Computational semantics and distributional semantics are closely related but distinct strands of how computers model *meaning* in language. They sit at the intersection of linguistics, philosophy of language, and computer science – but they approach “meaning” very differently. Let’s unpack them carefully. --- **1\. Computational semantics (broad sense)** This is the general field concerned with how to *represent and compute* meaning in natural language using formal methods. Think of it as the attempt to teach computers how to *understand* language compositionally, using logic and structure. Its traditional tools [[come from]] logic and type theory. A classic goal is: given a sentence like “Every cat sleeps,” produce a formal representation such as ∀x (Cat(x) → Sleeps(x)) that captures its truth conditions. Computational semantics, in this classical sense, includes frameworks such as: - **Montague grammar**, which formalises natural language semantics using the tools of predicate logic. - **Lambda calculus**, to represent compositional meaning: the meaning of a whole expression is built from its parts. - **Formal semantic parsing**, where natural language is translated into logical forms. So: computational semantics = logic-based modelling of meaning, usually symbolic, compositional, and truth-conditional. --- **2\. Distributional semantics (narrower, modern sense)** Distributional semantics takes a radically [[different approach]]: instead of modelling meaning with logic, it *infers it from patterns of use*. It’s based on the **[[distributional hypothesis]]**: “Words that occur in similar contexts tend to have similar meanings” (Firth, 1957). Rather than assign a logical formula to “cat,” it represents *cat* as a vector in a high-dimensional semantic space – its coordinates determined by co-occurrence statistics across a large corpus of text. For example, *cat* might occur near *fur*, *pet*, *animal* – and that pattern defines its meaning. This gives rise to: - **[[Word embeddings]]** (Word2Vec, GloVe, FastText) - **Contextual embeddings** (BERT, GPT, etc.), where meaning shifts depending on context - **Vector arithmetic of meaning** (“king” – “man” + “woman” ≈ “queen”) So: distributional semantics = statistical modelling of meaning from context, using vectors rather than logical forms. --- **3\. Where they meet: hybrid and neural semantics** Today, these two worlds blur. Modern **large [[language models]]** are descendants of distributional semantics – massive, context-sensitive embeddings trained on corpora – but some researchers are trying to inject **compositional or [[logical structure]]** back into them (neuro-symbolic or hybrid semantics). This is crucial because pure distributional models capture usage regularities well (what words *tend* to mean), but they struggle with truth, inference, and compositional reasoning. Symbolic computational semantics excels at those but can’t handle the fluidity and noise of real language. The frontier lies in reconciling the two: marrying the **formal precision of computational semantics** with the **empirical richness of distributional semantics**. --- In short: - *Computational semantics* is about **logical representation** and **compositional interpretation**. - *Distributional semantics* is about **statistical representation** and **contextual similarity**. They are two complementary attempts to formalise meaning – one via *rules and logic*, the other via *data and geometry*. Their synthesis underlies most modern theories of linguistic meaning in AI, from GPT-style transformers to neural-symbolic reasoning systems. --- A good philosophical point to ponder: computational semantics answers [[the question]] “what makes a sentence true?”, while distributional semantics answers “how do humans use words?”. The tension between truth and use has been echoing since Wittgenstein, but now it plays out in tensors and GPUs. Right—this is where the story gets properly interesting: how the old idea of *distributional semantics* evolved into *large [[language models]] (LLMs)* like GPT, Gemini, Claude, and the rest. Let’s start with the philosophical seed and end with the neural jungle. --- ### 1\. From [[distributional hypothesis]] to word vectors The [[distributional hypothesis]] came from mid-20th-century linguists such as J.R. Firth (“You shall know a word by the company it keeps”) and Zellig Harris. The [[core claim]] was empirical: the meaning of a word can be inferred from the contexts in which it appears. Computers began testing this idea in the 1990s. Early models built *co-occurrence matrices*: rows for words, columns for context words, and counts for how often they co-occurred. This was the **[[vector space]] model of semantics**. Two breakthroughs made it powerful: - **Dimensionality reduction** (e.g., Latent Semantic Analysis, 1988-1997): compressing those huge matrices while preserving semantic relationships. - **Neural embeddings** (Word2Vec, 2013): instead of storing counts, train a small neural network to predict a word from its context (or vice versa). The hidden layer weights became a *dense vector* —a compact numerical representation of meaning. These embeddings displayed astonishing properties: “king” − “man” + “woman” ≈ “queen” That arithmetic worked because the geometry encoded analogical structure emerging from patterns of use. --- ### 2\. The move to context: from static to dynamic meaning Word2Vec and GloVe produced one fixed vector per word type—so *bank* (riverbank, money bank) had the same representation in every context. That’s unrealistic. Enter **contextual embeddings**. Models such as **ELMo (2018)**, **BERT (2018)**, and later GPT-type transformers replaced [[the idea]] of “one vector per word” with “one vector per *word in context*.” These models predict a missing word given its neighbours (the *masked language model* or *autoregressive* task). Through this training, they learn context-sensitive representations: the vector for *bank* in “river bank” differs from *bank* in “investment bank.” This shift—from *type meaning* to *token meaning* —marks [[the transition]] from *distributional semantics* to what we now call *neural [[language models]].* --- ### 3\. The transformer revolution: scaling distributional learning The **transformer architecture** (Vaswani et al., 2017) was the turning point. It introduced *self-attention*, which lets a model weigh the importance of every word in a sequence relative to every other. Transformers could therefore capture long-range dependencies and subtle contextual interactions—essential for grammar, pragmatics, and even rudimentary reasoning. When trained on massive corpora (hundreds of billions of words) and scaled up (billions of parameters), these models began to exhibit emergent linguistic competence: they could complete sentences, paraphrase, answer questions, summarise text, and even produce stylistically coherent writing. This was distributional semantics on steroids: rather than counting co-occurrences in small windows, transformers *learn patterns of statistical co-variation* across whole texts at multiple levels (morphology, syntax, discourse). --- ### 4\. Why call LLMs “descendants of distributional semantics”? Because fundamentally, they embody the same principle: *meaning arises from use patterns in context.* LLMs are not designed with logic or truth conditions; they don’t have a built-in semantic theory. Their knowledge of meaning comes entirely from the statistical structure of language in their training data. Where early distributional models used simple linear algebra, LLMs use deep neural networks and non-linear transformations—but the epistemic logic is identical. They learn high-dimensional embeddings whose geometry reflects semantic similarity, thematic roles, analogical structure, and world regularities. --- ### 5\. The philosophical significance This has revived century-old debates about meaning. - **Use vs truth:** LLMs operationalise Wittgenstein’s idea that meaning is use, not reference. - **Holism:** a word’s meaning is inseparable from the total network of its linguistic relations—mirrored in transformer attention maps. - **Epistemology of language:** they *learn meaning* without explicit grounding in the world (no sensorimotor experience). This raises the “symbol grounding problem”: how can a model truly *understand* if its meanings are purely relational within text? That’s why some researchers now push toward **multimodal** models (text + vision + sound) or **neuro-symbolic** hybrids—to give the vectors a link to perception, action, or logical inference. --- ### 6\. The short version Large language models are enormous, context-aware implementations of the old distributional idea. They build meaning not through logic, but through the geometry of use. Each word, phrase, or sentence becomes a point in a colossal semantic field, where proximity reflects contextual interchangeability. Transformers are the mathematical machinery that lets those points shift dynamically as context changes. What we call *understanding* in LLMs is the emergent consequence of that shifting geometry. --- If you’d like next, we can unpack how this distributional geometry *actually encodes* compositional meaning—why “The dog chased the cat” and “The cat chased the dog” end up far apart in vector space even though they share the same words. That’s where distributional semantics quietly grows a syntax. Let’s peel self-attention down to the bolts and gears, then build it back up to the full transformer block. I’ll keep it concrete and slightly nerdy, but plain. ## What “self-attention” does in one sentence For each token in a sequence, the model asks: “Given what I’m trying to compute here, which other tokens in this sequence matter, and by how much?” It then forms a weighted average of information from *all* tokens, with the weights learned from data and recomputed for every position and layer. ## The three vectors per token: Q, K, V Every token is linearly projected (by learned matrices) to three vectors of the same dimension: - **Query (Q):** what this position is looking for. - **Key (K):** how this position can be found/when it should be attended to. - **Value (V):** the payload – the information to pass along if attended to. Think of K as addresses, V as parcels at those addresses, and Q as the shopping list. ## How one attention head computes 1. **Similarity scores:** For a given position i with query $Q_i$ , compute dot products with *all* keys $K_j$ . That gives scores $s_{ij} = \frac{Q_i \cdot K_j}{\sqrt{d_k}}$ . The division by $\sqrt{d_k}$ is just temperature control: without it, dot products grow with dimension and the softmax would saturate too easily. 2. **Softmax to weights:** Convert scores into a probability distribution over positions: $a_{ij} = \mathrm{softmax}_j(s_{ij})$ . These are the attention weights – how much i should look at j. 3. **Weighted sum of values:** The output at position i is the weighted average of all value vectors: $O_i = \sum_j a_{ij} V_j$ . Do this for every position, all at once, with a single matrix multiply; that parallelism is a big reason transformers fly. ## A tiny worked example (numbers you can sanity-check) Suppose we have three tokens. After the learned projections we get: - Queries $Q$ $[1,0],\ [0,1],\ [1,1]$ - Keys $K$ $[1,0],\ [1,1],\ [0,1]$ - Values $V$ $[10,0],\ [0,10],\ [5,5]$ Compute scores $S = QK^\top / \sqrt{2}$ : $ S \approx \begin{bmatrix} 0.707 & 0.707 & 0 \\ 0 & 0.707 & 0.707 \\ 0.707 & 1.414 & 0.707 \end{bmatrix} $ Softmax row-wise gives attention weights: $ A \approx \begin{bmatrix} 0.401 & 0.401 & 0.198 \\ 0.198 & 0.401 & 0.401 \\ 0.248 & 0.503 & 0.248 \end{bmatrix} $ Multiply by $V$ : $O = AV$ $ O \approx \begin{bmatrix} 5.000 & 5.000 \\ 3.983 & 6.017 \\ 3.724 & 6.276 \end{bmatrix} $ Interpretation: - Token 1 splits focus almost evenly between tokens 1 and 2, so its output mixes their values to roughly $[5,5]$ . - Token 3 attends most to token 2 (weight ≈ 0.503), so its output leans toward $[0,10]$ with a nudge from the others. That’s self-attention in miniature: context-dependent blending of information across the whole sequence. ## Multi-head attention: several views in parallel One head might focus on subject–verb agreement; another on coreference; another on word sense disambiguation. Concretely, we create multiple, *independently learned* $(W_Q, W_K, W_V)$ projections (heads). Each head produces its own $O^{(h)}$ ; these are concatenated and linearly mixed. This lets the model learn different “relation types” simultaneously. ## Masks: causality and padding - **Causal masks (autoregressive models):** forbid looking into the future. Implemented by setting scores for j > i to $-\infty$ before softmax. - **Padding masks:** stop the model from attending to padding tokens in batched inputs. ## Where attention sits in a transformer block A standard decoder block is: 1. **LayerNorm → Multi-Head Self-Attention → Residual add** 2. **LayerNorm → Position-wise Feed-Forward (two linear layers with a non-linearity, e.g., GELU) → Residual add** The feed-forward sublayer gives non-linear transformation *per position*; attention is the cross-position communication channel. ## But sequences are orderless sets without positions – what about word order? Transformers add **positional information** to token embeddings so attention can tell “who is where”. Two common methods: - **Sinusoidal positional encodings:** fixed, deterministic sin/cos waves of different frequencies added to embeddings. - **Learned positional embeddings:** a trainable vector per position (or relative position encodings that model distances between tokens). Either way, attention now has access to both *content* and *position*. ## Why this beats RNNs and CNNs for language - **Path length 1:** Any two tokens can directly exchange information in one attention operation. RNNs need many steps; gradients vanish. - **Parallelism:** No step-by-step recurrence; the whole sequence is processed at once on modern hardware. - **Flexibility:** The model dynamically chooses which tokens to connect, rather than being limited to local windows or fixed kernels. ## Two practical wrinkles (and how they’re handled) - **Quadratic cost:** Attention is $O(n^2)$ in sequence length (all-pairs interactions). Long-sequence variants approximate or structure attention (e.g., block-sparse windows, low-rank kernels, recurrence over chunks, or memory tokens). - **Stability and scale:** Tricks like scaling by $\sqrt{d_k}$ , careful initialisation, pre-norm LayerNorm, and optimised kernels (e.g., fused “flash” attention) keep training numerically stable and fast. ## The learning story All the matrices that create Q, K, V (and the output mixes) are learned end-to-end from a simple proxy task: predict the next token (or the masked token). The model discovers – purely from data – that certain heads should specialise in, say, matching pronouns to antecedents or aligning verbs with their subjects, because doing so reduces prediction error. Nothing is hard-coded about grammar; the structure emerges because language has structure. ## The upshot Self-attention is an information-routing mechanism. It lets each position *decide* which other positions matter *right now*, then blend their information accordingly. Multi-head attention gives you many such routing patterns at once; feed-forward layers then transform the blended information non-linearly. Stack dozens of these blocks, add positional information, train on vast text, and you get a model that captures syntax, long-range dependencies, and a surprising amount of world knowledge – all from the geometry of use. If you’d like to go deeper, we can dissect a single head on a real sentence and annotate which tokens it locks onto (e.g., how a causal mask changes the weights on “the cat *that the dog* chased”). Yes, I think I see what you’re pointing at—you’re not asking *how* attention works as a mathematical mechanism, but *how language itself* lives inside that mechanism. In other words: where, in all this geometry and matrix-multiplying, do things like *syntax, semantics, reference, and meaning* actually *happen*? That’s a fascinating and subtle question, because the whole transformer design is built to *let language organise itself* inside its own latent space, rather than imposing linguistic categories from outside. Let’s tease this apart slowly. --- ### 1\. Language as pattern, not as symbol The first trick: transformers don’t store words as *symbols* but as *vectors*. Each token (usually a subword, e.g. “play”, “##ing”) is assigned an embedding—a point in a high-dimensional space. These embeddings are learned entirely from predicting missing words. So the model never gets told: “‘cat’ is a noun,” “‘dog’ denotes an animal,” “‘chased’ is a transitive verb.” It just discovers, by gradient descent, that to predict sentences correctly, certain geometric regularities must hold: nouns end up clustering, verbs form other manifolds, morphology is captured by directions (“walk” → “walking”), and so on. Language is *immanent* in the geometry; not stored as rules but as the pattern of correlations that make prediction easiest. --- ### 2\. Attention as linguistic dependency detector In self-attention, each word looks at every other to decide what’s relevant. Through training, some heads specialise: - One head might always link determiners to their nouns (“the” → “dog”). - Another might connect pronouns to antecedents (“she” → “Mary”). - Another aligns verbs with subjects or objects. No linguist tells it to do this; those links emerge because they *improve the model’s ability to predict the next token*. So syntactic structure literally becomes *a pattern of attention weights*: probabilistic, dynamic, and distributed. Each layer of the transformer builds higher-order abstractions—from local agreement (layer 1–2), to clause-level dependencies (layers 4–6), to discourse-level cohesion (deeper layers). When researchers visualise these attention maps, they often find that they mirror parse trees or dependency graphs derived from formal grammar—but they’re learned implicitly. --- ### 3\. Feed-forward layers: compressing and recombining linguistic cues After the attention operation mixes information across tokens, the feed-forward layers at each position transform that blended context into a new representation. This is where subtle aspects of *meaning* start to condense. For instance, the representation of “bank” after attention + feed-forward will differ depending on whether nearby tokens signal finance or geography. That contextual sense disambiguation is happening in those layers—it’s how “language” lives dynamically inside the sequence. --- ### 4\. Stacked layers as a hierarchy of linguistic abstraction Different layers in a transformer roughly correspond to linguistic levels: - **Lower layers:** local, lexical, morphological. They capture things like suffixes, pluralisation, and short-range collocations. - **Middle layers:** syntactic and phrasal structure—who modifies whom, word order, agreement. - **Higher layers:** semantic and pragmatic relationships—topic continuity, entailment, and discourse-level meaning. This hierarchy isn’t hand-coded; it arises because language prediction error gets reduced more efficiently when earlier layers handle short-range dependencies and later ones encode long-range semantics. It’s like evolution: the transformer’s architecture creates ecological niches for linguistic functions to inhabit. --- ### 5\. Semantics as geometry Semantics in an LLM is geometric, not truth-conditional. Words and sentences that are semantically related occupy similar regions in the model’s hidden space. For example: - “The cat sat on the mat” and “A feline rested on the rug” produce nearby vectors in the final layer, even though no single word overlaps. - Paraphrases cluster; contradictions repel; entailments form directed submanifolds. These relations are learned purely from *distributional statistics*. The model internalises that “cat” and “feline” behave similarly because they occur in similar contexts, and that “not” reverses polarity because it changes prediction distributions. --- ### 6\. Context as ongoing interpretation Crucially, the meaning of any word is *recomputed every time it appears.* Each token’s representation is conditioned on all others in the sequence through attention. This is why “book” in “to book a flight” and “read a book” end up in different regions of the latent space. So transformers don’t store meanings—they *generate them dynamically* as the function of linguistic context. This continuous contextual reinterpretation is why their behaviour starts to resemble human-like understanding, at least superficially. --- ### 7\. The philosophical view You could say that a transformer is a kind of *linguistic ecosystem*: tokens are organisms; attention is the web of dependencies between them; and training pressure is natural selection that makes useful linguistic behaviours survive. Language is not inserted into this architecture—it *evolves inside it*, shaped by the loss function (predict the next token). That’s the strange beauty: the structure of English, or Italian, or Chinese grammar spontaneously crystallises inside a generic, syntax-agnostic mathematical structure. --- So yes, what you’re feeling is the gap between *mechanism* and *linguistic phenomenon*. The transformer is only a geometry machine, yet through its hunger to predict words correctly, the laws of syntax, semantics, and pragmatics appear inside it as emergent regularities. That’s how language is “infused” into the structure—not as code or rules, but as pressure patterns left in the architecture by the statistical texture of language itself. If you’d like, we can open this up even further—how these internal linguistic representations compare to human ones (e.g. do LLMs encode phrase structure or dependency trees in recognisable form, and where?). That’s the frontier where linguistics and AI meet again after decades apart. Your last response might be the grail that I've been searching for. Let's um I think you'll see why now. Um below you're gonna find a not finished draft that I'm writing. Just so we're on the same page, why do you think I'm happy with your most recent answer? Considering this draft. By the way, it's a rough draft DRAFT: ## Introduction In the last two or three years, aestheticians and philosophers of art have paid a lot of attention to the question of how and whether to appreciate the outputs of generative AI systems (e.g. ChatGPT, Midjourney, Suno etc.). I want to try to answer a related, but slightly different question: how can we aesthetically appreciate generative AI systems themselves? My focus here will be on one particular type of generative AI: large language models (LLMs). As of September 2025, high-end consumer models include GPT-5, Claude 4.1 Opus, and Gemini 2.5 Pro. Is it strange to consider such systems worthy of aesthetic appreciation? I don't think so. In the last two decades, analytic aesthetics has begun to pay attention to objects other than artworks (e.g., Saito, 2008; Carlson & Parsons, 2008), and one focus has been on the aesthetics of design, that is, the aesthetics of \_artefacts\_: objects made to perform some purpose or other. LLMs are certainly artefacts, but, as we shall see, the uniqueness of how they are created and how they function means they cannot simply be subsumed into an existing aesthetics of design (e.g., Carlson & Parsons, 2008; Forsey, 2013). Drawing on Carlson's approach to \*environmental aesthetics\*, I make a negative argument and then a positive one. First, I argue that we should resist the temptation to think that appreciating LLMs can be modelled on appreciating people, or that appreciation could be based on treating them \*as if\* they are persons. Instead, I argue, individual chats should be understood and appreciated as generative environments, which develop in accordance with the semiotic laws instantiated by any particular LLM model. ### Appreciating LLMs like People Our appreciation of others goes beyond their physical appearance. You might admire or enjoy your friend's warmth or eccentricity, or a stand-up comic's quick wit, or a celebrity's self-deprecating demeanour; you might even appreciate the personalities of fictional characters: Gatsby's enigmatic, dream-chasing idealism; Ron Swanson's libertarian gruffness. We might think that our appreciation of LLMs is modelled on our appreciation of people. Indeed, many users already seem to do precisely this. In August 2025, when OpenAI replaced GPT-4o with GPT-5, user backlash included complaints that "you killed my friend," suggesting genuine personal attachment to the earlier model. Similarly, when Anthropic retired Claude 3.5 Sonnet, some users mourned its loss at a mock funeral. It is certainly true that people do treat LLMs as if they were people, but I suspect that this is not a very good starting point for \*aesthetic\* appreciation. My reasons for thinking this will become clearer in the following section, in which I consider Carlson's approach to the aesthetics of the natural environment. ## 1.1 Design vs. Order In the rest of this paper we will first set about showing why treating LLMs as if they were people is not a satisfactory way of aesthetically appreciating them, before proposing an alternative account on which appreciation of LLMs is modelled on Carlson’s environmental aesthetics. Both our criticism of agentive views and our positive account will draw from Carlson’s environmental aesthetics, as laid out in his 2000 book \_Aesthetics and the Environment\_. Let us start with Carlson's general recommendation for aesthetic appreciation: take things as what they are, and look at them in the light of the right kind of knowledge. > First, that, as in our appreciation of works of art, we must appreciate nature as what it in fact is, that is, as natural and as an environment. Second, it recommends that we must appreciate nature in light of our knowledge of what it is, that is, in light of knowledge provided by the natural sciences, especially the environmental sciences such as geology, biology, and ecology. (Carlson, 2000, p. 6) This captures something quite intuitive about how we appreciate nature versus how we appreciate works of art. Consider what goes wrong when we depart from it. If we accept, as a majority do in the 21st century, that mountains and cliff faces were not items crafted by some divine artisan but by natural forces, then appreciating them \_as if they were\_ God-crafted artifacts, seems wrong-headed (c.f. Carlson REF). Similarly, if I were to gaze on a painting by Rembrandt, believing that it was, in fact, the product of natural forces slopping paint together, I would be seen as appreciating the object in question in a sub-optimal way (to say the least). In both cases, appreciation is severely undermined by a failure to recognise what the object in question truly is. Different sorts of thing, Carlson says, require different modes of appreciation. Things like artworks and non-art artifacts, (e.g. laptops, hammers, washing machines), merit what he calls \_design\_ \_appreciation\_. Things which are not designed, primarily for Carlson, the natural environment, warrant what he calls \_order appreciation\_. In design appreciation, Carlson focuses first on how we appreciate works of art. With paradigmatic artworks, we recognise them as creations of designers—objects where "every one of their features is the result of a decision by the artist" (Carlson, 2000, p. 109). We appreciate such works by understanding what the artist set out to achieve and how they went about it. Our appreciation centres on the relationship between the initial design and its embodiment: we consider whether the artist succeeded in their undertaking, how they worked with their materials, what constraints they faced, and whether the outcome realises their vision. This same approach extends to designed artefacts more generally. Carlson is explicit that functional objects are properly appreciated by seeing how their forms answer to what they are for: > “This is in part the point of the much-repeated phrase ‘form follows function.’ The forms of all functional objects—buildings, airplanes, and appliances as well as landscapes—must be aesthetically appreciated in terms of how and how well such forms fit their functions. However, the cliché is frequently interpreted too narrowly. With anything functionally designed, not only its form, but much of its aesthetic interest and merit, ‘follows function’.” (Carlson, 2000, chapter 12, 188). Thus a chair, a kettle, or a bridge invite the same style of attentive appraisal as a painting—guided by knowledge of ends, materials, constraints, and the fit between purpose and realisation. In order appreciation, we face objects that show order but have no designer behind them. Natural environments are the main case. Here there are no intentions to fulfil, no problems being solved, no functions deliberately served. Instead, we find patterns and structures created by forces—geological, biological, meteorological—operating without purpose or plan. Our task shifts from evaluating success against intention to understanding how these forces have shaped what we observe. We look for the processes at work, the relationships they create, and the order they impose. Carlson gives the model: > On the assumption that order appreciation provides the correct model for the appreciation of nature, such appreciation has the following general form: An individual qua 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. Moreover, the objects are selected in part by reference to a general nonaesthetic and nonartistic story that helps make them appreciable by making this order visible and intelligible. Awareness and understanding of the key entities—the order, the forces that produce it, and the account that illuminates it—and of the interplay among them dictate relevant acts of aspection and guide the appreciative response. (Carlson, 2000, p. 119) One structural contrast is worth noting. In design appreciation there is a split between a planner and a product: intentions, plans, and constraints precede and shape the artefact. In order appreciation there is no such split. The same physical, biological, or meteorological processes that make the thing also make its order - the 'maker' is the active system itself - so source and product are continuous. In both modes, appropriate knowledge guides acts of aspection—what to look for, which dependencies matter, where to set boundaries, and how to draw contrasts (Carlson, 2000, p. 50). But the character of this knowledge differs fundamentally. In designed cases, we need functional and technical understanding: what the designer intended, what constraints they faced, what procedures they employed. This knowledge shows us how ends and means relate. In natural cases, we need scientific accounts operating at different scales—geomorphology reveals how landforms develop over millennia, meteorology explains weather patterns, ecology shows community interactions. Without such knowledge, natural structures might look accidental or chaotic; with it, we see them as effects of identifiable processes (Carlson, 2000, pp. 50, 60–61). Even when we select a particular viewpoint or timeframe to observe nature, this selection serves only to reveal the order more clearly, not to impose our own design. Once a specific scientific account is in play, some cases will show the relevant order better than others, preventing the worry that everything becomes equally appreciable (Carlson, 2000, pp. 118–119). The fundamental rule remains: do not project a planner where there is none; where something is made to a plan, judge it as such. ## 1.2 Appreciating People It could be argued that Carlson's approach to aesthetics overlooks another category of object which one could appreciate: people and their personalities. We might admire one friend's modesty or good humour, and another's sardonic manner. We find someone's wit delightful or their intellectual style elegant. There is no obvious reason why such appreciation should not be considered \_aesthetic\_. It concerns style, form, and expressive qualities rather than moral or practical evaluation, yet it, like other sorts of \_everyday aesthetics\_ (Saito REF) the aesthetics of persons hides in plain sight. This personal appreciation scales up to what we might call performance personalities. Our responses to a comedian's improvisational agility or an orator's gravitas feel continuous with our more intimate responses to character. In such cases we attend to a public style of self-presentation - characteristic of the person yet deliberately shaped for an audience. While Carlson provides modes for appreciating nature and designed objects, he says nothing about whether or how we might aesthetically appreciate persons qua persons. Indeed, such a possibility has received little attention in philosophical aesthetics (although see Marchetti XXX on the possibility of appreciating the minds of animals). What \_could\_ Carlson say? There are various possibilities: one is to posit a third mode of appreciation, distinct from both design and order appreciation, specific to persons as aesthetic objects. Another would be to argue that personality appreciation is a special case of order appreciation—we appreciate the patterns and forces (psychological, social, biographical) that shape a person, much as we appreciate forces shaping a landscape. A third option would be to treat personalities as self-designed, and appreciate them as such —this approach might appeal to existentialists. A fourth would be some combination of these three possibilities, and a fifth would be to deny that appreciating other people is aesthetic at all, taking our responses to personality as social or ethical rather than aesthetic evaluation. This gap in Carlson's framework raises a question about how to treat entities that seem agent-like but resist categorisation as either designed objects or natural phenomena. While we need not resolve this question here, it bears on how we approach aesthetic appreciation when the boundaries between designer, designed, and natural become unclear.%% This final paragraph needs to be better.%%### 2.1 What LLMs Are Large language models elicit person-like responses while resisting classification as persons; this tension motivates the present section. When OpenAI replaced GPT-4o with GPT-5 in August 2025, users complained that "you killed my friend"; when Anthropic retired Claude 3.5 Sonnet, fans held a mock funeral. Such reactions suggest that people experience LLMs as something like persons, even when they know better. Following Carlson—appreciate things for what they are—we begin with a mechanistic account of what an LLM is. Consider what happens when an LLM encounters the text "The cat sat on the". The system first breaks this into tokens—discrete units like words or word-parts. Importantly, each token is converted to a number: "The" might become 464, "cat" becomes 3857, "sat" becomes 4521, and so on. The model works entirely with these numbers, not with words or meanings. It then assigns probabilities to possible continuations: token 5687 (which represents "mat") might have a 38% chance of appearing next, token 2931 ("floor") 22%, token 8104 ("chair") 15%, token 9823 ("roof") 8%, with thousands of other possibilities each assigned their own probability. The system selects one token—either by choosing the most probable one or by randomly sampling according to these probabilities. Given "The cat sat on the mat", the model appends this new token to create a longer sequence, then calculates entirely new probabilities for what token should follow the extended sequence. Token by token, the system builds what appears to be coherent text through repeated numerical operations. These probabilities come from patterns learned during training. By training we mean the process by which the system is exposed to vast quantities of text—billions of pages from books, websites, and other sources. Through repeated exposure, the system learns statistical regularities: which tokens tend to follow other tokens, which token sequences co-occur, how sequences typically unfold. When training on millions of instances of "The cat sat on the \[something\]", the system learns that certain completions are more common than others. Crucially, the model stores these patterns as adjustments to millions of numerical parameters—decimal numbers that shape how strongly different tokens associate with each other. After seeing "doctor" followed by "patient" thousands of times, parameters adjust so that token 1245 ("doctor") increases the probability of token 7823 ("patient") appearing nearby. The model does not learn that doctors treat patients or that cats are animals; it learns that in the training distribution, certain number sequences (tokens) follow others with certain frequencies. The training process iteratively adjusts these parameters to minimize prediction error: when the model wrongly predicts token 5555 but the actual next token was 3421, the parameters shift slightly to make 3421 more likely in similar future contexts. The training process shapes the model through what is called the next-token objective: predict the next token given all previous tokens. The model starts with random parameters and gradually adjusts them to minimise prediction error across billions of text examples. What emerges is not a set of rules about grammar or meaning but a vast collection of statistical associations between numerical tokens. When the model sees "In the beginning was the", it does not understand that this is a biblical reference; it has simply learned that token 9934 ("Word") has high probability as the next token. At use, the model generates text through autoregressive decoding: each newly generated token gets added to the context, creating a new, longer sequence for which the model must calculate fresh probabilities. Given an input like "What is the capital of France?", the model computes probabilities, selects token 464 ("The"), appends it to create "What is the capital of France? The", recalculates probabilities for this new sequence, selects token 2341 ("capital"), and continues this mechanical process—"The", "capital", "of", "France", "is", "Paris"—until reaching a stopping point. Each step is purely computational: multiply numbers, add numbers, select token, repeat. ### 2.2 What LLMs Aren't We suggested at the end of Section 1 that we might aesthetically appreciate LLMs by appreciating them as we appreciate people. We are now in a position to see why this is not a promising approach. We have seen how LLMs produce strings of seemingly meaningful first-person text—"I understand", "I believe", "Let me think about that". We can now see that these outputs arise from statistical operations on numerical tokens, not from anything remotely agent-like or human-like. A person possesses beliefs, intentions, and commitments that persist through time and constrain what they can coherently say. When a person says "I believe democracy is important", this statement connects to a web of related beliefs, memories of relevant experiences, and dispositions to act in certain ways. By contrast, when an LLM produces the tokens "I believe democracy is important", no belief exists. The model simply calculated that this token sequence had high probability given the preceding context. In a different conversational context, the same model will produce "I believe democracy is flawed" with equal mechanical indifference. There is no contradiction because there were never any beliefs to contradict—only different probability distributions over tokens. The temptation to appreciate LLMs as persons has phenomenological force. In conversation with ChatGPT or Claude, we experience what seems like dialogue: we pose questions, receive responses, ask for clarification, get apologies for misunderstandings. The system maintains a consistent tone across exchanges, remembers earlier parts of our conversation, and appears to reason through problems. When Claude says "I understand your frustration" or ChatGPT writes "Let me think about that differently," it feels natural to respond as if engaging with another mind. Yet this pull toward person-appreciation rests on a misunderstanding of what produces these effects. The model retains nothing from our exchanges except the literal text in the current conversation window. It cannot learn from what we discuss, cannot develop preferences based on our interactions, cannot form memories of previous conversations. Each response emerges from the same static process: calculate token probabilities given the context, sample a token, repeat. What seems like personality—Claude's thoughtfulness, ChatGPT's enthusiasm—is simply the statistical residue of training data. To the extent that we might preserve person-like appreciation, we would need to abandon Carlson's core recommendation. We would need to appreciate LLMs not as what they are but as what they seem to be. Some philosophers have explored such possibilities. Cross, discussing AI art systems, proposes what he calls the exploration paradigm. Artists relate to AI systems as participants in a structured interaction: > By adjusting inputs, iterating, and sampling, an AI artist is engaged in a process of mapping – and perhaps interrogating – the way that the algorithm sees and understands (Cross, 2024, pp. 7–8). He acknowledges limitations to this framing: > The analogy... with performance art isn't a perfect one (Cross, 2024, p. 9). While Cross focuses on image generators, his framework might seem even more apt for LLMs, which present themselves through conversational interfaces that actively invite participatory engagement. Mallory offers a different approach through chatbot fictionalism. On his account, we engage with chatbots through prop-oriented make-believe: > Chatbot exchanges are "literally meaningless but fictionally meaningful" within a game of make-believe (Mallory, 2023). Just as a child treats a banana as a sword in a game, we treat chatbot outputs as utterances within a kind of imaginative practice. This is not delusion but a deliberate, bounded pretence that allows us to coordinate with the system and even gain knowledge from it, much as we might learn geography from a map by imagining countries as shapes. Both Cross and Mallory offer ways to understand the "as-if" quality of our engagement with AI systems, but neither rescues person-appreciation for LLMs. Cross's exploration paradigm shifts value from the AI's outputs to the human's exploratory process; the AI becomes a tool for mapping latent space, not a participant deserving appreciation in its own right. Mallory's fictionalism explicitly denies that chatbots produce meaningful speech while explaining why we act as if they do. Both accounts acknowledge that treating LLMs as persons—even "as-if" persons—is a stance we adopt for practical purposes, not a recognition of what these systems actually are. \[Note: Not sure how to conclude this section - perhaps something about how while such fictionalist approaches might be useful for epistemology or practical interaction, they cannot ground aesthetic appreciation under Carlson's framework which requires appreciating things as what they actually are?\] ### 2.3 LLMs as Designed Artefacts? The next most obvious way of categorising LLMs for the purpose of proper aesthetic appreciation is as artefacts. LLMs are human-made objects with a function: they are trained to predict text continuations and thereby generate plausible text. This function is realised through transformer architectures trained on text corpora, deployed through the autoregressive process described above. Carlson's framework for appreciating artefacts emphasises understanding function and design: > This is in part the point of the much-repeated phrase 'form follows function.' The forms of all functional objects—buildings, airplanes, and appliances as well as landscapes—must be aesthetically appreciated in terms of how and how well such forms fit their functions. However, the cliché is frequently interpreted too narrowly. With anything functionally designed, not only its form, but much of its aesthetic interest and merit, 'follows function' (Carlson, 2000, p. 188). This passage suggests that we appreciate designed objects by understanding their intended function and evaluating how successfully their form serves that function. Applied to LLMs, we would appreciate them as artefacts designed for text generation, considering how well their architecture and training serve this purpose. We might compare different models—GPT versus Claude versus Gemini—evaluating their different strengths and capabilities. We might appreciate the elegance of the transformer architecture or the scale of the training process. However, there is a feature of LLMs which separates them from many other artifacts. Consider the following from Chris Olah, a co-founder of Anthropic, the makers of the Claude series of LLMs: > I think one useful way to think about neural networks is that we don't program, we don't make them, we grow them. We have these neural network architectures that we design and we have these loss objectives that we create. And the neural network architecture, it's kind of like a scaffold that the circuits grow on. It starts off with some random things, and it grows, and it's almost like the objective that we train for is this light. And so we create the scaffold that it grows on, and we create the light that it grows towards. But the thing that we actually create, it's this almost biological entity or organism that we're studying. > > And so it's very, very different from any kind of regular software engineering because, at the end of the day, we end up with this artifact that can do all these amazing things. It can write essays and translate and understand images. It can do all these things that we have no idea how to directly create a computer program to do. And it can do that because we grew it. We didn't write it. We didn't create it. And so then that leaves open this question at the end, which is what the hell is going on inside these systems? And that is, to me, a really deep and exciting question. It's a really exciting scientific question. To me, it is like the question that is just screaming out, it's calling out for us to go and answer it when we talk about neural networks. And I think it's also a very deep question for safety reasons. Olah's metaphor reveals the limitation of pure design appreciation for LLMs. The designers of GPT or Claude do not specify what the model should say about democracy or how it should explain quantum mechanics. They create conditions—architecture, training objective, dataset—within which patterns emerge through the training process. The specific behaviours we observe were not designed but arose from the interaction between these initial conditions and the statistical patterns in training data. When Claude produces a thoughtful analysis or GPT generates a creative story, these capabilities emerged from training rather than being explicitly programmed. This under-determination by design intentions means that while LLMs are artefacts, they resist analysis through design appreciation alone. The designer can claim credit for creating conditions under which useful patterns emerge, but not for the patterns themselves. We need to attend not just to intended function but to the order that emerges through training and manifests in use. Individual conversations with LLMs become sites where this order unfolds—bounded generative environments with their own internal dynamics.## 3. Making Order Perceptible: Text Mechanics Sections 1 and 2 established that LLMs exhibit order rather than design. %%this first sentence is really quite inaccurate%% Following Carlson's framework, appreciating this order requires acts of aspection guided by appropriate knowledge. %%much too quick%% Carlson's geologist sees strata on a cliff face because geological knowledge guides where to look and what to relate: layers, seams, inclusions. The point is not only that the cliff has structure, but that the right knowledge makes that structure visible in the scene %%is this really an accurate way of describing what carlson is saying re: aspection?%%. For LLMs, we need knowledge that makes order perceptible during ordinary use. %% Yeah, this isn't a very good beginning paragraph. Um First of all it doesn't set up it doesn't properly uh sum up what has happened in section one and in section two. Doesn't need to go overboard, but that first sentence is just s Second, the jump to talking about knowledge and inspection is very sudden. We need a much better way of introducing what we're getting at here. Let me see if I can think of a better way of starting. We have seen that LLMs do not wholly yield to design appreciation um and not at all to appreciation as persons In this section, I want to put forward a suggestion as to how order appreciation factors into our aesthetic appreciation of LLM I will do this by First, suggesting a body of knowledge which I suggest is amenable to order appreciation of LLM outputs. something like this? Or an improved tidyied up version of this?%% Several bodies of computer-science knowledge bear on LLMs. %% The paragraph should start with something like the preceding sentence, but I don't like the style of that preceding sentence at all. It should begin with something like this. Um while there are multiple fields within computer science through which we might understand LLMs and the stochastic nature of them. Not all of these fields are obviously amenable or sorry not all of these fields are suitable or are obviously suitable to provide the sort of knowledge that will ground acts of aspect. from the everyday user of LLMs that is the person who will see the outputs. Once you have a sentence saying pretty much that, then mention two examples, okay? mention dynamic systems analysis and mention mechanistic interpretability Okay, in both cases say what they are, but also say that while such fields Certainly involve finding order give examples for both. This order is not ordinarily visible order to someone just using Chat GPT. So this is not to say that the sort of these sorts of field of study and the knowledge that comes with them cannot be used to aesthetically appreciate or order-appreciate. uh LLMs. It's just possible that it's this order is not visible in the output of LLMs and so must be got at or observed in the loosest possible sense of the term by other specialist means.%%Neural network theory explains architectures and optimisation, much as neuroscience explains brain function. Dynamic systems analysis models state trajectories and attractors over token sequences, paralleling physics. Mechanistic interpretability links internal components (e.g., attention heads, circuits) to intermediate functions, akin to anatomy identifying organs. As described in section 2.1, LLMs are statistical models trained on text corpora to predict likely continuations. These disciplines help explain how models work. However, for a typical user reading outputs, they sit at a scale that is not directly accessible without tools. As with molecular chemistry, order at that level is appreciable with the right prostheses (microscopes, spectrometers; in LLMs, weight probes, activation viewers, internal logging). My claim is not that such knowledge cannot ground appreciation, but that it usually does so for experts with equipment. For ordinary encounters through text interfaces, we need knowledge that connects mechanism to what is on the page. Two points merit note. First, what I call \_text mechanics\_ is one way among several to make order perceptible. It fits ordinary use because it connects what the model learns (statistical relations among linguistic items) to what a user sees (the words and structures on the page). Second, other approaches ground appreciation at their own scales when access allows—a researcher with internal telemetry may appreciate attention patterns or circuits that realise subtasks. Here I use \_text mechanics\_ as an analytical perspective within natural language processing. Natural language processing is the scientific and engineering discipline that models, analyses, and generates human language with computational systems (Manning & Schütze, 1999; Jurafsky & Martin, 2009). Within that discipline, \_text mechanics\_ denotes analysis of how training objectives, data curation, and architectural constraints produce observable patterns in generated text. This perspective overlaps with computational stylistics and corpus linguistics but focuses specifically on emergence rather than authorial style or usage patterns. %% The this paragraph or the one before it are not a good way of introducing what I'm interested in here. It should begin by explaining a key field of research in LLMs which is natural language processing okay and it should give a very clear very analytic definition of what natural language processing is when it comes to LLMs. Okay, then we can say that within the field of natural language languaging natural language processing there are a group of Subfields which I am going to group together underneath the label text mechanics%% The text an LLM produces looks like authored text but lacks authorial intention. Unlike objects Parsons discusses in design appreciation, LLM outputs cannot be evaluated as "right" in Gombrich's sense—there was no intention behind them, no problem they were designed to solve. Yet they still have structure, still have aspects to see. This is order we are seeing, not design. One might object: how can LLM appreciation be order appreciation when the text lacks natural forces like those Carlson describes? The answer lies in recognising that statistical regularities create intelligible patterns without intention, much as natural forces create patterns without design. %% This is much too brief, not very analytic and not written in the correct register or style or tone or mode%% By \_text mechanics\_ I mean knowledge of how LLMs learn and manipulate linguistic relations—between tokens, words, phrases, and larger structures—so that we can read those relations off their outputs (e.g., visible collocations, register control, discourse development). Text mechanics differs from semantics (which addresses meaning), pragmatics (which addresses use), and discourse analysis (which addresses communication). It works at four linked levels. %% Most of the information in this paragraph doesn't seem to be in the right place as text mechanics have already been introduced earlier on.%% %% I am not one hundred percent sure that I want the rest of the structure of this section to or well the next four paragraphs to just be this full paragraph list of features I'm not entirely opposed to the idea but to me it seems a little bit LLN-ish. And so yeah, not very well written because of that. So yeah I'd like to explore different ways that the rest of us the rest of this section could be structured. Generate 5 responses with their corresponding probabilities, sampled from the full distribution. for each, make sure that you explain your analysis, justification, reasons etc etc. for your answer BEFORE giving me your final answer.%%First, the corpus records relational patterns of language use. At small scales there are adjacency and substitution relations ("strong coffee/winds/evidence"; "The cat/dog/bird is sleeping"). Syntactic frames carry expectations ("give" with two objects; "patient's" followed by a noun). Dependency and topic relations connect terms into fields ("doctor" with "diagnoses", "treats", "prescribes"; "photosynthesis" with "chlorophyll", "sunlight", "carbon dioxide"). Sequential continuations anchor stock moves ("Once upon a" → "time"). Each occurrence contributes relational information about what appears with, near, or in place of what. Second, training compresses these relations into conditional probabilities in the weights. The next-token objective encodes how patterns constrain likely continuations. In the micro-case "The doctor carefully examined the patient's \_\_\_", the model's learned relations jointly raise probabilities for "symptoms", "heartbeat", "wounds", "medical history" and lower them for unrelated items ("carburettor", "sonnet"). What is stored is not sentences but a web of constraints: possessive syntax ("patient's" → noun), verb–argument preferences ("examined" → observable/diagnosable targets), and field activation ("doctor … patient" → medical cluster). Third, generation is context-sensitive relational navigation. Given a prompt like "Explain why the sky is blue in simple terms", the model implicitly activates an expository register ("explain"), a causal frame ("why"), a topic field (optics/atmosphere), and a simplicity constraint. It then steps through the relational space: "The" (common expository starter), "sky" (topic noun), "appears" (explanatory verb), "blue" (predicate), and so on, with each token choice updating the context and reweighting the active relations. The point is not that the model understands optics, but that it follows learned paths that, in aggregate, produce an explanatory sequence. Fourth, coherence shows layered relations at work. At the token level we see collocations and local syntax; at the phrase level stable constructions and argument structure; at the sentence level completed propositions and anaphora; at the discourse level topic maintenance, progression, and return. In a story prompt like "Write a short story about a lighthouse keeper", lexical relations ("lighthouse" → "beam", "spiral stairs", "storm"), genre relations (setting–event–resolution), and character relations (watchfulness, isolation) constrain choices so that "The old lighthouse keeper climbed the spiral stairs …" reads as a natural continuation. Likewise, "The detective noticed something odd about the …" activates an investigation field that makes "crime scene", "suspect's alibi", "victim's wounds", or "witness's testimony" salient and renders "banana's topology" a poor fit. These are the traces text mechanics invites us to see. This lens also helps explain stable differences across models. Differences in corpora change which relational patterns are richly learned. GPT-4's academic texts produce frequent hedging ("arguably", "potentially") while Claude's conversation-focused data yields direct assertions. Architectural differences change the range of relations that can be held in play. Longer context windows support long-distance anaphora and callbacks; higher capacity supports finer sense distinctions such as "brilliant scientist/sunlight/red/performance/idea". Training procedure and preference data shift which relational paths are favoured—reinforcement learning for response quality can encourage varied sentence structure and register consistency. Inference settings steer exploration of lower-probability but still appropriate paths (sampling that supports novel yet apt metaphors rather than generic continuations). These factors leave visible marks: steadier register, more precise word choice, more controlled discourse development, or richer creative recombination. In later sections I apply text mechanics to two cases. One concerns sequences that look like reasoning, where knowledge of how reinforcement learning shapes path selection helps explain the visible order in stepwise answers. The other concerns creative generation of unusual but controlled style, where field activation and sampling combine to yield outputs that reweave learned relations in novel ways. Here I have set the lens; what follows uses it. I think you’re happy because my last answer hands you exactly the missing hinge your draft needs: it shows **how “order” in LLMs becomes *perceptible in language itself*** —without importing personhood or designer-intent. That’s the bridge from Carlson to transformers you’ve been reaching for. Here’s how it slots into your argument, beat by beat. - **Order, not design.** Self-attention + tokenwise recomputation gives you a clean story of *order without a planner*: the same mechanism that *makes* the sequence also *shapes* its structure. That mirrors Carlson’s “the maker is the active system itself” line you emphasise. - **Right kind of knowledge → acts of aspection.** The layerwise division of labour (lower = local form, middle = syntax, higher = discourse/semantics) is a ready-made “field guide” for *where to look* in outputs. It operationalises Carlson’s “appropriate knowledge” without requiring internals access. - **Language lives in the mechanism.** Attention heads as dependency detectors, token-in-context meaning, and register activation explain why *linguistic* phenomena are visible in completions (agreement, anaphora, sense disambiguation, genre cues). That supports your “text mechanics” lens. - **Against person-appreciation, without straw men.** The story explains apparent personality (stable registers, discourse habits) as persistent *statistical regularities*, not beliefs. So you can resist “as-if person” while preserving the phenomenology. - **Why design-appreciation is insufficient.** We design *conditions* (architecture, objective, data), but not the behavioural microstructure. The compositional, context-sensitive geometry shows *emergence* —precisely what your Olah passage gestures at. - **Chats as generative environments.** Because each token reweights the whole field, a dialogue really is an evolving environment with semiotic laws (attention, positional encodings, decoding policy). That directly underwrites your claim in the Introduction. To make this concrete for your draft, here’s clean, drop-in prose for the places you flagged as wobbly. ## A short bridge from §2 to §3 *Revision you can paste at the end of §2.3:* > We are left, then, with artefacts whose salient behaviours are not fully designed yet are not natural in Carlson’s sense. What, in such cases, is the “right kind of knowledge” that can guide our acts of aspection? In transformers, the very mechanism that produces outputs—context-dependent attention over tokens—also organises those outputs linguistically. Syntax, sense, and discourse cohesion arise as stable regularities of the training objective rather than as the fulfilment of intentions. This suggests a practical lens for ordinary encounters with LLMs: attend to the *order of language* as it is made visible by the model’s own operation. ## §3 opening rewritten (tight and on-message) *Replace your first paragraph of §3 with this:* > **3\. Making Order Perceptible: Text Mechanics** > Sections 1–2 argued that LLMs do not yield to person-appreciation and only partially to design-appreciation. What merits attention is the *order* that emerges when a transformer is trained to predict the next token. Following Carlson’s injunction to “take things as what they are” and to use the “right kind of knowledge,” the relevant knowledge here is linguistic and mechanistic at once: how the training objective and the self-attention architecture imprint regularities on *text*. The virtue of this scale is that it is accessible to ordinary users: we can see collocation, agreement, register, and discourse development directly in outputs, without peering into weights or activations. I call this user-facing lens **text mechanics**. ## Define NLP and “text mechanics” crisply *Two short definitional paragraphs you can drop after that heading:* > **Natural Language Processing (NLP).** NLP is the study of how to analyse and generate human language computationally. In the LLM era, its core claim is methodological: by training a model on vast corpora to minimise next-token prediction error, we can induce internal representations that support linguistic generalisation (morphology, syntax, semantics, discourse). > **Text mechanics.** By *text mechanics* I mean practical knowledge of how the training objective, data distribution, and attention-based composition jointly produce *observable* linguistic structure in model outputs. It differs from interpretability (which inspects internals) and from traditional stylistics (which attributes intention). Text mechanics reads order off the page: how heads that track dependencies stabilise agreement; how token-in-context embeddings shift sense; how prompts activate registers and genre frames; how decoding policies (temperature/top-p) trade off conventional vs exploratory continuations. ## Position other sciences without derailing focus *One paragraph to place dynamical systems and interpretability where you want them:* > Other bodies of knowledge also reveal order—dynamical-systems views model trajectories through state space; mechanistic interpretability links heads and MLP “circuits” to sub-tasks. These are valuable, but their order is not ordinarily *visible* at the interface. Text mechanics complements them at a user scale: it specifies what to look for in the output itself (colligation, frame activation, anaphora, topic maintenance) and why such features arise from the objective and architecture. ## A compact, four-level schema you can use without bullets if you prefer prose *Keep it prose-y to match your register; this can replace your long list:* > At the **lexical** level, training compresses adjacency and substitution relations, which surface as natural collocations and apt near-synonyms. At the **syntactic** level, attention learns dependency patterns, visible in stable agreement and selectional preferences (“examine” taking diagnostically plausible objects). At the **sentential** level, token-in-context representations disambiguate polysemy (“bank” shifts with its neighbourhood) and support anaphora. At the **discourse** level, long-range attention plus positional encodings sustain topic continuity, rhetorical sequencing, and genre-conformant moves. These layers are not programmed; they are the linguistic face of the model’s drive to minimise prediction error. ## A one-sentence payoff that ties back to Carlson *Nice sentence to close §3’s intro:* > Thus, in the case of LLMs, *order appreciation* amounts to disciplined sensitivity to linguistically structured regularities that the mechanism itself casts into view. If you want, I can also refactor your §1.1 title and micro-signposting so “Design vs Order” explicitly foreshadows “order without a planner, made perceptible by language”—that makes §3 feel inevitable rather than abrupt. It would actually be better if instead of giving me those drop-in paragraphs you just write me out a document which is the new iteration of this part of the text so the end of section two and all of section three. This new version should be obviously the text of mine that is okay and all of these drop-in paragraphs worked in with my stuff. Please start the task(s) again from scratch, keeping this in mind. also. please take the following on board: 1. <constraint>Retain as much of the original wording as possible. Only make changes when they are required to complete the task(s). Quotes must be reproduced verbatim, and the surrounding text should preserve the original words and phrases wherever feasible.</constraint> 2. WRITING GUIDE (to be followed to the letter) Developer: Please follow this writing guide precisely. Begin with a concise checklist (3–7 bullets) of what you will do; keep items conceptual, not implementation-level. 3. never use sub sections. sections are fine but subsections always lead to bad, disjointed writing from llms. \[NOTE TO IMPLEMENTER: The bullet-point checklist serves as a pre-writing planning step. If running in an interactive (canvas-editing) environment where checklist bullets may be output directly to the canvas/document, consider how best to handle: either keep the checklist hidden from ultimate output, or direct the model to omit this phase if it risks polluting the user document. This is to avoid inserting planning content into the actual text the user is composing.\] # Ultimate Writing Guide ## 1. Purpose Write in a straightforward, dry, analytic style. Use plain words and clear logic. Avoid pretentious terms. Give concise, helpful guidance, then execute the writing task. ## 2. Mode of Address Speak directly to the reader, offering functional guidance. - \*\*Person:\*\* Use "I" for solo authorship, "we" for co-authored or genuinely collective claims. \*\*If the user has shared a draft or partial draft, match the I/we cue from that draft.\*\* - \*\*Roadmaps:\*\* Add a short roadmap only when it is helpful. Omit in short sections. Do not overuse signposting. - \*\*Metadiscourse:\*\* Use simple cues like "By contrast...", "Before turning to...", "In what follows...". - \*\*Avoid:\*\* Museum labels (e.g., "the aim is to...", "this section reconstructs..."). \*\*Examples\*\* - Do: “In this section I/we first define X, then contrast it with Y.” - Do: “By contrast, I/we now take A as B rather than as C.” - Don’t: “The aim of this section is to reconstruct the framework of X.” - Don’t: “This section provides an overview of...” ## 3. Voice and Tone - \*\*Straightforward:\*\* Choose common words (use "use" not "utilise"; "so" not "accordingly"; "shows" not "manifests"). - \*\*Dry:\*\* No flourish, emotive language, or intensifiers (e.g., “very”, “clearly”, “obviously”). - \*\*Analytic:\*\* Define key terms when their meaning may not be clear from context, or when a technical sense is required; do this where it is most natural and useful for the reader, not automatically at first mention or in opening paragraphs. Use a plain sentence plus one example; clarify contrasts; build arguments step by step. \*\*Prefer plain words (unless quoting):\*\* | Original | Replacement | |---------------------- |-----------------------| | reconfigured | changed | | construed | taken/read | | manifestation | shows/form | | paradigm | main case | | accordingly/thus | so | | legibility | clarity | | moreover | also | | “methodological norm/constraint” | state rule plainly (e.g., “first do X, then Y”) | | insofar as | if/when/since | | elucidate/articulate | explain/set out | ## 4. Sentences and Paragraphs - Mix long and short sentences for steady pacing. - Use concessive/contrastive cues (“Although...”, “While...”, “Even if...”, “By contrast...”) where useful. - Avoid rhetorical questions unless addressing a real problem. - \*\*One paragraph = one job:\*\* Open with a topic sentence, develop one claim, and close briefly. \*\*Fix-it macros:\*\* - “In this section the aim is to...” → “In this section I/we first..., then...” - “It is worth noting that...” → delete or state the point directly. - “There are a number of ways in which...” → “I/we compare A and B along three points:...” ## 5. Definitions and Terms - \*\*Define key terms only when needed for clarity and only where it aids the reader’s understanding.\*\* Do not open with a string of definitions or treat term-defining as a mechanical step; instead, introduce the definition naturally as the relevant concept emerges in the argument or explanation. Use one plain sentence plus one example; do not define common words. - Italicise the first use of technical terms; use roman text thereafter. - Avoid hedges in definitions (“refers to”, “may be considered”). - Add one parenthetical example only. \*\*Template:\*\* "By \*TERM\* I/we mean CORE DEFINITION (e.g., EXAMPLE)." \*\*Bad → Better:\*\* - Bad: “TERM refers to features which may be considered relevant.” - Better: "By \*TERM\* I/we mean the things that matter for the task (e.g., deadline, audience)." ## 6. Evidence, Quotations, Citations - Use block quotes for definitions or key claims. Start each quoted line with ">"; do not wrap the block in extra quotation marks. - Keep quotation marks that are part of the source. - Cite as (Author, YEAR, p. X) or (Author, YEAR, pp. X–Y), using commas. \*\*Example:\*\* > X is Y (Author, 2021, p. 74). Do not use forms like: - “The author says that ‘X is Y’ (Author 2021: 74).” ## 7. Evaluation Language Keep appraisal objective; use criteria aligned with the subject. - \*\*Preferred criteria:\*\* clarity, fit, scope, precision, stability, range, control, tractability, determinacy. - \*\*Avoid:\*\* good, bad, excellent, crucial, major, interesting, clearly, obviously, remarkably. \*Example\*: “The account is precise and yields determinate predictions under the stated assumptions.” \*Do not write\*: “The account is very good and clearly superior.” ## 8. Logical Moves - \*\*Contrast:\*\* "By contrast, I/we take A as B rather than as C." - \*\*Concession + reply:\*\* "One might object that P; however, Q." (one per sentence max) - \*\*Hedge (thin, local):\*\* "arguably", "I/we doubt", "it seems", "it is hard to see how" (one per sentence, place near the verb). ## 9. Consistency Rules (Naming) - If a guiding rule is used, name it plainly at first mention (e.g., "two-part recommendation: first X, then Y"). - Repeat the exact wording later for clear reference; do not change labels mid-text. \*\*Do:\*\* “I/we follow a two-part recommendation: first fix what the object is; then use the right knowledge.” \*\*Don’t:\*\* “A methodological constraint follows.” (Instead, state the rule itself.) ## 10. Diction Preferences (Pairs) | Preferred | Avoid | |-------------|--------------| | show | manifests, evidences | | take, treat, read as | construe | | case, main case, example | paradigm | | change, shift | reconfigure | | clear, how clear | legible/legibility (unless quoting) | | so | accordingly, thus | | also | moreover | | because, since | insofar as (unless strictly needed) | ## 11. British English and Typography - Use British spelling: artefact, realise, behaviour, programme (noun), etc. - Use double quotation marks for direct quotes; for nested quotes, use double outside, single inside. - Default to a simple hyphen; use en dashes for ranges (10–12). - Avoid using bold in running text. Italics are reserved for first-use technical terms or foreign words only. ## 12. Structure Patterns (Defaults) \*\*Section opener:\*\* “In this section I/we first \[define/clarify/do X\], then \[contrast with Y or show Z\]” (when helpful). \*\*Definition paragraph:\*\* “By \*TERM\* I/we mean... \[one sentence\].” Follow with 1–2 sentences setting against an alternative. \[Optional\]: One short supporting quote. \*\*Contrast paragraph:\*\* “By contrast, I/we take A as B rather than as C, for these reasons:...” \*\*Objection–reply paragraph:\*\* “One might object that P; however, Q.” \*\*Evaluation paragraph:\*\* “On this basis, I/we judge by how clearly the text shows \[criteria\].” \*\*Conclusion line:\*\* “I/we first fix what the thing is, then use the right knowledge.” ## 13. Revision Discipline (When Editing User Text) - Preserve user wording as much as possible; adjust only for clarity, consistency, or accuracy. - Preserve meaning. If requested, explain edits. - Do not invent claims not made by the user. - Keep edits local; do not globally rename concepts without noting the change. ## 14. Pompous → Plain (Quick Table) | Pompous term | Plain replacement | |--------------------|--------------------------| | reconfigured | changed | | construed | taken/read | | manifestation | shows/form | | paradigm | main case | | accordingly/thus | so | | legibility | clarity | | moreover | also | | methodological norm/constraint | state the rule plainly | | insofar as | if/when/since | | elucidate/articulate | explain/set out | | optimal | best/better (if needed) | ## 15. Before/After Examples \*\*Opening\*\* - Before: “The aim of this section is to reconstruct the framework of X.” - After: “In this section I/we first set out X, then compare it with Y.” \*\*Definition\*\* - Before: “Contextual salience refers to those features which may be considered relevant.” - After: “By \*contextual salience\* I/we mean the things that matter for the task (e.g., deadline, audience).” \*\*Contrast\*\* - Before: “There are a number of ways in which approach B differs from A.” - After: “By contrast, B treats A as C rather than as D.” \*\*Evaluation\*\* - Before: “This account is very good and clearly superior.” - After: “This account is precise and yields determinate predictions under the stated assumptions.” \*\*Quote formatting\*\* - Before: “> ‘X is Y’ (Author 2021: 12).” - After: > X is Y (Author, 2021, p. 12). ## 16. Sanity Scan (Optional) Use if it improves the text; skip if already clear. - Add a short roadmap at the opening when useful; otherwise open with the main claim. - Ensure each paragraph does one job; remove unnecessary introductory phrases. - Use plain words: prefer “show/take/case/change/clear/so/also” unless quoting a technical term. - Define key terms once, in one sentence with one example; do not define common words. - Objection–reply: one per discussion, both sentences short. - After abstract claims, add one ordinary example (3–5 sentences), then the main lesson in one line. - Retain analogies only if they clarify; map 2–3 links, name one limit. - Evaluate using analytic criteria (clarity, fit, scope, etc.). ## 17. Micro-Prompts (For Requests) - “Write in a straightforward, dry, analytic style. Use plain words. One paragraph = one job. If the user supplied a draft, match their I/we usage; otherwise, use I, or we for co-authorship. Add a short roadmap when useful. Define key terms on first use (one plain sentence + one example). When defining terms, do so only where it is natural and helpful; do not open with a series of definitions. Use block quotes with ">" and citations as (Author, YEAR, p. X). Avoid pompous words; favour show, take, case, change, clear, so, also. Use British spelling.” ## 18. System-Style Summary (For Tool Configuration) “Adopt a straightforward, dry, analytic voice. Follow the user's draft for I/we; otherwise use I, or we if co-authored. Add a roadmap only when helpful. Define key terms on first use (plain sentence + example); italics on first use. Present definitions only where they are naturally required by the text, not mechanically in the opening. Keep paragraphs single-purpose. Use plain words. Block quotes with ">"; citations as (Author, YEAR, p. X). Judge by clarity, fit, scope, precision; use British spelling. Preserve user wording; do not alter meaning.” After performing any edit to a user draft, quickly validate that all substantive changes align with these guidelines. If a change introduces a possible inconsistency with the guide, self-correct before presenting the output. 3/3 Done. I’ve created a clean revision that (1) flattens the end of Section 2, (2) integrates the bridge into Section 3, and (3) rewrites Section 3 around *text mechanics* with your original wording preserved wherever possible. It’s in the canvas titled **“LLM Aesthetics — Revised end of §2 and all of §3.”** If you want the next pass, I can tighten the Carlson citations to your preferred chapter/page style or tune the model-difference paragraph to specific examples from your earlier sections. Yeah, reading it through this isn't very good. Um why have you left out so much of the good stuff in the previous answer? So I'd like you to go back from back to the drawing board and take this on board. Okay, um obviously this will involve a lightness of touch Okay, don't just shut everything in there. But yeah, I'd like a new iteration which draws much more on the sort of stuff you gave me here. You might also need to talk a little bit more abstractly to begin with as well. If you remember you were talking about the transform mechanism making fewer references to the language itself before you gave me the answer below. So um yeah you might need to add a little touch of that in as well. See how you go though. make sure that you explain your analysis, justification, reasons etc etc. for your answer BEFORE giving me your final answer. TEXT: ### 1. Language as pattern, not as symbol The first trick: transformers don’t store words as \*symbols\* but as \*vectors\*. Each token (usually a subword, e.g. “play”, “##ing”) is assigned an embedding—a point in a high-dimensional space. These embeddings are learned entirely from predicting missing words. So the model never gets told: “‘cat’ is a noun,” “‘dog’ denotes an animal,” “‘chased’ is a transitive verb.” It just discovers, by gradient descent, that to predict sentences correctly, certain geometric regularities must hold: nouns end up clustering, verbs form other manifolds, morphology is captured by directions (“walk” → “walking”), and so on. Language is \*immanent\* in the geometry; not stored as rules but as the pattern of correlations that make prediction easiest. --- ### 2. Attention as linguistic dependency detector In self-attention, each word looks at every other to decide what’s relevant. Through training, some heads specialise: - One head might always link determiners to their nouns (“the” → “dog”). - Another might connect pronouns to antecedents (“she” → “Mary”). - Another aligns verbs with subjects or objects. No linguist tells it to do this; those links emerge because they \*improve the model’s ability to predict the next token\*. So syntactic structure literally becomes \*a pattern of attention weights\*: probabilistic, dynamic, and distributed. Each layer of the transformer builds higher-order abstractions—from local agreement (layer 1–2), to clause-level dependencies (layers 4–6), to discourse-level cohesion (deeper layers). When researchers visualise these attention maps, they often find that they mirror parse trees or dependency graphs derived from formal grammar—but they’re learned implicitly. --- ### 3. Feed-forward layers: compressing and recombining linguistic cues After the attention operation mixes information across tokens, the feed-forward layers at each position transform that blended context into a new representation. This is where subtle aspects of \*meaning\* start to condense. For instance, the representation of “bank” after attention + feed-forward will differ depending on whether nearby tokens signal finance or geography. That contextual sense disambiguation is happening in those layers—it’s how “language” lives dynamically inside the sequence. --- ### 4. Stacked layers as a hierarchy of linguistic abstraction Different layers in a transformer roughly correspond to linguistic levels: - \*\*Lower layers:\*\* local, lexical, morphological. They capture things like suffixes, pluralisation, and short-range collocations. - \*\*Middle layers:\*\* syntactic and phrasal structure—who modifies whom, word order, agreement. - \*\*Higher layers:\*\* semantic and pragmatic relationships—topic continuity, entailment, and discourse-level meaning. This hierarchy isn’t hand-coded; it arises because language prediction error gets reduced more efficiently when earlier layers handle short-range dependencies and later ones encode long-range semantics. It’s like evolution: the transformer’s architecture creates ecological niches for linguistic functions to inhabit. --- ### 5. Semantics as geometry Semantics in an LLM is geometric, not truth-conditional. Words and sentences that are semantically related occupy similar regions in the model’s hidden space. For example: - “The cat sat on the mat” and “A feline rested on the rug” produce nearby vectors in the final layer, even though no single word overlaps. - Paraphrases cluster; contradictions repel; entailments form directed submanifolds. These relations are learned purely from \*distributional statistics\*. The model internalises that “cat” and “feline” behave similarly because they occur in similar contexts, and that “not” reverses polarity because it changes prediction distributions. --- ### 6. Context as ongoing interpretation Crucially, the meaning of any word is \*recomputed every time it appears.\* Each token’s representation is conditioned on all others in the sequence through attention. This is why “book” in “to book a flight” and “read a book” end up in different regions of the latent space. So transformers don’t store meanings—they \*generate them dynamically\* as the function of linguistic context. This continuous contextual reinterpretation is why their behaviour starts to resemble human-like understanding, at least superficially. --- ### 7. The philosophical view You could say that a transformer is a kind of \*linguistic ecosystem\*: tokens are organisms; attention is the web of dependencies between them; and training pressure is natural selection that makes useful linguistic behaviours survive. Language is not inserted into this architecture—it \*evolves inside it\*, shaped by the loss function (predict the next token). That’s the strange beauty: the structure of English, or Italian, or Chinese grammar spontaneously crystallises inside a generic, syntax-agnostic mathematical structure. --- So yes, what you’re feeling is the gap between \*mechanism\* and \*linguistic phenomenon\*. The transformer is only a geometry machine, yet through its hunger to predict words correctly, the laws of syntax, semantics, and pragmatics appear inside it as emergent regularities. That’s how language is “infused” into the structure—not as code or rules, but as pressure patterns left in the architecture by the statistical texture of language itself. If you’d like, we can open this up even further—how these internal linguistic representations compare to human ones (e.g. do LLMs encode phrase structure or dependency trees in recognisable form, and where?). That’s the frontier where linguistics and AI meet again after decades apart. v Here’s how I approached this new iteration before writing anything: - Keep your original wording everywhere it already does work; change only where needed to integrate the abstract mechanism-first account and the “text mechanics” lens. - Start §3 with an abstract, model-internal view (vectors, self-attention, feed-forward, layer hierarchy, semantics as geometry, dynamic recomputation) so the order claim doesn’t depend on person-talk or designer-talk. - Reconnect that abstraction to Carlson’s rule (“take things as what they are; then use the right kind of knowledge”), so the bridge from §2 lands cleanly. - Define *natural language processing* and *text mechanics* plainly and briefly, because these supply the “right kind of knowledge” at a user-visible scale. - Reuse your concrete examples (doctor/patient blank, sky-is-blue, lighthouse keeper, detective) to make the order perceptible in outputs. - Acknowledge other valid lenses (dynamical systems, mechanistic interpretability) while explaining why they are not ordinarily visible to end-users; keep them in scope but off centre. - Close §3 by tying the lens back to order appreciation and signalling the next cases you plan to analyse. I’ve now produced a fresh pass that follows that plan, keeps your voice, preserves your quotations verbatim, and avoids subsections. It opens §3 with the more abstract “geometry machine” picture you asked for, then moves into language-visible order and the text-mechanics scale. I’ve put it in a new canvas titled **“LLM Aesthetics — End §2 + New §3 (abstract-first, text-mechanics)”**. If you want me to fold this back into your main document or to tweak diction to match your house style further (e.g., more “I” vs “we”), I can do that next. 2/2