# 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 --- # plan for section 5 #paper/environmentalaestheticsofai **Paragraph 1** (existing opening, verbatim) Carlson's requirement: appropriate appreciation depends on a general [[nonaesthetic story]] that makes [[order visible]]. The appreciator focuses on "[[the order]] imposed on these objects by the various forces, random and otherwise, that produce them" guided by "a general nonaesthetic and nonartistic story that helps make this [[order visible]] and intelligible." For natural environments, this story is supplied by environmental sciences—geology, ecology, meteorology. **Paragraph 2** (existing second paragraph, verbatim) [[The level]]-of-description point. Chemical physics could in principle explain cliff faces, but geology makes Carlson's idea of aspection easier to explain. Geological categories (strata, faults, 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 to attend to. The microstructural patterns described by chemical physics do not map as straightforwardly onto features available to ordinary perception. For present purposes, we start from sciences whose categories already function as ways of looking, even if more fundamental levels could in principle support a similar story. **Paragraph 3** Transition to LLMs. What is the analogous [[nonaesthetic story]] for LLM-mediated chats? What kind of knowledge could make [[the order]] in generated text visible and intelligible? Section 2 characterised LLMs in computational terms: tokenisation, embeddings, attention mechanisms, pretraining on prediction error, post-training through RLHF. These are genuine sciences of LLMs. Mechanistic interpretability (MI) research has made substantial progress in identifying what specific circuits, attention heads, and [[internal representations]] do. This knowledge is relevant to Carlson's first recommendation—it tells us [[what LLMs are]]. **Paragraph 4** However, for purposes of the second recommendation—the [[right kind]] of knowledge for appreciation—MI stands to LLM outputs roughly as chemical physics stands to cliff faces. Nothing in principle rules out developing an aesthetics grounded in attention-head activations or circuit-level features. But the categories of MI (specific attention patterns, feature directions, circuit structures) do not map straightforwardly onto what ordinary users perceive when reading generated text. Since our concern is with how users can appreciate LLM-mediated chats, we focus on a different level: [[the level]] of the text itself, which is what users actually encounter and respond to. %% paras 3 and 4 can be combined so as to make things more succinct. Also, for this part of the final text feel free to use some of the phrasing sentences from this paragraph from an old draft, I think they articulate things quite well, but perhaps need to be suppliemented: Similarly, while mechanistic interpretability reveals important truths about neural networks, the gap between descriptions of weight matrices and [[the experience]] of reading text is wide. The impediment is one of perceptibility: how to connect sub- symbolic, mathematical descriptions to the surface features of generated text that readers encounter.%% **Paragraph 5** Introduce Janus and Picca. Recent work has proposed frameworks for understanding LLMs that operate at this textual level. Janus (2022) suggests thinking of GPT-style models as systems that have learned to propagate text according to regularities induced from [[training data]]—generating trajectories through text-space from prompt-specified [[initial conditions]], much as physical systems evolve states according to physical laws. The model is not an agent pursuing goals but something more like a dynamics governing how text unfolds. Picca (2025) develops a parallel idea through semiotics: LLMs are "semiotic machines" that recombine and circulate signs according to learned statistical patterns, without understanding or intention. Both shift focus from cognition and agency to the question of what regularities govern textual transitions. **Paragraph 6** Announce the development. Drawing on these ideas, we develop what we call "semiotic physics": an account of the regularities governing sign-transitions in LLM outputs. This account is pitched at the level of text—what patterns govern how tokens follow tokens—rather than at the level of computational mechanism. Our claim is that semiotic physics can serve as the nonaesthetic story that makes the emergent order in LLM-mediated chats visible and intelligible, fulfilling the role that geology and ecology play for natural environments in Carlson's framework. %%be careful on these paragraphs (5 and 6) frame the issues. paragraph 5 should begin with something like the following: We suggest that a viable candidate for the right kind of knowledge can be found in work such as REF and REF. What follows is our articulation of what we take to be a somewhat similar framework for understanding LLMs. Then you need to describe the ideas that can be found in Janus that our version of semiotic physics will use/borrow/endorse. Use quotes from Janus to illustrate exactly what he is talking about, becareful not to get too lost in the details. the stuff you take from janus should be chosen with an idea of what comes next in the rest of this section. Once you have done this with janus, do exactly the same procedure for Picca.%% %%I worry that the next few paragraphs are simply a list of features rathewr than any sort of an argument. Can you give me a new iteration which takes my comments into account, and has a more elegant, more analytic framework for the structure of the rest of this section.%% **Paragraph 7** What are semiotic regularities? They are constraints on how signs (tokens) follow signs in sequence. When an LLM predicts the next token, it draws on patterns learned from its training distribution—which is predominantly human text production. Human text production is governed by regularities at multiple levels: syntactic constraints on word order and agreement; semantic associations that make certain words probable in certain contexts; discourse-level patterns that structure how paragraphs and arguments unfold; genre conventions that govern what counts as appropriate in different types of text; pragmatic norms concerning implicature, politeness, and speech acts; stylistic patterns characteristic of registers, periods, and individual authors. To predict text well, a model must learn to track these regularities. The totality of what it learns constitutes its semiotic physics. **Paragraph 8** These regularities are hierarchically organised. At the lowest level are sub-lexical patterns: which letter combinations are common in English, which are rare or impermissible. Above this are lexical co-occurrences: "doctor" raises the probability of "patient" and "prescribe"; "inclement" collocates with "weather" but not "food." Then syntactic patterns: determiners are typically followed by nouns or adjectives, not verbs; after "the tall," a noun is highly probable. Then sentence-level coherence: subjects and verbs agree in number, pronouns require antecedents, presuppositions must be satisfied or accommodated. Then discourse-level structure: paragraphs maintain topics, arguments move from premises to conclusions, narratives establish setups before payoffs. Then genre conventions: academic prose hedges and cites, dialogue turns and interrupts, technical documentation enumerates and instructs. Then stylistic patterns: short declarative sentences in one register, complex periodic sentences in another, passive voice in scientific writing. These levels interact and mutually constrain. **Paragraph 9** The regularities are context-sensitive. The probability of "bank" being followed by "account" depends on whether the preceding text concerns finance or rivers. The probability of formal register depends on genre cues established earlier. What counts as a coherent continuation depends on what has been established. Some regularities depend only on local context—the immediately preceding tokens, the current syntactic state. Others depend on the entire preceding text: the topic, the apparent speaker, what claims have been made, what questions remain open. A model with a long context window can, in principle, condition on all of this. Context-sensitivity is what gives semiotic physics its depth: the relevant regularities are not context-free rules but patterns that interact with arbitrarily complex prior text. **Paragraph 10** The regularities are probabilistic. Given a context, many continuations are consistent with the semiotic physics; what the physics determines is which are more or less probable, not which is uniquely correct. This is not a deficiency. Human text production is genuinely variable—the same person in the same situation might produce different sentences. Semiotic physics determines distributions over trajectories, not unique trajectories. This probabilistic character is reflected in how models are trained (to minimise cross-entropy, a scoring rule that rewards accurate probability estimates) and in how they generate (sampling from output distributions, not simply selecting the most probable token). **Paragraph 11** The regularities are not explicitly represented in the model. There is no module labelled "syntax" or "genre conventions" or "pragmatic norms." The regularities are implicit in the weights—they are what the weights collectively implement when the forward pass is computed. The model behaves as if it respects syntactic constraints, tracks semantic associations, maintains discourse coherence, and observes genre conventions, because doing so minimises prediction error on human text. This is analogous to how a physical simulation embodies physical laws without explicitly representing them: the rules of Conway's Game of Life are implicit in the update function, not stored as a list. Similarly, semiotic physics is implicit in the transformer's learned parameters. **Paragraph 12** Connect to Section 4's emergent-order point. Section 4 argued that the aesthetically interesting order in LLMs is emergent rather than designed—the lesson of the raku and Pollock analogies. Designers specify scaffolds and objectives; what grows on those scaffolds through optimisation is not the execution of a detailed plan. But "emergent" names the phenomenon without explaining it. Semiotic physics fills this explanatory gap. The order in LLM outputs is the order of regularities learned from human text production. These regularities were not programmed into the model; they were induced through gradient descent on prediction error across billions of examples. But they are not random or inexplicable. They reflect how humans actually write: the patterns of syntax, the associations of meaning, the structures of argument and narrative, the conventions of genre and register. When a model produces text that exhibits coherent discourse structure or appropriate genre conventions, it is not executing a design specification; it is propagating regularities that were distilled from the training corpus. The "forces" producing the order—to use Carlson's term—are semiotic forces: patterns of sign-transition learned from the distribution of human language. **Paragraph 13** The prompt-as-initial-conditions framing. Section 2 described prompts as inputs that the model processes but did not characterise what role they play in the generative process in terms that connect to order appreciation. Semiotic physics provides this. A prompt specifies a point in text-space—an initial configuration from which trajectories will be propagated. The model's learned regularities then govern how the trajectory unfolds, determining a distribution over possible continuations. This is directly analogous to how physical laws propagate physical states from initial conditions. The model is not being asked a question in the sense relevant to oracles; it is being given conditions that its semiotic physics will evolve. Different prompts activate different regularities: a prompt establishing an academic genre activates hedging, citation patterns, formal register; a prompt establishing casual dialogue activates informality, turn-taking, contractions; a prompt constructing a fictional persona activates whatever regularities are associated with how such personas typically speak. The same model (same semiotic physics) produces different outputs from different prompts (different initial conditions). **Paragraph 14** Variation between models. Different models instantiate somewhat different semiotic physics. They are trained on different corpora (predominantly scientific papers, predominantly social media, mixed web text), with different architectures (varying context lengths, attention patterns), at different scales (parameter counts affecting how fine-grained the learned regularities can be), through different post-training regimes (RLHF targeting different response profiles). They therefore overlap on robust, high-frequency regularities—basic English syntax, common semantic associations, widespread genre conventions—but diverge on fine-grained patterns: characteristic approaches to hedging, handling of particular topics or genres, collocational preferences, typical response structures. What users perceive as a model's characteristic manner or style is a pattern in how its particular semiotic physics manifests across diverse prompts. This is not personality in the sense relevant to Section 3's person-aesthetics; it is the characteristic order of a particular set of learned regularities. **Paragraph 15** Conclude by setting up Section 6. Semiotic physics provides the nonaesthetic story that can make order in LLM-mediated chats visible and intelligible. Its categories—syntactic patterns, semantic associations, discourse coherence, genre conventions, pragmatic regularities—describe features that can be perceived in generated text and thus can guide aspection. An appreciator equipped with this framework attends to how regularities are being propagated: whether coherence is maintained, how genre conventions are being handled, where semantic associations are tight or loose, how the trajectory unfolds from the prompt-specified initial conditions. In Section 6, we turn to concrete cases, showing how this framework applies to particular episodes of LLM-mediated text generation.