Tokenized Rules of Inference as Physical Memes: From Metaphor to Mechanism

Since Dawkins introduced the concept of memes as cultural replicators, memetics has struggled with a fundamental problem: identifying the physical substrate of memes. Unlike genes, which have clear molecular instantiation, memes have remained frustratingly vague—patterns of neural activity, behaviors, ideas—without measurable, isolable physical form. This paper argues that transformer weight matrices in Large Language Models (LLMs) provide the first genuine candidate for physically instantiated memes. Specifically, when rules of inference—both valid and invalid—are tokenized, they exist in the exact representational space that transformers operate on. The weight configurations that reliably regenerate these tokenizable inference patterns constitute measurable, quantifiable information structures that replicate across training contexts. This moves memetics from evocative analogy to empirical science, enabling mechanistic study of how inference patterns compete, propagate, and evolve in computational substrates. Mathematical expressions are in marked down format.