How Language Models Work: A Multi-Register Investigation

A language model tokenises text, computes attention-weighted representations through 96 layers of transformation, and generates output one token at a time by selecting the most probable continuation from distributions learned through gradient descent on hundreds of billions of tokens of human writing. The result: systems that write legal briefs, debug code, translate between languages, sustain philosophical argument, and pass professional examinations, without understanding a single word they produce. The engineering is fully documented. The mathematics is precise. The training procedures are reproducible. An engineer can trace every operation from input to output. And yet the engineers who build these systems cannot explain why the conjunction of next-token prediction and sufficient scale produces outputs that pass for comprehension, reasoning, and creativity. The how is known. The why is not. The two questions have never been addressed together, in the same investigation, at the level of precision that each demands. This paper addresses both through four registers. Part I presents the engineering through concrete explanation, analogy, and step-by-step description, for the reader with no technical background. Part II presents the same architecture at mathematical precision: the attention equation, the training objective, the scaling laws, the sampling strategies, for the engineer who wants to verify every claim against the published literature. Part III conducts a philosophical investigation into why the engineering works, testing six frameworks in the philosophy of language (Saussure, Derrida, Wittgenstein, Quine, Austin, Davidson) against the empirical fact of the models' success. Part IV draws out the operational implications for the claims currently being made about these systems: the scaling hypothesis, the nature of hallucination, the "cognitive abundance" thesis, and the question of what these systems can and cannot become. Each register is self-sufficient. A reader may enter at any point. The cumulative finding: language produces distributional regularities that are learnable by exposure and resistant to formalisation. The models succeed by approximating the traces that language leaves behind. The traces are rich, structured, and informationally dense. They are not the meaning that produced them. Understanding this distinction, across all four registers simultaneously, is the only adequate response to the confusion that currently dominates public discourse about artificial intelligence.