In recent weeks I’ve been working on a paper inspired by Offert et al.’s “Synthesizing Proteins on the Graphics Card: Protein Folding and the Limits of Critical AI Studies,” which I wrote about in my last post, alongside another piece from Offert, “Can We Read Neural Networks? Epistemic Implications of Two Historical Computer Science Papers”. The latter is in American Literature, which is paywalled (I can’t even grab it through my institution; I only have a copy thanks to a scholarship that provides access to NYU’s digital holdings). It’s fascinating, and I’m happy to send it along to anyone who asks, along with the two papers referenced in the title.
I’ll be presenting an early version of this paper at The 3rd International Conference on the Ethics of Artificial Intelligence, which is taking place online and in the Autonomous Region of the Azores, Portugal next month. My talk will be online, on either the 19th or 20th of September.
The paper’s working title is “Wrong Answers Only: Transformer Architectures and the Evacuation of the Negative.” Here are two versions of the abstract, brief and extended:
***
Brief abstract
This paper addresses the epistemic implications of the rise of transformer architectures (transformers) in neural network-based artificial intelligence. Transformers are perhaps best known for their use in natural language processing (NLP) applications. Their theoretical deployment as “universal approximators” which can execute any sequential computing function has been invoked as proof of the possibility that reality is fundamentally structured as a kind of language. By this rationale, transformers’ original success in the context of NLP counts as a posteriori justification for the premise that “all the world is language,” so long as their capacity for universal approximation — thus far only borne out in theory — is not disproven.
I begin by reviewing literature which challenges this view and additionally suggests that while transformers fail to prove that all is language, they produce a new and distinct class of knowledge. This novel epistemological territory encompasses everything that can be expressed in natural language and computable numbers, and it exists independent of programmatic instantiations. From there, I consider whether the advent of transformer-based computing and its concomitant epistemic turn amounts to a final triumph of epistemology over other philosophical lenses. I argue that a confound to this possibility lies in the fact that transformers, as universal approximators, cannot “fail” or “negate” in any absolute sense. I conclude by indicating avenues for further research in metaphysics and aesthetics.
Extended abstract
In their article “Synthesizing Proteins on the Graphics Card: Protein Folding and the Limits of Critical AI Studies,” Fabian Offert, Paul Kim, and Qiaoyu Cai argue that transformer architectures (transformers) are structurally non-linguistic despite their successful use in natural language processing (NLP) programs. On their account, transformers introduce a simulated mathematical continuity to data “that could not be more language-unlike,” as they write. Consequently, transformers confound “the intuitive (humanist, and implicitly structuralist) interpretation that the many capabilities of contemporary artificial intelligence systems are a result of the fact that ‘all the world is language.’” As a case study, Offert et al. explore how transformers are used in protein-folding applications. Their account of AlphaFold, an AI program that predicts how protein structures will unfold, highlights the technological root of transformers’ broad range of applicability. The authors demonstrate that the simulation of continuity is exactly what allows for the same architecture to be used for purposes ranging from the “cultural” — e.g., large language model programs like ChatGPT — to the “scientific” — e.g., AlphaFold — and beyond. This seeming universality does not mean that “all the world is language” for empirical reasons: as they observe, “amino acids are not letters and amino acid sequences are not sentences.” Indeed, digital representations of language and digital representations of protein structures are only alike in that both comprise sequential parts, which they argue is too weak a likeness to ground claims regarding the language-like nature of reality.
As an alternative, they argue that transformer architectures disclose “a new class of knowledge” which does not privilege the structural-linguistic and which exists independent of particular applications.This novel epistemology is the outcome of what Offert describes as the rise of “postsymbolic computing,” his term for the nascent widespread deposition of transformers that emulate “the fundamental building blocks of computation” in neural networks. In the central event of this shift, all forms of computing would stand to be relegated to neural networks operating as universal emulation machines. To approach the epistemological territory of the postsymbolic turn, he argues that critical AI studies must transcend its current “cultural” orientation which privileges language-based knowledge. This is because “nothing [in particular] is language” in this new territory, including language itself — which also means that “everything is language.” More precisely, when transformer architectures are treated as universal emulators, the hierarchical distinctions between layers of computational abstraction vanish. Here, natural language would be no less “basic” or fundamental to computation than what is traditionally understood as relatively less refined symbols of digital representation, such as binary zeros and ones. In this instance, coding may be supplanted by the production of natural language with none of the loss typically associated with abstraction.
My paper considers whether the advent of postsymbolic computing amounts to a triumph of the epistemological, vis-à-vis the informational, over other areas of philosophical study. After addressing philosophical debates on the relationship between the informational and the epistemic, I turn to a possibility raised by the representative capacities of universal approximation, which is that all forms of apprehensible knowledge might be perfectly replicable as computable information. As I argue, this suggests that the new epistemological territory of postsymbolic computing is superior to models provided by, e.g., ontology, ethics, aesthetics, and political philosophy, among other subdisciplines of philosophy and the theoretical humanities and social sciences.
I argue that one challenge to this possibility lies in the fact that transformers cannot “fail” in any absolute sense. The outputs of neural networks may be relatively far from a desired target, but they always exist within the scope of the digitally approximable, which includes all manifestations of correctness and their possible negations. Transformers thus relativize both correctness and incorrectness on the basis of two assumptions: first, that the former need not be absolute or self-identical; second, that the latter is never desirable and may therefore be relativized to no consequence. As they define correctness as nearness to given inputs, transformer architectures reduce incorrectness and negation to forces which only exist to affirm the relative value of the correct. This reduction serves as an arch-framework within which all forms of correctness and incorrectness (such as ethical correctness) become relative informational-epistemic positives, and negativity in itself ceases to exist. I conclude by suggesting aesthetics and metaphysics as areas of study where negativity in itself stands to remain both apprehensible and irreducible to epistemology in the age of postsymbolic computing.
***
As I move ahead with this project, there are a few issues I need to consider. First, is it really fair to posit a perfect correspondence between the epistemological and the computational-informatic, that is, to so casually write of “the epistemological vis-à-vis the informational?” I’ll do my best to address this problem, but I don’t have every move plotted out just yet.
The second issue is about the notion that “all forms of apprehensible knowledge might be perfectly replicable as computable information” and its implication that “that the new epistemological territory of postsymbolic computing is superior to models provided by, e.g., ontology, ethics, aesthetics, and political philosophy, among other subdisciplines of philosophy and the theoretical humanities and social sciences.” In the abstract, I don’t explain what it means to be “superior” to ontology, ethics, etc., and so I might be seen as wrongly presupposing a certain universal explanatory function to theoretical models from the humanities (especially including but not limited to philosophy) and social sciences. After all, these approaches may just as well be used not for the purpose of orienting thought towards the true and real, but to guide it in any number of other directions, such as the ethically good, the politically expedient, or the aesthetically sublime. I will need to clear this up, but for now I’ll just say that my concern isn’t so much that the alleged triumph of epistemology would be a “real” triumph, but rather that it would be a triumph in the sense that it intimates a lack of substance or reality to that which is more clearly illuminated through those other approaches (and which in many cases remains intractable to computation, even postsymbolic computation).
As part of my research, I’ve been reading Anna Kornbluh’s new book Immediacy, or The Style of Too Late Capitalism. Immediacy presents several arguments related to the central thesis that various aspects of contemporary culture — “high culture,” “low culture” and everything in between, although she gives credence to the idea that those distinctions haven’t mattered for a long while — should be understood as absorbed in a sort of stylistic anti-logic best glossed as “immediacy.” The style of immediacy forecloses slowness, deliberation, representation, and mediation in favor of the direct, visceral, and unrepresentable. According to Kornbluh, it’s both cause and effect of the characteristic late-capitalist condition of epidemic levels of anhedonia and nihilism.
In a chapter titled “Imaginary,” she argues that immediacy style deals death blows to the Lacanian realm of the symbolic, which she describes in sufficient enough detail that a non-Lacan scholar like myself can understand it:
Ordinarily — or, in less image-abundant times — the imaginary’s one-to-one equation is complicated in the symbolic, the realm of the signifier. The symbolic spans spoken and written language, with all of their detours, confusions, and metaphors; language as a system; institutions, laws, norms not of the individual’s making. The symbolic stopples the dyadic mirror liquidity of body and ego / I and other-I, introducing instead lack, inconsistency, and negation. Signifiers extrude flatness into volume, scaffolding the space of the social in greater amplitude than the plane of the image. The symbolic girds more than the identities of the imaginary: it is the terrain of alterity. Such alterity might assert itself as a kind of thirdness that disrupts the imaginary duo, auguring not just self and other but the other of the other. It might inscribe itself as the grammar or norms that structure the exchange between I and imaging, belying the nonverbal fluidity of one-to-one infantile relay. And it might make itself legible, in parapraxes and other symptoms, as the unconscious, the other side of imaginary self-consciousness. With its alterizing activation of the unconscious, contradiction, and lack, the symbolic evokes absence in contrast to imaginary presence, unpleasure in contrast to pleasure, death in contrast to life. (Immediacy, or The Style of Too-Late Capitalism, 50-51)
I think that there’s a continuity between this construction of post-symbolism — which seems to pertain primarily to politics, culture, and the psyche — and the post-symbolic computing whose epistemological terra incognita is just beginning to materialize, at least in Fabian Offert’s view. In the epistemological plane of post-symbolic computing, language fails to function as a system of irreducible signifiers (I explore this break from irreducibility towards the end of my last post, using the example of subword tokenization in AI transformer architectures). Put crudely, abstraction in computing seems to be under threat by AI, and I think it’s significant that this threat is contemporaneous with the dissolution of self-reflexive abstraction in human thought. Here, I’m following Kornbluh’s (and many other scholars’) understanding of abstraction as the mental act which affirms otherness (the other of other people; the other of society and culture as entities which do not always perfectly mirror the self) in its performance of vital mediating functions.
The working title for my paper is “Wrong Answers Only.” It’s a jokey reference to the verbal meme, and I think the reference fits because the paper is largely concerned with the transformation of the negative into relative positives. In the procedures of transformer architectures, all counterfactuals, contradictions, and other instances of negation are presupposed as immanent and are only directly invoked to measure relative distances to truth. In the same sense, one can never give an intrinsically wrong answer to a call for “wrong answers only:” all responses are either correct in the sense that they’re intrinsically wrong (2 + 2 = 5) and thus meet the criterion of the call, or they’re correct in the sense that they’re intrinsically correct (2 + 2=4), in which case their “wrongness” is nothing but an extrinsic attribute (they fail to be a “wrong answer”). Negation/wrongness in itself has no real substance or existence in “wrong answers only” epistemology, because it cannot assimilate that which is intrinsically negative, wrong, or other.
Per Kornbluh and other critical theorists (like Byung Chul Han and Theodor Adorno), negativity in itself is a prerequisite to any real encounters with otherness and to the construction and inhabitation of worlds not fully determined by the atomized and isolated self. Negative space affords points of distinction and orientation, and it constitutes the fertile ground from which ordering systems can emerge and take particular forms. In the epilogue to my doctoral dissertation, I suggested that the “dissecting table” envisioned in a famous line from Comte de Lautréamont’s Les Chants de Maldoror — a young boy is described as being “as beautiful as a chance meeting on a dissecting table of a sewing machine and an umbrella” — represents this negative space. The aesthetic effect of the juxtaposition can’t obtain without the table, because the table is the stable territory and space on which the sewing machine and the umbrella may be distanced from one another rather than collapsed into an undifferentiated mass.
So, I’m preparing to defend an argument that negativity in itself is under siege thanks to transformer architectures, and perhaps also thanks to “too late capitalism” — that is, if my instincts about the continuity between post-symbolic computing (per Offert) and the disintegration of the Lacanian symbolic realm (per Kornbluh) are correct. The “so what” of the argument is that negativity in itself furnishes human life with a sort of meaning that evades capture by the merely epistemological but that appears as increasingly, evanescent, weightless, and discountable in the age of digital capitalism and AI (which, of course, have a symbiotic relationship). As a brief aside, I’m planning to incorporate Wilfred Sellars’ notion of the manifest image of human life here. The manifest image, which he presents as anathema to theory, is purely correlative and natively available to perception, and it’s out to kill theoretical or “deep” constructions of human life from various intellectual domains, including theory-informed “hard” science.
As with the earlier two posts, this is an overture to something that won’t be on Substack. I hope it goes somewhere, anyway (beyond the conference, even). I am committed to articulating the importance of negativity in itself in such a way that underscores its current relevance. If for no other reason than this work can’t just be left to footnotes on Hegel and Deleuze.