Large language models and knowledge in pieces as epistemological laboratories: insights from physics education research

This thesis is situated within the field of Physics Education Research and investigates the relationship between Large Language Models (LLMs) and Knowledge in Pieces (KiP) around the problem of knowledge. Rather than reducing the comparison to assimilation or opposition, the work constructs a theoretical space in which artificial and natural knowledge can be examined through shared dimensions while preserving their specific differences. The first part reconstructs the historical and epistemological genealogy of Artificial Intelligence, from early times to LLMs. It then analyzes Transformer-based models with regard to their specific architectural features. In parallel, the thesis reconstructs KiP within the broader horizon of Knowledge Analysis. The central contribution lies in the bidirectional comparison between LLMs and KiP. On the one hand, KiP provides a vocabulary for interpreting LLMs as granular, sub-symbolic and context-sensitive systems whose coherent performances emerge without explicit rule-based organization. On the other hand, LLMs make more visible some theoretical tensions internal to KiP, especially the distinction between local activation and stable understanding, the role of language, and grounded experience in human knowledge. The thesis argues that LLMs do not possess understanding in the human sense, since they lack embodied experience and direct access to the world; nevertheless, they may function as epistemological tools for deepening into the analysis of human knowledge. Finally, the thesis outlines future concrete research directions in which LLMs could support Knowledge Analysis, particularly with respect to prediction and bootstrapping, large-scale linguistic validation, learning timescales, runnability and boundary crossing beyond the domain of intuitive physics.