Simulation and the epistemology of transformer models: operational similarity through sustained difference

Abstract Extending simulation beyond its usual semiotic application, this article steps through the process by which Large Language Models (LLMs) reproduce similarity, to argue that simulation produces a “doubling” effect: an operational similarity sustained through difference. Accordingly, this article draws on literary theory to “read” transformer models with the intent of understanding how they operate as an infrastructure for the production of linguistic similarity. It is argued that neural networks enact an epistemology of exchange in which tokenisation, vectorisation, and self-attention render language commensurable, thereby producing an operational similarity that manifests differently in each natural language response generated. The political and epistemic implications of doubling are thereafter discussed with reference to the relationship between simulation, models and subjectivity.