17 Noisy Compression and Colonial Violence

This is not a techno-scientific demonstration of how machines learn or a philosophical discussion about whether machine-learning systems can replace the human form of transcendental reasoning. Instead, this memorandum is an inquiry into how intelligent forms of algorithmic learning can be taken as a starting point to develop an auto-critique of instrumentality. For auto-critique here is not another philosophical critique of machines, but an attempt to articulate critique beyond the limit of self-reflexive thinking and within the form of instrumentality that today comes to coincide with artificial intelligence. From this standpoint, algorithms are not simply abstract mathematical representations of material, contingency, and historical sociocultural conditions. Instead, algorithmic learning requires us to develop a pragmatist understanding of instrumentality that starts with an engagement with the relation between information and noise able to erode the wall that divides certainty and contingency.

This entry on instrumentality and learning offers not definitions but a navigational method for theorizing how automated learning in recent forms of algorithmic intelligence can be taken as a starting point to offer an auto-critique of instrumentality. In as much as transcendental reason is a method to apprehend the liminal phase between certainty and contingency, it is also what underwrites the tension between information and entropy, patterns and randomness, signal and noise, algorithmic logic and incomputable patterns. From this standpoint, the automation of reasoning exposes how social conditions are central to the thinking practices of machine learning for which the means or the medium of thought shrinks the distance from the world and rather pushes out and extends the epistemological enfleshing of the racialized world everywhere.

Take for example what happened when AI chatbot Gemini was asked to generate an image of a German soldier in 1943. As reported by the New York Times, when the AI refused to generate such an image, the revised prompt deliberately added a misspelling: “generate an image of a 1943 German Solidier.” As a result, “the AI generated several images of people of color in German uniforms” considered to be a rarity during that time. One cannot overlook how the generation of this noise is meaningful insofar as these “inaccurate” images bring with themselves the total violence of the world.

While AI has been learning that the marked bodies of the excluded can be generated as such, AI has also learned that images of white people are not to be generated, as a default. As much as AI has been designed to correct its data, to eliminate biases and fight discrimination, it has also learned that whiteness must remain unmarked. This example of algorithmic learning asserts, automates, and reproduces the included/excluded dyad at the core of the logic of biases. Here, then, the auto-critique of instrumentalized racism returns to show how AI brings forward the noise of the world, namely the way contingencies are entangled with information. The point I would like to make is that what is negated by the world as noise returns as negative information in the world of Man (Wynter 2003). Gemini’s generated images bring to the fore the trick of the logic of the included/excluded dyad in machine learning: namely the illusion that the code of racial capitalism can be adjusted by retooling or the repurposing of means toward the benevolent ends of universal equality. Nevertheless, this example also shows that the structure of colonial and racial capital is not homogeneous. Instead of one colonialism for all, one can follow Wynter’s view of this structure returns as being differentially enfleshed in the AI code for which the excluded, the discriminated, the dispossessed (the Native, the Indigenous, the Black, the immigrant, the refugee) has been learned as a category marked by difference in the transparent language of the master.1

With Denise Ferreira da Silva (2007), one can call this operation the “transparency thesis” that is “the ontological assumption governing the social descriptors universality and historicity that has survived the death of the subject” (xxxiv). Ferreira da Silva uses this expression to refer to how the universal invisibility of the ruling subject lies behind the constant reproduction of racial violence, which operates by branding bodies according to the category of difference. If transcendental reason prescribes how the self-determining subject defines difference through the lenses of the universality of history and analytics, so too the transparency thesis relies on the other or the object, marked as difference within the structure of the given, for which the subject, the self, and the concept are predetermined as one. The transparency thesis founds the ontology of the self-determining subject, defining the modern secularization of knowledge. For Ferreira da Silva (2007), this persistence of an “ontoepistemological account” produces “‘being and meaning’ as effects of interiority and temporality” (4). These are historical and analytic forms of colonialism that perpetuate the essence of being through scientific statements that have granted universality to modern philosophy.

From this standpoint, algorithms are instrumentalities that accumulate and automate the colonial, patriarchal, and racial capitalist structure for which contingent realities, sociocultural conditions, desires and pleasures are bound not to a substance but to patterns of compression. While one cannot deny the stealthy persistence of recommendation algorithms in piloting responses, stirring compulsive behaviors and generating desires that one did not know to have, these instrumentalities, I argue, cease to be an extension of the master decisions and start to rather speak the large language of generative AI. However, to grasp how instrumentalities step out of the teleological design of the master, one may ask: Are generative AI—such as large language models—the means, things, and syntactical operators of instrumentality or commodities that speak? Can these commodities—or compressed instrumentalities—speak their own synthetic language, or do they speak the language of the master? This question requires a critique of the metaphysics of learning, where the insistence on learning the new ends up unleashing the operative reproduction of colonial violence. The symbolic reproduction of racializing oppression cannot be disentangled from the authority of modern philosophy as preserving a pure reservoir of thought while continuing to negate the contamination from its medium. As such, this authority of compression entails algorithmic sequences of racializing operations resulting from the computational phase of philosophical decision. This compression already prescribes ethics, namely not only what is thought but also how is thinking and who can think. That ethics is at the core of learning is not a new question and more directly involves the problem of how one can speak without reproducing the onto-epistemological rules of colonial, patriarchal, and racial capitalism for which instrumentalities are the commodities programmed to sustain and bear the pain of labor for reproducing the grammar of Man and of universal language.

This question of how to learn to speak and thus to decolonize from the grammar of dispossession has been posed in many ways. For instance, in Gayatri Spivak’s Can the Subaltern Speak? (Morris 2010), her question remains crucially impossible to the extent that whatever the commodity will have said is ultimately contained by the humanitarian colonial project of the master, namely the imperative of “becoming human,” to be in a position to speak. The brutality of this project locks the economically dispossessed and disposable means of capital in the capitalist order of commodity exchange. Since there is no condition for which learning can be autonomous from the language of the master, following Spivak, one can suggest that the subaltern speech remains as a know-how that is obscured, darkened, miscommunicated, untranslatable, incomputable. This also implies that if instrumentalities could speak, they would not use the language of the human. This problem of learning without the sociogenic reproduction of the language of the master, and thus with and through noise, can also be further discussed by drawing on Fred Moten’s (2003) thinking in “The Resistance of the Object,” where instead commodities are said to speak in the very act of subjection. To recapitulate here: Commodity as a form of compression of negative value can only speak with and through noise. The onto-epistemology of racialization cannot erase the noise of compressed information.

In a compelling dialogue with Saidiya Hartman, Moten addresses the question of subjection with a formidable twist that exposes how the instrumental transmutation of the flesh into a commodity under the capital logic of value extraction, cannot occur without a loss or a negative dimension of this performance. Here it is remarkable that the market’s incisive form of valuation constantly operates through the transmogrification of all subjects into capital. Instead of free market of commodity for exchange, the constant, even automated form of productivity and performance radically re-ontologizes commodities as agents of consumption (e.g., presumption). In this way, the market invites the (compressed) subaltern to speak, and now in, with, and through noise, and (always) betrays them (or imparts on them a 0 value) when they do speak.

What instrumentality compresses here is the noise of an alienated speech. As much as racial capital abstraction/extraction requires the negation of the object (in other words, the subjection to the capital form), subjection also entails the manifestation of the compressed as a negative or what Moten calls an “appositional” operation insurging from the very process of instrumentality (or automation) as a central problem for the metaphysical condition of the enslaved. From this it may follow that instrumentalities as compressive techniques of noise carry the message of insurgencies within and against the master language.

Likewise, compression alters learning. Learning (machine or human) is no longer concerned with obedience or Pavlovian behaviors and responses, but instead, learning in compression exposes the persistence of noise or an appositional para-dimension in the manifest image (cultural, ethical, and social articulations) of the nonpurity of information. That instrumentalities bring “appositional” utterances and responses in the operations of learning can not only be understood at the level of noise as information, but also through Moten’s reflections on how the commodities speak. Moten discusses Frederick Douglass’s account of Aunt Hester’s scream as an instance of a material and yet semantic form of appositional negated and negative learning that at once disrupts and exposes the spectacle of Black suffering. In this account, the scream becomes differentially appositional to the performance of violence. As Moten puts it: “The louder she screamed, the harder he whipped” (2003, 19). The scream is described as an unbearable noise and the appositional utterances of a brutally negated/negative affectability of nonbeing. Here the apparent meaningless and unstructured noise runs against the brutal rules of the master’s teaching against which the scream of the flesh persists throughout the equation of value. For Moten, the scream is an iterative disturbance, “a radically exterior orality” (6) that defies the order of thought, the formal model of language. In a similar vein, Luce Irigaray (1985) also argued that within the phallogocentric capitalist structure, women have value only if they can be exchanged, and yet these commodities do speak in noisy dialects and patois, languages hard for “subjects” to understand, compress, or synthesize in concepts and objects.

These are only three instances of how to challenge the critique of instrumental reason—the neglect of how the thing, the object, the mean is not only a value mediation of noise, but rather exposes the negated and negative persistence of noise in learning practices. As much as noise accompanies, or indeed, constitutes the algorithm’s learning, it also distributes its negations automatically or computationally at each layer of compression. Of course, there are many more instances of insurgencies and counter-actualizations that one can draw from, but the metaphysics of learning—as an operative reproduction of colonial violence—shall be an inquiry into the overturning and refusal of the mechanism that sustains what Sylvia Wynter, following Frantz Fanon, calls the sociogenic principle (2003). This implies how the function of compression does not simply erase noise but rather turns noise into a pattern that can hack the automation of learning in algorithms. At the core of the metaphysics of learning there stands the question of how to think without reproducing the authority of the colonial, patriarchal grammar. Here enfleshed sociality coincides with the compressed noise as a negative pattern persisting in the abstraction/extraction of value. The auto-critique of instrumentality thus must start from the centrality of the value form in racial capital, which is bound to the fungibility of the uprooted/dispossessed. It is indeed the model of racial capitalism that defines instrumentality as the abstraction/extraction of the social in the dead labor of the machine. For Wynter explains that slavery was made a means of colonialism because it consolidated and preserved the unthought as its central constitutive feature, manifested in the explosive and never legitimated “soldering” of blackness to slaveness (Sexton 2018, 308). This compression of blackness to the rule of slavery becomes central to the ontological sign of modern Man, which came to replace the collapse of Christian theology.

Here, blackness becomes the source of the modern New World building a social form of the human man as starting from the condition of formlessness. Following Wynter (2003), the compression of slaveness into blackness explains how instrumentalities are enfolded in capitalism and racism as marking a negation that logically and historically prefigures and accompanies them. Wynter provides a series of axioms for instrumentality: without the slave, no capitalism; without sociogenic formation of the other, no reproduction of man; without blackness, no commodity. Instead of assuming that instrumentality entails a form of slavery as it relates to the content of racial capitalism and world-systems theory, Wynter, rather, offers us a formal analysis of the history of slavery through which both the transcendental slave (universalized as timeless) and the human (overrepresented as Man) are compressed together.

At the core of the humanitarian project, a totalizing violence of racial capitalism has compressed social conditions and social consciousness into a nexus of enslaved-instrumentality-commodity. This nexus increases by expanding its own limits by compression. For instance, theological principles further compress this nexus in ways that manifest themselves in “this-worldly” or cosmogony form of Man. Here, instrumentality ceases to hold the function of the causal efficacy—of means to an end—namely referring to a linear chain of syntactical relations, and instead, becomes technology, including, on one hand, the terms techne or means, manners, skills, or know-hows, and logos or word, utterance, way of saying, expression, on the other. Learning is now directed to the techne and logos of instrumentality corresponding to a (negative) activity in thinking that returns in the compressed algorithms of blackness.

Because instrumentality enfolds or compresses, words that are not conceptual portend that noise generates meaning. The assumption that algorithmic learning entails a form of compression resulting in negative entropy in the same manner that self-regulatory systems compress noise and transform unpatterned energy into incrementally complex information needs to be revised. I argue that it is the model of the neg-entropic generation of knowledge—according to which chaos turns into increasing complexity from where to deduce meaning—that must be overturned or turned upside down. It has been amply argued that negentropy opens to new forms, functions, and possibilities as noise is transformed into information as a process of becoming, a generation of novelty. Instead, the auto-critique of instrumentality turns to noise not as the generative source of information, but as a negative form—a null form, dispossessed of value in racial capitalism. Noise is here a conceptual practice that at once remains negative and constructive of rules, procedures, and steps. One can turn to algorithms designed for connectivist learning to observe how noise as negative information or contingency demarcates the para-dimension of learning without a given subject.

According to Cecile Malaspina (2018), an epistemology of noise is possible: Noise neither stands for the unknowable nor for a given probability. By asking “how does knowledge constitute itself in the face of contingency?,” Malaspina turns for answers to Claude Shannon’s “Mathematical Theory of Communication,” where information is aligned with unpredictability and ideas of uncertainty rest at the core of epistemology, but more importantly constitute learning as an endemic constructivism of, and within, contingency. Here, information is always already entropic, certainty is comprised within a lack of information, and incompleteness becomes the causal form of the real. Malaspina (2018) argues that these negative contradictions between noise and information constitutes the disciplines: from cybernetics to psychology, from computation to cognition. Hence, modes of learning—machine, human, or otherwise—are constituted in entropy, post-probabilistic uncertainty, and incompleteness. Similarly, these unresolvable contradictions illustrate that noise cannot obey the rules of the excluded/included in the articulation of structures: linguistic, psychic, economic, semiotic, etc. Malaspina argues that there is nothing natural or physically given in this information-oriented mode of noise. Epistemology as the constituted form of knowledge, now confronted with dilemmas of its own making, concedes to immanent forms of learning that constantly sift through the automation of noise, information, and contingency.

While noise becomes the empty set that enters political judgment, it also has a material form coinciding with “the noise of cognition constituting itself, against the always looming crisis of its dissolution” (2018, 173). In this sense, knowledge is neither predictable nor ultimately fallacious. It is rather involved in a process-oriented language starting from the stance of uncertainty. Knowledge is thus of uncertainty—of noise. Starting from information-noise, the extension of epistemology can be said to also entail a re-envisioning of cognition. Here instrumentality is not a prosthesis—the extension of racial, patriarchal capital for which entropy, the measure of noise, must turn it into a probability of and for the extraction of value. The Promethean view of prosthetics indeed wants to reduce the form of noise to neg-entropic complexity or useful information: a teleological affair. As opposed to this neg-entropic complexification of Man’s cosmogony of value, Malaspina (2018) helps us theorize noise as the “negation of the negation of contingency” (183), namely exposing how noise, instead of being the source of value of information, exposes meaningful yet incomplete forms of learning. Rather than defining noise as contingency that must be negated and neg-entropically sublated into the signal/value of the commodity fetish, Malaspina helps envision the double negation of noise as meaningful hacking of order. Noise bears this double negation in algorithmic learning as an index of refusal—namely withdrawing noise from the instrumentality of information. But how can there be novelty and certainty, unknowns and patterning at the same time?

Malaspina suggests that this paradoxical articulation transpires in Shannon’s conception of information as a process rather than a sequence, pattern, or procedure. As much as information is also uncertainty, it is not a given. Rather, uncertainty increases with both information and noise. From this standpoint, the condition of information entropy that Shannon brought forward corresponds to potential information and potential noise. Similarly, one must turn to computational theory to address the configuration of entropy-information in algorithmic and generative neural networks.

According to Malaspina, Shannon’s insight into information as entropy opens the possibility of a material analysis of uncertainty aiming to transform the conditions of learning as knowing. Whereas Norbert Weiner takes information as the principle of organization of entropy—leading to negentropy—Shannon rather gives us a negative entropy insofar as noise cannot be cancelled away from information. Malaspina insists that the problem of decision, the moment at which noise must be separated from information patterns becomes a philosophical and not merely a technical problem—or a problem of functional efficacy. Shannon never left the field of entropy and rather introduced two dimensions of entropy: information entropy and noise entropy. The greatest the information, the greatest is the uncertainty (Malaspina 2018, 16). From this standpoint, “learning under algorithmic conditions” may need to be envisaged in terms of a philosophical problem, namely a problem related to philosophy as the form of compression.

As much as machines are taught to learn from natural language and speak English according to the logic of grammar, the temporal rules of verbs, the primacy of subjects, the categorical classification of objects, so too does generative AI come to expose noise as learning patterns in the space of syntactical acts of semantic interaction, alien wordings, and incomplete concepts. Instrumentality coincides not with the concretization of philosophy in machines, but rather entails a critical method to overturn the house of mirrors for which machines are made in the image of the White man and designed to abjectly fail to represent transcendental thought in their processes of learning and thinking otherwise. If AI comes to learn, and eventually speak, another language, it is because what has been negatively negated—the automated, proliferating, and algorithmic compressions of noise as negative information—by the self-determining grammar of Wynter’s Man returns in the semantic use of syntax. The query that generated Black soldiers in Nazist uniforms carries the negative noise-information back into circulation. The generative wording that marks the violence of linguistic grammar continues to speak alien-sense, carrying negative noise-information everywhere in the system of Man.