unity, and unity is, in turn, a terminal point that can be reached only if indeterminacy ingresses the processing of data and makes the algorithmic procedure both a process and a processing. In other words, the ‘terminus’ of an algorithm can be approached only if the incomputable partakes in its procedure. Recapitulating, we can think of the incomputable as the indeterminacy that is necessary for completion, or for the unity of a form of process. However, our revisiting of computation through the problems of completion and indeterminacy cannot be fully grasped without discussing the dominance of a new paradigm within computing: the interactive paradigm. Arguably, interaction has become an all-encompassing principle of explanation in contemporary technoscience. Fields of enquiry such as biology, social theory, neuroscience and robotics have all pushed forward an alternative understanding of thought – understandings that see the latter as always already
P. 13
Do Algorithms Have Fun? 121 environmentally ‘situated’. Interaction is, in this respect, the interactive behaviour of multiple distributed agents. In computing, prediction and simulation now have to account for situations that are outside the confines of the black box and are instead always embedded into and dependent upon the environment. This interactive approach has clashed with the algorithmic constraints of the Turing Machine. Turing’s model of computation has been considered insufficient or unable to cope with the complexity of the empirical world – a complexity that, one could say, philosophically speaking, has its own non-representational logic. In this respect, efforts have been advanced in computer science to revisit Turing’s algorithmic modelling.14 While Turing’s conceptualization of mechanism based on a priori instructions can arguably be said to correspond to first-order cybernetics (due to its closed system of feedback), the combination of environmental inputs and a posteriori instructions proposed by the interactive paradigm more clearly embraces second-order cybernetics and its open-feedback mechanisms. Returning to our investigation of computational completion, we can claim that what we defined earlier as the algorithmic drive towards completion has perhaps changed now. In particular, we want to stress that if interaction is more powerful than algorithms,15 it is because indeterminacy, from the point of view of interaction, might now imply and represent something else. The volatility and malleability of lived situations are no longer obstacles for mechanisms of prediction. The goal of interaction is indeed to account for variation and novelty, and to enlarge the horizon of calculation to include qualitative factors as external variables of the mechanism. We want, thus, to suggest that, according to an ‘interactivist’ approach, the problem of the incomputable might be partly eluded. This is because completeness is no longer a solipsistic affair; rather, completeness becomes a prospect that is achievable only by virtue of the contribution of the outside world. In this sense, the algorithmic procedure might be incomplete per se. However, it reaches completeness by virtue of its interactive execution. From the standpoint of interaction, then, the successful running of an algorithm is a performance in the environment (i.e. computation is embedded in the world) and of the environment (i.e. computation needs the world and the data extracted from it to fulfil the algorithmic task). In our opinion, various approaches to interactive computing share a common goal of pointing towards a new mode of mechanization of a procedure, in which the starting condition of the program does not dictate the procedure’s final output. We can see this mode everywhere, from the imperative of participation in artworks, to social media strategies for user input geared to data-mining via
P. 14
122 Fun and Software neuro-cognitive mappings that now involve actions and perceptions. Taking into account such ubiquity and pervasiveness of computational interaction, it should be commented here that interactive computing is the result of the challenges that computation has to confront today: a quick and efficient responsiveness to vast data spaces, the quantification of desires, beliefs, inclination and knowledges that underpins the statistical calculation of trades and prices. We are keen to stress that, in this contemporary scenario of computational interactivity, completion is attained by adding new levels of quantification of variation. These new levels afford completeness insofar as the function of the algorithm is extended by external inputs, and is, thus, able to bypass its internal limits by simply posing the limit of computation elsewhere. The interactive paradigm, then, concerns the capacity of algorithms to respond and adapt to its external inputs. However, we want to suggest that this interactive form of adaptation does not serve to overcome, but only to postpone the threat of computational indeterminacy. In order to understand further how these new dynamic attributes of the interactive algorithm work, we can ask whether the interactive algorithm could be seen as being conceptually closer to Whitehead’s actual occasion. Is this discussion of interaction in computation a way to prove our parallel between the two notions? The answer is no: we believe that there is an inherent problem with the interactive paradigm’s reworking of indeterminacy aiming to calculate the variations of lived situations. Indeterminacy, we argue, is only approximated by the interactive algorithm; the problem of the incomputable is not solved but merely optimized through the addition of an external input. In this sense, the interactive algorithm harnesses the speculative power of the incomputable for computational processing. Going back to Chaitin, we can argue that the incomputable is not simply a limit that must be resolved by the addition of environmental variations, but that – as a problem – it remains an active ingredient or element of every computation, however many variables and data may be involved in it. Incomputability, like an eternal object, conditions the finality of the algorithmic occasion. We are thus saying the interactive paradigm is not so different from the Turing one: both are geared towards the fulfilment of a task. From both perspectives the pre-established procedure clashes with the unknown quantities of the incomputable. While in Turing computation stops, with the interactive algorithm the problem is bypassed by the addition of more and more data so as to enrich, diversify and vary the goal of the function. It could then be argued
P. 15
Do Algorithms Have Fun? 123 that the interactive paradigm interprets the incomputable from a human (or humanist) standpoint: it tries to solve the limit from the viewpoint of the human capacity to comprehend or compress data, or as how a human would cognitively cope with the unknown. Instead, our contention is that in fulfilling a task the algorithmic procedure inevitably confronts information indeterminacy, and by this encounter it fully realizes itself. This is completion from the standpoint of a computational process that cannot ‘not’ be incomplete. In other words, the limit of computation cannot be eluded, as it is intrinsic to computation itself. Computation may account for variations from the environment and thus become more powerful because of this inclusion. Yet the internal limit remains as the mark of the radical indeterminacy that constitutes computation. Satisfaction and the autonomy of rules The question of confronting indeterminacy and fully realizing oneself in relation to it needs to be explained by returning yet again to Whitehead. Whitehead, we said, calls the capacity of an actual occasion to constitute itself through the prehensions of data coming from other actual occasions and eternal objects ‘satisfaction’. An actual occasion reaches satisfaction as it reaches completion. For Whitehead, this condition of ‘exhaustion’ corresponds to the constitutive function of every actual occasion, and is not simply an option. Satisfaction should thus be interpreted not as emotional gratification, but as the final fulfilment of an appetite. In this sense, the actual occasion’s tendency to reach completion corresponds to its final determinations, which, however, as we have seen, can only be achieved through the ingression of indeterminacy in actuality. It is exactly at this point in our discussion that we would like to go back to the debate about fun in software. The long treatment of algorithms as actual occasions was in our opinion necessary to add an ulterior challenge to the conceptualization of fun in computational culture. Here, we return to our initial question: ‘Do algorithms have fun?’ So far we have proposed that we can articulate this question in the Whiteheadian terms of satisfaction and enjoyment, and, thus, in relation to the question of completion in computation. An algorithm that has fun is an algorithm that ‘enjoys’ its own process of determination. As anticipated earlier, in computation this process of determination is explicated in the processing of data on behalf of sequences of instructions. Drawing on Chaitin, we have said above that this processing involves a
P. 16
124 Fun and Software tendency of information to increase in size (i.e. algorithmic randomness). Via Chaitin, it is also possible to argue that the introduction of randomness, entropy or indeterminacy within computational processing describes algorithms that select a multiplicity of data which gets unified into an output. This output is a unity that is still a part, but that, however, can be bigger than the whole set of instructions from which the computation started. From this standpoint, and yet again following Whitehead, we can add that satisfaction can only correspond to the final expression of such constitutive indeterminacy within computational processing. Here, one can see that satisfaction, understood as algorithmic completion, might bring a further level of speculation to the debate about fun in software. From this perspective, there is fun in software because the algorithmic procedure entails a dimension of enjoyment that derives not from the breaking down of the mechanization, but instead from the fulfilment of the internal dynamic of completion via indeterminacy in computation. We do not deny that there is fun in going outside of the grains of formal logic, as the epitome of representational thought in computing. We maintain, however, that computational processing already implies a dimension of enjoyment or satisfaction – one that is being understood here as a process of completion determined by the ingression of indeterminacy within algorithmic procedures. Despite the fact that the process by which an algorithm enjoys itself might still remain obscure, we believe that this proposition offers an opportunity to address one of the main challenges that software poses. Crucial to the history of computation is the quest for an ultimate procedure for the mechanization of thought, and thus the possibility of finding a new form of conceptual function. To propose that algorithms are actual occasions means to confront both this quest and this possibility, and to reformulate what the mechanization of thought could be from the standpoint of the Whiteheadian notion of ‘form of process’. We want to argue that, in a Whiteheadian sense, the algorithmic mechanization of thought does not aim to establish an equivalence between emotional intelligence and the execution of a rule. By taking algorithms and not androids as our object of investigation, we can say that we are ultimately interested in re-directing the question of the autonomy of thought from emotion to reason. As mentioned earlier, we suggest that algorithms, taken as the contemporary epitome of automated thinking, reveal that the execution of rules (indeed, the mechanical) might have another order of autonomy – one that can be found not without or against but within reason. This is because algorithmic reason is already mechanical, insofar as it is the processing of rules.
P. 17
Do Algorithms Have Fun? 125 This last sentence should, however, be introduced with a disclaimer: we are not suggesting that algorithms exhaust all forms of thinking, nor are we asserting that all thought is determined by reason or by the processing of rules. We are rather making the case for a mechanization of thought that is already a form of reason. Algorithms might have an autonomy that is not to be understood as mere replication of what makes us autonomous (this is instead what Dick’s androids were aiming to do). Algorithms, we believe, have their own form of autonomy: one that pertains, we are keen to stress, to the final aim of executing rules (i.e. completion). The question of whether algorithms have fun becomes, from this broader reading of the mechanization of thought, an enquiry into instructions that operate by means of decision. It could be argued that our vision of fun in software shares, to a certain extent, some of the assumptions that motivated Dick’s critique of mechanical thought. His dreaming androids were meant to think beyond the rule, so as to deviate from what the traditional idea of mechanized thought was based on (i.e. the repetition of a rule and the constant return to initial conditions). Dick was thus exposing the limits of representation to contain the reality of thought. For Dick, a computational unconscious was a way to argue for non-representational thought, or for drives that cannot be reduced to symbols and steps, and which, thus, cannot be mechanized. In this respect, our contention that algorithmic fun needs to be understood in terms of computational completion similarly opposes the idea that rules represent thought. Ultimately, representation is for us, just as it is for Dick, insufficient to explain the reality of a process. The difference between our proposition and Dick’s, however, is that while his androids oppose rules to imagination by emphasizing the indeterminacy of dreaming and/or affective thought, our algorithms/actual occasions reveal that procedural and mechanical thought already contains much more indeterminacy than one could imagine. This indeterminacy is not simply readable via representational means, but remains nonetheless logical, formal and computational. This is the sense in which we understand the mechanization of thought as a rule-based mechanism that does not simply exclude or optimize indeterminacy, but instead cannot avoid expressing it. This is also to say that the incomputable, far from being the ultimate indeterminacy that eludes computation, is instead the indeterminate condition of every information system and, as such, it is constantly realized in algorithmic procedures. Our emphasis on the incomputable vis-à-vis Whitehead’s eternal objects therefore aims to put forward the
P. 18
126 Fun and Software idea that something has changed in the nature of the rule itself. We suggest that the computational rule is dynamic, but this dynamism is not uniquely derived from the interaction that the rule entertains with its outside when executed. In our view, the rule is instead dynamic by virtue of its processual function. Computational processing is thus also a process, even though it cannot be reduced to it. In conclusion, we suggest that computational processing can correspond to a non-human intelligibility of indeterminacy. From this perspective, algorithmic rules are what they are: instructions. As such, they are meant to fulfil a task. Yet, we believe that this instrumentality is not a symptom of an undefeatable power that programs our very human capacity of thought. Instead this instrumentality can be understood as defining an irreducible, and fundamentally algorithmic, form of process. In this respect, the fact that the rule accomplishes a function is what for us justifies the autonomy of such a function. To ask whether algorithms have fun is then a speculative exercise to test the hypothesis of the autonomy of algorithms as a new form of mechanical thought. Perhaps the question remains absurd, nevertheless it must be posed. Notes 1 2 3 4 Dick, Philip K. Do Androids Dream of Electric Sheep? (New York: Ballantine Books, 1996). See, among others, Munster, Anna Materializing New Media. Embodiment in Information Aesthetics (Hannover: Dartmouth College Press, 2006); Hansen, Mark B. N., New Philosophy for New Media (Cambridge, MA: MIT Press, 2004). Such understandings would instead take thought to be a mere aggregation of fixed rules. This occurs, for instance, in Putnam’s computational theory of mind: Putnam, Hilary Mathematics, Matter and Method (Cambridge: Cambridge University Press, 1979). The notion of ‘sense’ here is derived from Deleuze’s project to challenge the logic of representation from within. Deleuze, Gilles, The Logic of Sense (London: Continuum, 2004). Against the legacy of representational thought and theories of signification, Deleuze addresses and overcomes the linguistic turn in philosophy by proposing a ‘logic of sense’, according to which the genesis of value and significance is more important than truth and meaning conditions. For Deleuze, sense is neither original nor final, but an ontological event that is expressed in its effects – effects that are intensive, affective, material.
P. 19
Do Algorithms Have Fun? 5 6 7 8 9 10 11 12 13 14 15 127 See Whitehead, Alfred N. Modes of Thought (New York: The Free Press, 1968), in particular, refer to ‘Lecture Five’. Whitehead, Modes of Thought, 86. Whitehead, Alfred N., Process and Reality (New York: The Free Press, 1978), 211. Massumi’s elaboration of process and processing in terms of an opposition between the analog and the digital (both conceived as modes of articulating the real) has arguably influenced many juxtapositions of these notions in computational culture. For Massumi, analog processes are ontologically ‘superior’ to digital processing. The latter, for him, can never fully account for the virtual potential of thought and experience, a potential that is continuous (therefore, analog) and not discrete (or digital). See Massumi, Brian Parables for the Virtual. Movement, Affect, Sensation (Durham, NC and London: Duke University Press, 2002). See, specifically, ‘On the Superiority of the Analog’. Whitehead, Process and Reality, 85. Whitehead, Process and Reality, 23. Whitehead, Process and Reality, 108. Gödel, Kurt ‘On Formally Undecidable Propositions of the Principia Mathematica and Related Systems I’, in The Undecidable: Basic Papers on Undecidable Propositions, Unsolvable Problems and Computable Functions, ed. Martin Davis, trans. E. Mendelson (Mineola, NY: Dover Publications, 2004), 4–38. See Chaitin, Gregory, Meta Maths. The Quest for Omega (London: Atlantic Books, 2005); Chaitin, Gregory, Exploring Randomness (London: Springer-Verlag, 2001). For an overview of some of the key issues in interactive computing, refer to Goldin, Dina, Smolka, Scott A. and Wegner, Peter (eds) Interactive Computation: The New Paradigm (New York: Springer, 2006). This hypothesis has been advanced, for instance, in Wegner, Peter, ‘Why Interaction is More Powerful than Algorithms’, Communications of the ACM, 40(5) (1997): 80–91.