The previous sections have shown why neither person-appreciation nor pure design-appreciation adequately captures what is involved in appreciating LLMs. Person-appreciation fails because LLMs are not agents but statistical systems operating on numerical tokens. Design-appreciation proves incomplete because the specific patterns LLMs exhibit emerge from training rather than being directly designed. [[The question]] now is: what kind of knowledge would allow us to properly appreciate these systems—to see and understand [[the order]] that emerges in their outputs? Carlson's framework suggests that appropriate appreciation requires knowledge suited to [[the nature]] of [[the object]]. For LLMs, this means knowledge that makes visible the patterns arising from statistical learning. The field most directly suited to this task is natural [[language processing]] (NLP), the discipline that studies computational models of [[human language]]. Within NLP, we can identify a perspective we might call [[text mechanics]]—the study of how training objectives, data patterns, and architectural constraints produce observable regularities in generated text. Unlike fields such as mechanistic interpretability or [[dynamical systems]] analysis, which require specialised tools to observe internal model states, [[text mechanics]] focuses on patterns visible in the text itself. This makes it particularly suited for the everyday user who encounters LLMs through their outputs rather than through technical interfaces. To understand how [[text mechanics]] reveals order in LLM outputs, consider what the model actually learns during training. The corpus contains millions of instances where "doctor" appears near "patient", "diagnosis", "treatment", "hospital". Through training, these co-occurrence patterns become encoded in the model's parameters. When generating text, if "doctor" appears, the model raises probabilities for medically-related tokens. This is not knowledge that doctors treat patients but rather that certain tokens cluster together in [[the training]] distribution. Similarly, the model learns syntactic patterns—that "the" often precedes nouns, that "gave" typically appears with two objects ("gave him the book"), that questions beginning "Why" tend to be followed by explanations. These patterns layer together [[during generation]], creating text that appears meaningful despite emerging from purely statistical processes. [[Text mechanics]] operates at multiple scales simultaneously. At the smallest scale, we see collocations—words that habitually appear together. "Strong" collocates with "coffee" but also with "wind" and "evidence", each collocation carrying different associative patterns that influence subsequent generation. At the phrase level, we find constructional patterns—recurring templates like "not only X but also Y" or "on the one hand X, on the other hand Y". At the sentence level, the model maintains syntactic coherence and semantic consistency, ensuring pronouns have clear antecedents and verb tenses align. At the discourse level, the model exhibits topic maintenance and progression, stylistic consistency, and genre-appropriate structures. A story maintains narrative progression; an explanation maintains its explanatory register; an argument maintains its [[logical structure]]. These patterns become visible when we [[know what]] to look for—what Carlson calls acts of aspection. When reading LLM output, we can attend to how the model maintains register throughout a response, how it handles transitions between topics, how it patterns its explanations. We notice that certain phrasings recur, that certain conceptual moves repeat across different contexts. These are not signs of understanding but traces of statistical learning. When ChatGPT begins responses with "Great question!" or Claude opens with "Let me think through this systematically", these are not expressions of enthusiasm or thoughtfulness but high-probability patterns for conversation openings in their respective training distributions. Understanding text mechanics also explains differences between models. GPT-4, trained heavily on academic text, exhibits frequent hedging—"arguably", "potentially", "it seems that". Claude, with training emphasising helpful assistance, produces more direct assertions and clearer structure. These are not personality differences but statistical signatures of different training distributions. Longer context windows allow models to maintain references and themes across extended passages. Larger models distinguish finer gradations of meaning—"brilliant" appearing in distinct contexts for scientists versus performances versus ideas. These differences leave traces in the generated text: more consistent register, more precise word choice, more sophisticated syntactic variation. This perspective also illuminates what happens during conversation with an LLM. Each exchange adds to the context, shifting the probability landscape for subsequent generation. Early mentions of technical topics activate technical vocabulary; casual phrasing encourages casual responses. The conversation develops not through mutual understanding but through accumulating statistical constraints. The model is not tracking what we mean or building a mental model of our discussion. It is navigating a high-dimensional probability space, with each token choice influenced by all previous tokens in the conversation. Text mechanics thus provides the knowledge needed for appropriate aesthetic appreciation of LLMs. It reveals the order in what might otherwise seem like mysterious intelligence or mere randomness. It shows us patterns to attend to, dependencies to trace, regularities to observe. Through this lens, we can appreciate the intricate statistical patterns that emerge from training, the subtle ways context shapes generation, the remarkable coherence that arises from purely mechanical processes. We appreciate LLMs neither as persons nor purely as designed tools, but as systems exhibiting complex order arising from simple rules applied at massive scale. This order—statistical, emergent, neither intended nor random—becomes the proper object of our aesthetic attention.