Research programs: Computational Communication & Culture, Digital Publics
Abstract:
This thesis employs the concept of “doubling” to analyse the process of simulation as it occurs in Large Language Models (LLMs). Doubling is an effect of simulation that reveals a constitutive paradox: there must always be something missing from a simulation to prevent it from being the exact same as the entity being simulated. Doubling emerges from this paradox, and accordingly I define it as a functional act of resemblance that is sustained through difference. For instance, the output of an LLM must at once resemble and differ from both its training data and the prompt that activates the process of dialogue generation. I argue that this doubling effect is inescapable for LLMs, and that attempts to circumvent it are symptomatic of a culture immersed by the idea of a complete computational rationalisation of the world that renders events, objects, and people entirely predictable.
Analytically, the work of this thesis is directed toward understanding what is lost or what escapes in the process of simulation. I trace this constitutive gap across three sites where it produces observable effects: in the technical operation of the transformer architecture, in the design and operation of the user interface, and in corporate discourse around autonomous systems that promise hyper-intelligent, predictive machines. At each level, doubling operates differently. That is, doubling emerges in the vector spaces within the transformer architecture, as an experiential effect of interface design, and as an ideological tension in claims about AI capabilities. Yet across all three levels, the gap between simulation and simulated reveals the limit of computational representation and reinforces the elusiveness of human language, and subjectivity.
Drawing on Continental philosophy and computer science, this thesis offers a novel account of how simulation operates across computation, communication, and ideology in Generative AI. Through close readings of technical documentation, a walkthrough analysis of the creative writing interface of the creative writing application Sudowrite, and a trade press analysis of corporate discourse around LLMs, I demonstrate that recognizing simulation’s constitutive limits challenges the assumption that language and meaning can be fully formalised, predicted, and reproduced through computation alone.
Principal Supervisor
Associate Supervisors
