# created ```dataview LIST WITHOUT ID file.link FROM -"windsurf" WHERE file.cday = date(this.file.name) AND !startswith(file.folder, "windsurf") SORT file.cday ASC ``` # modified ```dataview LIST WITHOUT ID file.link FROM -"windsurf" WHERE file.mday = date(this.file.name) AND !startswith(file.folder, "windsurf") SORT file.mday ASC ``` --- # [[diary and thoughts]] #thought #diary --- # Notes ## 5. Making [[Order Perceptible]] – [[Text Mechanics]] Carlson's account of [[order appreciation]] depends on a specific [[knowledge requirement]]. To appreciate an environment 'as what it is,' we need a general non-aesthetic story that identifies the relevant processes and materials, and gives us categories that guide _acts of aspection_—telling us what to look at and how to organise our perception. In natural environments, geology and ecology play this role. They provide concepts like strata, erosion channels, and habitats, which map directly onto features we can see in a landscape. By contrast, while chemical physics offers a more fundamental description of the same matter, its categories—molecular bonds, reaction kinetics—do not correspond to the patterns available to someone standing in a valley or before a cliff face. For [[aesthetic appreciation]], the mid-level description is the right one because it makes the visible order clear.%%the last sentence is an incorrect interpretation of what my view is.why do LLMs have so much trouble articulating the subtlties of this point. Please see these paragraphs from my old draft for my actual more subtle view. You can use A LOT of this text, the precise words phrases sentences, in the next iteration of this section, because it is better and clearer and a much [[more accurate]] reflection of my views on this particular idea than [[what you]] have produced so far: Carlson’s starting point was that appropriate appreciation of nature depends on a general, non-aesthetic story that makes its [[order visible]]. As he puts it, the appreciator “selects objects of appreciation from the things around him or her and focuses on [[the order]] imposed on these objects by the various forces, random and otherwise, that produce them”, and does so “by reference to a general nonaesthetic and nonartistic story that helps make this [[order visible]] and intelligible” (Carlson, 2000, p. 119). In the natural case, this story is typically supplied by the environmental sciences. Geology explains how cliffs and valleys are shaped by erosion and sedimentation [[over time]]. Ecology explains how the distribution of plants in a meadow follows from competition, cooperation, and niche structure. Meteorology explains large-scale patterns in clouds and weather systems. With such accounts in mind, what might otherwise [[look like]] accidental shape or colour can be read as the trace of identifiable processes, and this guides what we attend to when we appreciate an environment. One might note that nothing in principle rules out more fundamental sciences here. Chemical physics also explains the cliff face, in terms of [[the structure]] and interaction of molecules, and in principle one could build an order-based aesthetics of nature around such accounts. For present purposes, however, geology makes Carlson’s idea of aspection easier to explain. Geological categories such as strata, faults, and erosion channels can be used directly as ways of looking: they tell the appreciator which bands to treat as distinct, which lines to trace, which contrasts of texture and angle to attend to when they look at a cliff with the naked eye. The microstructural patterns described by chemical physics do not map as straightforwardly onto features that are available to ordinary perception, so it is harder to show how such knowledge might guide someone’s acts of aspection in the field. This is enough for the present point: Carlson’s recommendation can be introduced most naturally by starting from sciences whose categories can already function as ways of looking, even if more fundamental levels could in principle support a similar story. The LLM case presents a similar range of options. On the one hand there are low-level approaches such as mechanistic interpretability and causal intervention studies. These track the behaviour of particular attention heads or neurons, and how changing individual weights alters outputs. On the other hand there is naïve reading, which treats outputs as the sayings of a subject with beliefs, intentions, and projects. Section 3 argued that the latter view conflicts with Carlson’s rule that we should not project a planner where the best background knowledge tells us there is none. The former can, in principle, also provide the kind of non-aesthetic story Carlson has in mind: it explains how internal structures and circuits give rise to patterns of behaviour. For present purposes, however, it plays [[the role]] that chemical physics did in the natural case. Mechanistic interpretability works with categories that are far from the patterns that appear in ordinary reading, so it is harder to show how such knowledge might guide the kinds of acts of aspection that most users actually perform when they read model outputs. If we want an analogue of geology or ecology in the LLM case, it is more straightforward to start from a way of using background knowledge about model behaviour that does not personify the system, but that connects directly to the textual patterns that are available to a reader.%% We face a similar choice of level when appreciating LLMs. As described in Section 2, these systems can be analysed at the level of individual neurons, attention heads, and circuits. This work, often termed _mechanistic interpretability_, is important for scientific understanding and safety engineering. However, like chemical physics in the landscape case, its categories are too fine-grained to guide the appreciation of outputs. A reader encountering generated text does not perceive the activation of specific neuron clusters; they perceive discourse structures, stylistic shifts, reasoning patterns, and breakdown modes. To guide acts of aspection in LLM-mediated chats, we need a description that connects the technical reality of the model to these perceptible textual patterns. This section sets out such a framework. We need a level of description that respects the non-agentive nature of the system while focusing on _linguistic structure_, since text is what the system produces. The order we are concerned with is not the arrangement of weights on a hard drive, but the stable patterns that emerge in the generated text itself. %%this is really unclear and lacking in detail%% This order is linguistic and statistical. As we saw in Section 2, LLMs are trained on vast corpora of human text to predict next tokens. In doing so, they encode the _statistical regularities of language use_: which words tend to appear together, how syntactic structures unfold, how register and genre constrain word choice. These regularities are not explicit rules; they are patterns inherent in the collective linguistic practice recorded in the training data. No individual speaker consciously represents all these statistical dependencies, yet they structure our language use. LLMs compress these regularities into a set of mathematical parameters.%%this is really unclear and lacking in detail%% However, we are not dealing with "language" in the abstract. Different models are trained on different data mixtures, with different architectures and objectives. Each model therefore encodes a specific _parameterisation_ of these linguistic regularities. One model may capture certain stylistic nuances that another misses; one may be heavily constrained by safety tuning to refuse certain prompts, while another is more permissive. When we look at the outputs of a specific model, we are seeing the statistical structure of language as captured, approximated, and constrained by that particular system. %%this is really unclear and lacking in detail there is also an increasing problem with jargon entering the picture%% This suggests two levels at which we can appreciate LLMs. At the first level, LLMs as a class give us access to a _shared linguistic order_: they show the large-scale statistical patterns of language use that are normally too diffuse for us to grasp. At this level, we appreciate the system as a way of exploring the "statistical unconscious" of collective linguistic practice—the deep regularities that govern usage without any speaker's explicit intent. At the second level, each individual model represents a specific _design_ for accessing and shaping that order. Here, we can appreciate the functional beauty of a particular system: how well its architecture and alignment choices allow us to explore that linguistic substrate, or how distinctively it structures the generative possibilities.%%lacking in detail and so very uncompelling as part of an argument%% We call the body of knowledge that supports this appreciation _text mechanics_. By _text mechanics_, we mean a mid-level understanding of how a specific model's learned constraints over tokens give rise to _recurrent patterns in its generated text_. It studies patterns such as typical discourse formats, characteristic habits of hedging or refusal, ways of sustaining or dropping threads, and systematic modes of breakdown. It explains these patterns not by appealing to a "mind" or "personality," but by referencing the system's components: its embedding geometry, attention dynamics, training data composition, and decoding procedures. %%i hate the way this paragraph is written. "It studies" as though we are starting a new academic discipline. It's just shit. %% Text mechanics sits at the junction between the technical and the textual. It does not require tracking individual weights, nor does it treat the output as the speech of a person. Instead, it treats the output as the result of _trained mathematical operations on linguistically structured units_. The substrate—the token—is not a neutral physical particle but a unit of language, already saturated with cultural usage patterns. The processes—attention, sampling—are mathematical transformations that preserve and recombine these patterns according to learned probabilities. In this sense, text mechanics functions for LLM-mediated chats as geology and ecology do for natural environments. It identifies the relevant processes (training, architectural constraints) and materials (linguistically structured tokens). It provides categories—such as model-specific generative profiles and recurrent textual structures—that map onto what a reader actually observes. By understanding this knowledge, a user can look at a stream of generated text and see it not as a miraculous burst of intelligence, nor as random noise, but as the clear trace of a generative system acting on a linguistic corpus. %%jargon jargon, what a load of shit%% This view shares some ground with other recent accounts but differs in focus. Picca (2025) describes LLMs as "dynamic semiotic machines" that recombine signs within a "semiosphere." We share Picca's rejection of the cognitivist view that treats LLMs as minds. However, while Picca focuses on interpretation, ideology, and meaning-making, our focus here is on _structural order_. We are interested in how the statistical constraints learned by a model produce stable patterns of generation, regardless of their semantic or ideological content. Text mechanics is a pre-hermeneutic framework: it describes the generative texture of the text that interpretation then works upon. %%this is really unclear and lacking in detail%% Similarly, Janus (2022) characterises GPT-style models as "simulators" that can generate various "simulacra" (processes or personas) under different prompts. We accept the distinction between the underlying model (the simulator) and the specific outputs it generates. However, our interest is not in the ontological status of the simulacra as agents, but in the _generative profile_ of the simulator itself. We appreciate the system by seeing how its learned constraints shape the evolution of the text over time.%%this is really unclear and lacking in detail%% We now have a description of the kind of order LLMs instantiate—statistically emergent, linguistically grounded, and model-relative—and an account of the mid-level knowledge, text mechanics, that makes this order available to aspection. With this framework in place, we can turn to specific episodes of LLM-mediated interaction. In the next section, we will examine how attention guided by text mechanics yields aesthetic appreciation of these episodes as forms of a model's generative order.%%this is really unclear and lacking in detail%% ## another version of the same section Making Order Perceptible – Text Mechanics In this section we first return to Carlson’s knowledge requirement and his use of geology as an example of order appreciation. We then explain what kind of order LLMs instantiate, and how this varies between models. On that basis we introduce _text mechanics_ as a mid-level way of understanding LLM behaviour that can guide acts of aspection, and we close by situating this view in relation to Picca’s semiotic approach and Janus’s simulator picture. Carlson’s starting point is that appropriate appreciation of nature depends on a general, non-aesthetic story that makes its order visible. As he puts it, the appreciator “selects objects of appreciation from the things around him or her and focuses on the order imposed on these objects by the various forces, random and otherwise, that produce them”, and does so “by reference to a general nonaesthetic and nonartistic story that helps make this order visible and intelligible” (Carlson, 2000, p. 119). In the natural case, this story is typically supplied by the environmental sciences. Geology explains how cliffs and valleys are shaped by erosion and sedimentation over time. Ecology explains how the distribution of plants in a meadow follows from competition, cooperation, and niche structure. Meteorology explains large-scale patterns in clouds and weather systems. With such accounts in mind, what might otherwise look like accidental shape or colour can be read as the trace of identifiable processes, and this guides what we attend to when we appreciate an environment. One might note that nothing in principle rules out more fundamental sciences here. Chemical physics also explains the cliff face, in terms of the structure and interaction of molecules, and in principle one could build an order-based aesthetics of nature around such accounts. For present purposes, however, geology makes Carlson’s idea of aspection easier to explain. Geological categories such as strata, faults, and erosion channels can be used directly as ways of looking: they tell the appreciator which bands to treat as distinct, which lines to trace, which contrasts of texture and angle to attend to when they look at a cliff with the naked eye. The microstructural patterns described by chemical physics do not map as straightforwardly onto features that are available to ordinary perception, so it is harder to show how such knowledge might guide someone’s acts of aspection in the field. This is enough for the present point: Carlson’s recommendation can be introduced most naturally by starting from sciences whose categories can already function as ways of looking, even if more fundamental levels could in principle support a similar story. The LLM case presents a similar range of options. On the one hand there are low-level approaches such as mechanistic interpretability and causal intervention studies. These track the behaviour of particular attention heads or neurons, and how changing individual weights alters outputs. On the other hand there is naïve reading, which treats outputs as the sayings of a subject with beliefs, intentions, and projects. Section 3 argued that the latter view conflicts with Carlson’s rule that we should not project a planner where the best background knowledge tells us there is none. The former can, in principle, also provide the kind of non-aesthetic story Carlson has in mind: it explains how internal structures and circuits give rise to patterns of behaviour. For present purposes, however, it plays the role that chemical physics did in the natural case. Mechanistic interpretability works with categories that are far from the patterns that appear in ordinary reading, so it is harder to show how such knowledge might guide the kinds of acts of aspection that most users actually perform when they read model outputs. If we want an analogue of geology or ecology in the LLM case, it is more straightforward to start from a way of using background knowledge about model behaviour that does not personify the system, but that connects directly to the textual patterns that are available to a reader. So far we have only said that we need a mid-level, language-facing account that respects what LLMs are technically, but meets the reader at the level of text. To get clearer on what that account needs to cover, we now say more directly what sort of order is present in LLM outputs. As Section 2 stressed, current LLMs are trained on large corpora of human-produced text to predict the next token in a sequence. The training procedure pushes the model to match, as well as it can, the conditional structure of the training data: which units tend to follow which others in which contexts. The units in question are _tokens_: words, subwords, or characters which, taken together, form the sentences, paragraphs, and documents of the corpus. By fitting this prediction task, the model comes to encode a wide range of statistical regularities of language use: which expressions co-occur, which syntactic frames are common in which registers, how certain discourse moves tend to follow others. These regularities are not written down as rules anywhere in the model; they are implicit in the way the parameters have been adjusted during training. Nor are they typically represented as explicit rules by individual speakers. They are properties of large collections of usage that become visible only when we aggregate at scale. However, we should not say that a trained model simply “gives us language itself”. Different models are trained on different corpora, with different filtering, different proportions of genres and languages, and different fine-tuning objectives. Their architectures also differ in depth, width, and attention patterns. Post-training procedures such as instruction tuning and reinforcement learning from human feedback further bias their outputs towards particular interaction styles. The result is that each concrete model encodes a specific way of capturing and organising the statistical structure present in its training data. Two models may both be “English assistants”, but diverge in how they handle humour, or in how quickly they hedge, or in how they prioritise safety over directness. The order we meet in any particular chat is therefore not Language with a capital “L”, but _language-as-modelled-by-this-system-under-these-constraints_. Even with that caveat, there is still a useful sense in which LLMs as a class open up a shared object of appreciation. All of them, if they are at all successful, approximate large-scale regularities in human language use. All of them can be probed interactively, in real time, by giving prompts and watching continuations. In that sense they provide, for the first time, a kind of dynamic access to the “long-range” structure of a language community’s practice: to patterns which are latent in corpora but not normally experienced as a coherent field. At this _shared_ level we can talk about a linguistic order that LLMs make perceptible: the way contemporary English (or some mix of languages) tends, taken in the large, to connect topics, develop explanations, switch registers, and so on. At the same time, because each model instantiates its own parameterisation, there is a second, _model-specific_ level at which order can be appreciated. Here, what matters is not only that the system gives us access to some portion of the linguistic order, but the particular way in which it does so: which genres it tends to favour, how it balances directness and politeness, how it structures long answers, what kinds of drift or breakdown it displays. This is where design appreciation, in Carlson’s and Forsey’s sense, has a natural foothold. Given a general function—“serve as a general-purpose assistant based on English text”—different labs and projects realise it in recognisably different ways. We can then ask which designs make the underlying linguistic regularities more or less tractable, more or less unified, more or less satisfying to engage with. To talk about these two levels in a more systematic way, it is helpful to have a label for the kind of mid-level understanding that connects model internals to textual patterns. By _text mechanics_ we mean a way of using existing computer science knowledge about LLMs to explain and organise their outputs for appreciation. It is not a new discipline, but a way of bringing together several strands of work that already exist. On the one side, we have accounts of distributional semantics and representation learning, which describe how tokens are placed in high-dimensional embedding spaces so that items used in similar contexts lie close together. We have analyses of transformer architectures and attention heads, which show how information about earlier tokens is carried forward, how long-range dependencies are maintained, and how different heads specialise. We have studies of training data composition and scaling laws, which show how increasing data size, parameter count, or training steps changes the stability and breadth of learned patterns. We have work on decoding strategies and temperature, which describes how sampling policies alter the balance between safe, high-probability continuations and more adventurous, low-probability ones. We have accounts of post-training and reinforcement learning from human feedback, which show how a base model’s raw predictive tendencies are nudged towards a cooperative assistant profile. Text mechanics, as we use the term, draws selectively on this work to make sense of recurrent structures in a model’s outputs. It treats as orders in the text any stable patterns in how a given model tends to continue, structure, or lose grip on a line of writing across prompts and tasks. Examples include familiar “shapes” of explanation (list-plus-summary, tree-like branching, cautious step-by-step reasoning), characteristic refusal and hedging templates, typical ways of handling ambiguity, and common breakdown trajectories under long or conflicting prompts. The point is not to classify these patterns exhaustively, but to have enough understanding of the underlying training and architecture to say: “this is the sort of thing we expect from this model, for these technical reasons”, and to use that understanding to guide what we attend to. This puts text mechanics in the right place to play Carlson’s geology role. On the one side, it keeps faith with the technical story told in Section 2. It does not pretend there is a subject behind the screen; it does not attribute beliefs or desires to the system. It speaks instead of embeddings, attention, loss functions, sampling policies, and feedback training. On the other side, it stays at a scale where its categories can be brought to bear on what we actually read. It tells us that a certain style of abrupt topic switch is a known failure mode of this architecture at this context length; that a certain repetitive hedging style reflects a particular alignment regime; that a certain sort of graceful degradation under noise reflects robust training on heterogeneous web data. These are the kinds of links that can guide acts of aspection: they tell us which features of a transcript are expressive of the model’s learned constraints, and which are likely to be contingent on the prompt or trivial surface noise. In that sense, text mechanics is for LLM-mediated chats what geology and ecology are for landscape appreciation. Geology identifies processes such as uplift and erosion, and materials such as rock strata and sediments, and gives us categories—faults, ridges, terraces—that we can literally trace with our eyes. Ecology introduces categories like successional stage, niche, and competition, which help us see a wood or meadow as an organised field rather than a random clump of plants. Analogously, text mechanics identifies the processes by which a model is trained and shaped, the linguistic “material” it operates on, and the characteristic ways in which that material is organised in use. It gives us categories like “this model’s long-answer template”, “this family’s refusal profile”, “this architecture’s breakdown mode”, which can be used in reading just as “stratum”, “fault” and “alluvial fan” can be used in looking. This picture overlaps in some ways with recent semiotic accounts but has a different focus. Picca (2025) suggests that LLMs should be seen as “dynamic semiotic machines” operating within Lotman’s semiosphere: they recombine signs, respond to prompts as semiotic acts, and produce outputs that invite interpretation. We agree with Picca on two important points: that LLMs are not minds, and that they should be understood as systems that manipulate signs rather than as bearers of original thought. Where our approach differs is in emphasis. Picca is primarily concerned with meaning-making, ideology, and the role of LLM outputs in an ongoing ecology of interpretation. Our concern here is prior to that. We want a clear account of how a particular model’s training and architecture shape the _structure_ of its outputs as text—how it tends to extend, qualify, or derail a line of writing—so that order appreciation in Carlson’s sense has something determinate to latch onto. Questions about how those structures are then taken up in interpretation and politics are important, but they lie downstream of the present project. Janus (2022) offers another useful frame by describing GPT-style models as _simulators_. On this view, the trained network is a rule that, given a text prefix, generates continuations in line with a learned distribution, and different prompts “call forth” different simulacra—fictional agents, argument lines, or worlds. We take over one key distinction from this picture: the distinction between the simulator (the underlying model with its training story) and the simulacra (the specific processes and personae that can be generated under prompts). Our interest, however, is not in the simulacra as quasi-agents, as Section 3 already argued against treating them as persons for aesthetic purposes. Instead, we treat episodes of interaction as episodes of the simulator’s characteristic dynamics under constraint. What matters for us is how a particular model, given a certain framing and sampling policy, tends to produce order, lose it, or hold it at a certain level of complexity. Text mechanics is the way of talking about that without sliding back into agent-talk. We are now in a position to say, with some precision, what kind of order LLMs instantiate and what kind of knowledge makes that order available to appreciation. The order is not natural in Carlson’s sense, nor simply designed. It is statistically emergent from large corpora of linguistically and culturally structured text, as captured and organised by specific training and alignment regimes. The right kind of knowledge is not a full micro-level account of every neuron, nor a naïve projection of human minds into the machine, but a mid-level understanding—text mechanics—that links technical facts about training and architecture to the patterns a careful reader can actually see. In the next section we put this framework to work by looking at particular LLM-mediated episodes and showing how, under text-mechanical aspection, they can be appreciated as forms of a model’s generative order. ## old version 22 Nov 2025 %%no references, mention the semiotics paper and some more technical paper. Chalmers on 2002 survery consciouness %% In this section I bring Carlson’s order-based framework back to the LLM case. I first recall how, in natural settings, scientific knowledge makes order visible for appreciation. I then introduce _text mechanics_ as a way of using existing computer science work on LLMs to play a similar role for model outputs, and I close by explaining how this prepares the ground for the extended example that follows. Carlson’s starting point was that appropriate appreciation of nature depends on a general, non-aesthetic story that makes its order visible. As he puts it, the appreciator “selects objects of appreciation from the things around him or her and focuses on the order imposed on these objects by the various forces, random and otherwise, that produce them”, and does so “by reference to a general nonaesthetic and nonartistic story that helps make this order visible and intelligible” (Carlson, 2000, p. 119). In the natural case, this story is typically supplied by the environmental sciences. Geology explains how cliffs and valleys are shaped by erosion and sedimentation over time. Ecology explains how the distribution of plants in a meadow follows from competition, cooperation, and niche structure. Meteorology explains large-scale patterns in clouds and weather systems. With such accounts in mind, what might otherwise look like accidental shape or colour can be read as the trace of identifiable processes, and this guides what we attend to when we appreciate an environment. One might note that nothing in principle rules out more fundamental sciences here. Chemical physics also explains the cliff face, in terms of the structure and interaction of molecules, and in principle one could build an order-based aesthetics of nature around such accounts. For present purposes, however, geology makes Carlson’s idea of aspection easier to explain. Geological categories such as strata, faults, and erosion channels can be used directly as ways of looking: they tell the appreciator which bands to treat as distinct, which lines to trace, which contrasts of texture and angle to attend to when they look at a cliff with the naked eye. The microstructural patterns described by chemical physics do not map as straightforwardly onto features that are available to ordinary perception, so it is harder to show how such knowledge might guide someone’s acts of aspection in the field. This is enough for the present point: Carlson’s recommendation can be introduced most naturally by starting from sciences whose categories can already function as ways of looking, even if more fundamental levels could in principle support a similar story. The LLM case presents a similar range of options. On the one hand there are low-level approaches such as mechanistic interpretability and causal intervention studies. These track the behaviour of particular attention heads or neurons, and how changing individual weights alters outputs. On the other hand there is naïve reading, which treats outputs as the sayings of a subject with beliefs, intentions, and projects. Section 3 argued that the latter view conflicts with Carlson’s rule that we should not project a planner where the best background knowledge tells us there is none. The former can, in principle, also provide the kind of non-aesthetic story Carlson has in mind: it explains how internal structures and circuits give rise to patterns of behaviour. For present purposes, however, it plays the role that chemical physics did in the natural case. Mechanistic interpretability works with categories that are far from the patterns that appear in ordinary reading, so it is harder to show how such knowledge might guide the kinds of acts of aspection that most users actually perform when they read model outputs. If we want an analogue of geology or ecology in the LLM case, it is more straightforward to start from a way of using background knowledge about model behaviour that does not personify the system, but that connects directly to the textual patterns that are available to a reader. %%enrico like this a lot%% %%but he hates this. there needs to be more stuff on acts of aspection%%By _text mechanics_ I mean this kind of use of existing computer science knowledge about LLMs to explain and organise their outputs for appreciation. It draws on several strands of work that are already in place. Distributional semantics and representation learning study how training arranges tokens in high-dimensional embedding spaces, so that tokens with similar patterns of use are close together. Work on transformer architectures and attention studies how different attention heads and layers track dependences between tokens over short and long ranges. Work on scaling laws and optimisation studies how training on larger data sets and models changes the stability and breadth of the patterns a model can reproduce. Work on decoding, temperature, and nucleus sampling studies how different sampling policies trade off between safe, high-probability continuations and more varied, low-probability ones. Work on instruction tuning and reinforcement learning from human feedback studies how a base model’s behaviour is changed when it is fine-tuned to follow instructions and to exhibit an assistant-like profile. Text mechanics is not a new discipline that replaces these areas. It is a way of taking selected results from them together, and using them to make sense of the recurrent orders that show up in a model’s outputs. This role can be seen by looking at the kinds of generalisations that text mechanics supports. It treats as orders in the outputs any recurrent patterns in how a model continues, structures, or degrades text across prompts and tasks: for example, stable families of discourse templates, typical ways of building up and then dropping a line of thought, characteristic styles of inventing and reusing made-up words, or familiar breakdown patterns under length or pressure. The details differ between models and deployments, but in each case we are dealing with repeatable textual structures that show up at the surface and that can be tracked by a reader. A text-mechanical account then links such orders to non-intentional processes described in computer science: how the training corpus is composed, how optimisation in a transformer yields particular embedding geometries and attention patterns, how decoding settings and reinforcement learning from human feedback bias the model towards some continuations rather than others. Taken together, these links supply what Carlson calls the “general nonaesthetic and nonartistic story” (Carlson, 2000, p. 119). For the appreciator, this story guides aspection: it tells them which textual patterns to treat as diagnostic of how a particular system is built and trained, and where to look if they want to compare two systems or to understand what kind of order they are encountering. These examples show how text mechanics fits Carlson’s pattern for order appreciation. In each case there is a perceivable order in the outputs: in hedging frames, in the shallow list-like structure of arguments, in recurring ways of failing. There are forces that produce that order: the way training data is composed, the dynamics of optimisation in a transformer, the decoding and feedback procedures that shape which continuations are favoured. There is a general non-aesthetic story, drawn from computer science, that links the two. And that story dictates acts of aspection, in the sense that it tells the appreciator which patterns are worth attending to, which contrasts are informative, and where to expect both strengths and weaknesses. Text mechanics is that story, applied not to rock strata or ecosystems but to the behaviour of a trained language model in use. In practice, we never encounter the generative regularities of these systems “in the abstract”. We only ever meet it in particular episodes: runs of interaction by a concrete system, under a specific system prompt and safety setup, with a particular prompt history. Often we do not know that history in detail, especially when we study striking outputs that circulate without full context. Text mechanics responds to this by focusing on features of an episode that are robust across a reasonable range of prompts, rather than on details that could be written into the prompt itself. When a user instructs a model to answer in bullet points, we should not treat the presence of bullet points as revealing its order. When a model produces patterns of neologism, internal echo, or gradual breakdown that match what we know about its embedding geometry, attention span, and decoding settings, it is more plausible to treat these as displays of its generative order under the given boundary conditions. The aim is to ascribe only those aspects of an episode to the model that cannot be straightforwardly traced to explicit user constraints. From this standpoint, it is natural to adopt an environment framing. A chat system based on an LLM is not a subject behind the screen but a large, trained field of dispositions to continue text in certain ways under certain inputs. Each episode of interaction is a path through that field, shaped by the prompts and by the system prompt and safety regime. One might worry that this simply replaces one metaphor (“person”) with another (“environment”). I do not think the two are on a par. Personifying talk suggests a subject with beliefs, projects, and responsibilities, and so brings in kinds of state and norm that are not part of the technical account in Section 2. Environment talk, by contrast, reuses a term that already has a role in Carlson’s framework and asks us to treat the system as a domain in which non-intentional generative processes produce orders that can be explored. It does not posit any further kinds of entity beyond the processes that computer science already posits for LLM training and operation; it only changes how we group and attend to the outputs. %%he likes some of this paragraph%% On this reading, the aesthetic appreciation of LLMs as environments is not a new kind of response. It is Carlson’s order appreciation applied to an artificial, linguistic case: we attend to perceivable patterns in model outputs as the expression of underlying generative processes described by text mechanics. In the next section I turn to an extended example, drawn from an unusual piece of output by Claude Opus 3, to show how this way of looking can be used in practice. There I treat the text not as the work of an authorial mind, but as a revealing episode in the behaviour of a trained system under prompts that loosen some of its usual constraints.