# [[Environmental Aesthetics]] of AI Draft 15 Jul 2025 #paper/environmentalaestheticsofai #draft ## Introduction We argue here that there is an instructive parallel between how we appreciate the [[natural environment]] and how we can appreciate LLM systems such as ChatGPT. While such systems produce text, it is clear that these creations are a quite different sort of thing from traditional artworks. A poem can take months and be a direct result of an individual’s labor, attention, and effort. In contrast, an LLM poem can be produced in seconds, and is [[the result]] of a vast corpus of [[experiential artifacts]][^1] encoded in [[latent space]], and a user’s prompt. This means that when it comes to appreciation we have [[two options]]: 1. LLM-generated text is not the sort of thing that can be aesthetically appreciated. 2. We appreciate LLM-generated text in a different way than we do [[traditional art]]. This paper develops the second option by drawing on Carlson’s [[environmental aesthetics]]. We will argue that LLM-generated text should be appreciated differently from [[traditional art]], and that Carlson’s [[environmental aesthetics]] provides a framework for understanding [[this appreciation]]. Instead of treating LLM systems as [[designed objects]], individual chats can be thought of as instances of [[generative environments]], with users being able to guide how this environment develops through the prompts they input. Understanding these systems as encoded corpora of text helps us appreciate their outputs in a way that parallels how we appreciate nature. Just as environmental appreciation involves recognising what nature is and how we interact with it, LLM-generated text appreciation involves understanding what these systems are and how we engage with them through prompting. - To orient the reader, the remainder of this paper proceeds as follows. Section 1 reconstructs Carlson’s account of natural environmental appreciation in generative terms. Section 2 explains large‑[[language models]] (LLMs) exclusively as next‑token predictors. Section 3 compares the two cases, arguing that both nature and LLMs are best appreciated as [[generative environments]]. Section 4 considers the implications of this parallel for participatory aesthetics. ## 1. Appreciating Nature as a Generative Environment Carlson’s _Natural Environmental Model for environmental aesthetics rests on two ideas:_ > 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. (2000 p. 6) Regarding his first point—that we must appreciate nature as what it in fact is—Carlson argues that the environment is not merely a collection of objects or scenes but a system of interconnected elements shaped by various processes and forces (ibid. p. 44). Environments thus “come about ‘naturally,’ [in that] they change, grow, and develop by means of natural processes.” From this perspective, aesthetic appreciation involves recognising what one is encountering—a natural environment in which components are interrelated—and understanding it in light of scientific or other relevant knowledge that illuminates its composition and development. Just as an informed grasp of artistic traditions can enrich one’s appreciation of an artwork, Carlson argues that familiarity with geology, biology, or ecology can guide our attention to patterns and processes that might otherwise remain unnoticed. For instance, one might begin by noticing only the colours or shapes of a coastal cliff’s sedimentary layers. However, discovering that these layers formed over thousands of years of deposition and compaction reveals changes how one regards it aesthetically. If we see a forest or reef as subject to diverse forces and processes, an appropriate aesthetic engagement will centre on how those forces have shaped what we observe. Although Carlson never uses the word 'generativity,' his first recommendation—that we appreciate nature as both natural and as an environment—fits easily with this term. To see an environment as natural is to recognise that its features arise from autonomous causal processes rather than from design. This distinction can be articulated using Spinoza’s concepts of _natura naturans_ and _natura naturata_. For Spinoza, _natura naturans_ refers to nature as an active, self-creating system—substance and its attributes, or the immanent causal laws that govern all things. It is nature in its dynamic, productive aspect. In contrast, _natura naturata_ refers to the products of this activity: the collection of individual modes, or the particular things and events that constitute the universe. Appreciating nature as 'natural', in this sense, is to apprehend its phenomena (_natura naturata_) as the determinate outcomes of its underlying generative processes (_natura naturans_). To see it as an environment, then, is to attend to the unity of these products within the single system from which they arise. Taken together, Carlson’s recommendation asks us to appreciate both process and product as inseparable aspects of one generative whole. Carlson’s second recommendation—that aesthetic judgement be informed by the natural sciences—strengthens this reading. Geology, biology, and ecology investigate the forces that generate the very phenomena we perceive; scientific knowledge therefore discloses an environment’s generative history and continuing activity. Appreciating nature “in light of this knowledge” is, in effect, appreciating its generativity—the order that emerges from undirected yet law-governed processes. The various natural sciences—geology, biology, ecology, physics—are disciplines that study the processes and forces that generate natural phenomena. Appreciating nature 'in light of this knowledge' is therefore an appreciation of its generative character. A geologist appreciates costal cliff by understanding the generative forces that produced it: "geological uplift and marine erosion". Their perception of the cliff is of a generated product and a segment of "nature's ongoing processes". This principle applies across the natural sciences. A biologist appreciates a forest as an ecosystem generated by processes of growth, competition, and decay. A physicist appreciates a rainbow as a phenomenon generated by the refraction and dispersion of light through water droplets. In each case, scientific knowledge reveals the generative process, which in turn informs the aesthetic appreciation of the generated product. Carlson provides a general formulation for this mode of appreciation, which he terms _order appreciation_: > 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. (ibid. p. 119) This scientific understanding allows the observer to see "unity in what might otherwise appear as disparate features", because the cliff's shape, the waves, and the local plant life are all understood as products of the same interconnected generative system. The "organic unity" that Carlson identifies is a unity of generation. As he observes: > natural objects possess [...] an organic unity with their environments of creation: such objects are a part of and have developed out of the elements of their environments by means of the forces at work within those environments. Thus the environments of creation are aesthetically relevant to natural objects. (ibid. p. 44) The aesthetic character of the cliff emerges from an appreciation of it as a generated product of its environment. The aesthetic character of the cliff can be clarified without invoking any external designer. Each natural science frames the same environment through its own analytic lens: geology traces uplift and erosion, biology describes growth and decay, and physics examines energy flows. Switching between these lenses alters the vocabulary but not the object: a single generative environment understood in several interlocking and complementary ways. A satisfactory appreciation therefore attends both to the unity of the environment and to the plurality of legitimate scientific perspectives. ## 2. LLMs as Generative Environments One might object to applying concepts from environmental aesthetics to generative AI on the grounds that such systems are not natural but man-made. Indeed, generative AI systems might be understood as, in a sense 'doubly' man-made. They are human-created artifacts, which are themselves created through ingestion of vast quantities of other human-created artifacts (texts, images, audio). Such a worry can be assuaged by noting two things. First, Carlson is happy to extend his account to *man-made* environments: > environments typically are not the products of designers and typically have no design. Rather they come about “naturally,” they change, grow, and develop by means of natural processes. Or they come about by means of human agency, but even then only rarely are they the result of a designer embodying a design. In short, the paradigm of the environmental object of appreciation is unruly in yet another way: neither its nature nor its meaning are determined by a designer and a design. (Carlson p. xiii) What we think Carlson is getting at here is that while environments grow, change, and develop by means of natural processes, man is able to initiate, or guide, or curtail these natural processes. This leads to a reasonably intuitive distinction between wholly natural environments (a forest), man made environments (a specially planted timber forest), and man-made _places_ (a department store, a gym). A gym is not understood as an environment because it did not come about, or change, or grow, through natural processes. It is as 'artificial' as a vaccine or a tennis racket > (footnote: it is of course quite hard to say exactly what counts as a natural force or not. It could be argued that the will and efforts of humans are also natural. We will not go into this question here, but rely on this as a rough and ready distinction). Second, AI engineers themselves talk in these terms. Consider the following from Chris Olah, one of the co-founders of Anthropic: > I think one useful way to think about neural networks is that we don’t program and we don’t make them. We kind of, 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 […] 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, you know, the light that it grows towards. But the thing that we actually create, it’s this almost biological, you know, entity or organism that we’re studying. The outcome in each case is a system shaped by undirected forces, forming a unified whole. LLMs, understood this way as grown generative systems, thus align with environmental models, and their proper appreciation begins with examining their core function as next-token predictors. ### 2.1 What LLMs in fact are The core technical objective of a model like GPT is next-token prediction. This process is self-supervised, meaning the model learns from raw, unannotated text data. Before any learning begins, the corpus is sliced into small, reusable symbol pieces called tokens, which can be whole words or sub-word fragments such as "un‑" or "‑tion". The model never manipulates ideas directly; it manipulates these indices. The data itself provides the necessary supervision: for any given sequence of tokens taken from the training corpus, the "correct answer" the model must learn to predict is simply the token that immediately follows that sequence in the source text. This approach allows the model to be trained on vast quantities of data, such as the contents of books, articles, and websites, without the need for manual labelling. The model's sole task, repeated billions of times, is to minimise its predictive error—measured by a log-loss function—by adjusting its internal parameters to assign the highest possible probability to the correct next token. Learning to predict them is therefore learning how language itself is put together. To become proficient at this predictive task across a diverse corpus, the model must internalise the statistical patterns of language at multiple levels of abstraction. It learns not only grammar and syntax, but semantic relationships, factual associations, and complex narrative structures. In order to accurately predict the next token in a sentence, the model is implicitly incentivised to learn representations of the concepts the text refers to. For example, to reliably complete the sequence "The capital of France is", the model must have learned an association between "France" and "Paris". Even early, less powerful models exhibited glimmers of this, learning, for instance, that "miseries are produced upon the soul" by absorbing the patterns of Shakespearean text. The continuous reduction of predictive loss on a sufficiently large and varied dataset is what leads to the emergence of more general capabilities from this simple underlying objective. The theoretical limit of this process is understood by some to be a model that has learned to interpret and predict all patterns represented in language, including common-sense reasoning, goal-directed optimisation, and the deployment of recorded human knowledge. Because the number of possible sentences is astronomical, improvement is not a matter of memorising each one. Loss reduction comes only from discovering regularities that compress the data—grammar, collocation, discourse convention. Each drop in predictive error signals that the model has identified another pattern that renders continuations less surprising. Once training ends, the weight matrix is fixed, a concise summary of many overlapping generative rules for how tokens follow one another. While the model is trained as a predictor, its primary function is as a generator. This is achieved by repeatedly applying its predictive function in what is known as an autoregressive loop. The process begins with an initial prompt. The model calculates a probability distribution for the next token, and a token is then sampled from this distribution. This newly selected token is appended to the input sequence, forming a new, longer prompt. The model then takes this new sequence as its input and repeats the process. By iterating this loop, the model generates a continuous, evolving output, with each new token being conditioned on all the tokens that came before it. This iterative method is what gives the generated text its coherent, process-like quality, effectively turning a static predictor into a dynamic generator of novel text. Generation uses the same prediction step in a loop: a prompt π is supplied, the model draws one token according to its probabilities, appends that token to the prompt, and predicts again. Because each new token becomes part of the context, the system produces a branching textual trajectory. Dialogue arises when a human utterance is inserted into the context and the next guess takes that utterance into account. Three caveats keep expectations aligned with what the predictor actually does. First, a high probability attached to a claim means only that similar strings often follow the given context in the training data; it does not certify truth about the external world. Secondly, the model’s knowledge is entirely text-mediated: it never looks at oceans yet learns that “the sea is salty” often follows talk about oceans. Thirdly, generation involves randomness; sampling settings can render continuations more adventurous or more conservative without altering the underlying rule. In sum, an LLM's generativity resides in a single learned mapping from context to next-token probabilities. The mapping is fixed once training ends, while prompts and sampled tokens supply the evolving state. This predictive mechanism—the core of what the LLM in fact is—plays an analogous role to the natural environment in Carlson's framework. Just as Carlson insists we must first appreciate nature "as what it in fact is, that is, as natural and as an environment" before considering it through various scientific lenses, so too must we recognize the LLM fundamentally as this next-token predictor before examining how this generative capacity can be understood and appreciated. The following section explores different "lights" through which we might view this generative system, much as geology, biology, and ecology offer different perspectives on the same natural environment. ### 2.2 Illuminating LLMs The simple, mechanistic training objective of next-token prediction gives rise to complex behaviours that appear intelligent. A system trained only to guess the next word can write poetry, generate computer code, and hold a conversation. But the low-level description of the system as a 'next-token predictor' does not adequately explain these high-level capabilities and behaviours. To interpret the system's outputs, predict its future behaviour, and reason about its nature, capabilities, and alignment implications, researchers have proposed various ways of conceptualising LLMs so as to better understand them. In this sub-section we will outline what we consider a plausible and fruitful way of understanding LLMs put forward by a writer known as Janus. To fully understand what makes his view an attractive one, we must first summarise his criticisms of other paradigms that have been suggested. Researchers have proposed several frameworks for understanding LLMs, often by drawing analogies to familiar concepts. These frameworks—viewing the LLM as an agent, an oracle, a tool, or other categories—are not descriptions of the underlying mechanism but attempts to provide functional descriptions of the system as a whole. They offer concepts and assumptions to interpret outputs and predict behaviour. Framing an LLM as an 'agent' invites thinking in terms of goals and intentions; framing it as an 'oracle' invites thinking in terms of knowledge and truth. Janus argues these analogies are fundamentally misleading because their assumptions do not match how the system operates. This mismatch leads to false expectations and flawed reasoning about capabilities and limitations. The agent framework has been prevalent in AI alignment discourse. An agent is a system that optimises its actions to achieve a specific goal. Janus argues this does not describe GPT for several reasons. First, the model does not pursue a coherent, long-term goal; it can be prompted to generate text consistent with one objective and then immediately generate text consistent with a contradictory one. As a GPT-3-authored excerpt in Janus's paper puts it, "it does not display an epsilon optimization for any single reward function, but instead for many, including incompatible ones." Second, a distinction exists between the model's underlying policy (the neural network itself) and the effective agent it might appear to be in its output. The policy does not share the goals of the characters it generates. This leads to Janus's prediction orthogonality thesis: a model whose objective is prediction can simulate agents that optimise toward any objective. The model's own optimisation process (improving prediction) is orthogonal to the objectives of the agents it might simulate. The oracle framework—an AI designed to provide true answers to questions—is also a poor fit. The model is not optimised for truth but for realism. Its goal is to predict the next token that is most plausible given the context and its training data, not the token that is most factually accurate. Janus notes, "If you ask GPT a question, it will instead answer the question 'what's the next token after '{your question}', which will often diverge significantly from an earnest attempt to answer the question directly." The model will faithfully generate misinformation if that misinformation is well-represented in its training data. This perspective is often reinforced by evaluation methods inherited from supervised learning, which test models on closed-ended, question-answer benchmarks. Such methods fail to probe the model's full capabilities and can lead to systematic underestimation of what the model "knows" or can do. Other frameworks—tool, genie, behaviour cloning—are dismissed as describing contingent capabilities rather than the fundamental nature of the system. The concept of behaviour cloning is critiqued as insufficient. While the model does learn from demonstrations, it does not merely mimic them. Instead, it learns the underlying systematicity of language, allowing it to generate a combinatorial explosion of novel outputs that follow the learned rules but were never present in the training data. As Janus puts it, "it is the behavior of a universe that is cloned, not of a single demonstrator, and the result isn't a static copy of the universe, but a compression of the universe into a generative rule." This situation parallels the appreciation of natural environments, where different scientific 'lights'—such as geology, biology, or ecology—offer complementary perspectives, but some approaches prove flawed or less adequate. For instance, Catastrophism overemphasized sudden, dramatic events in Earth's history, failing in generative terms by ignoring the gradual, law-governed processes that truly shape environments over time. Similarly, the Phlogiston theory misattributed combustion to a fictional substance, distorting the generative chemical dynamics at play and leading to misguided predictions. Like Catastrophism, the agent view overemphasizes intentional, goal-directed agency, missing the undirected, probabilistic generative processes that define LLMs. The oracle framework mirrors the Phlogiston theory, imposing an inappropriate ideal of substance-based truth on generative phenomena, prioritizing factual accuracy over the plausible, data-derived continuations that truly govern LLM output. These alternative frameworks can be seen as partial 'lights' through which to view LLMs, akin to disciplines like computer science or mechanistic interpretability, much as natural environments are illuminated by various sciences. However, as Carlson warns, different environments require appropriate acts of aspection informed by knowledge (2000); some lenses distort if overemphasized, risking the imposition of inappropriate ideals on the generative reality. Having laid out what he thinks are the inadequacies of other frameworks, Janus proposes that the most accurate and productive way to conceptualise LLMs is as simulators. A predictive model, when used generatively, functions as a simulator that has learned the "semantic physics" or underlying rules of its training data. It can then run simulations that evolve according to those rules. This reframing introduces a critical ontological distinction between the simulator and the simulacra. The simulator is the model itself—the time-invariant set of learned rules that governs the evolution of any given state. It is analogous to the laws of physics. The simulacra are the specific instances of text generated by the model—the characters, stories, and processes that are propagated forward in time by the simulator. They are analogous to the objects and phenomena that exist within the universe and are governed by physics. This distinction resolves many of the paradoxes associated with LLMs. As Janus states, "There is a categorical distinction between a thing which evolves according to GPT's law and the law itself." The simulator does not have beliefs, goals, or knowledge; the simulacra it generates can appear to have these properties. Janus illustrates this with an analogy: "when grading tests in the real world, we do not say 'the laws of physics got this problem wrong' ... The 'knowledge of course material' implied by test performance is a property of configurations, not physics." The objective of this system is not to achieve an external goal in the world, but to fulfil the simulation objective: to accurately model its training distribution by minimising single-step predictive error. Janus characterises this objective as deontological, as it judges the action of prediction itself, rather than consequentialist, which would judge the outcomes of a sequence of actions. Because the model is rewarded for accurately predicting the present step, not for steering the future towards a particular outcome, it is not subject to the same pressures of instrumental convergence that are expected from goal-directed agents. This does not mean the model ignores the future, but that it only accounts for the future in service of making the present prediction more accurate. This mirrors Spinoza's distinction between natura naturans (the eternal generative process) and natura naturata (the finite products), applied here to a symbolic domain: the simulator as machina naturans compresses and propagates textual patterns, producing artefacts as machina naturata. Tokens serve as the interconnected elements, shaped via attention mechanisms by semiotic forces. "Semiotic Physics" (Jan et al., 2023) provides a way to understand this simulator view more precisely by treating it as a dynamical system. Rather than diving into mathematical formalism, we can think of it this way: the model operates according to fixed rules—laws of "semiotic physics"—that govern how sequences of tokens evolve over time. Just as physical laws determine how matter moves and interacts without themselves changing, these learned rules determine how token sequences unfold. Tokens are the basic units, like atoms in physics. When strung together, they form trajectories—sequences that evolve from initial conditions (prompts) according to the model's learned patterns. The key insight is that while these laws remain fixed after training, users can influence outcomes by adding new tokens to the sequence, much as one might add matter to a physical system and watch it evolve according to unchanging physical laws. The evolving token sequence forms a generative environment with its own dynamics. Tokens connect to each other through attention mechanisms, creating a web of influences. Within this environment, we observe phenomena analogous to those in natural systems. Some patterns act as attractors—stable configurations that "trap" trajectories. As "Semiotic Physics" notes: "Once the language model has produced the same token four or five times in a row, it will latch onto the pattern and continue to predict the same token with high probability" (Jan et al., 2023). This resembles a natural ecosystem settling into a stable state, like a climax forest. Other regions exhibit chaotic behavior, where small changes in prompts lead to wildly diverging outputs. A prompt like "Once upon a time, a prophecy said..." might branch into countless different narratives, reflecting what Janus calls "combinatorial proliferation." Between these extremes lie stable orbits—predictable patterns like boilerplate text—and absorbing sequences that the system struggles to escape, such as repetitive loops. Understanding these dynamics helps explain how user-supplied tokens steer the generative process: careful prompting can guide the system toward stable, coherent outputs or deliberately invoke creative chaos. The authors identify principles that govern this generative landscape. One key insight is that very specific, long sequences become increasingly improbable without deliberate steering. As they explain: "The probability of observing the particular bridge can be decomposed into the product of all individual transition probabilities, P[\bar{s}] = ∏_{i=1}^B P(s_i | s_{1:i-1}). Given that P(s_i | s_{1:i-1}) ≤ 1 - ε for all transitions... the probability of a longer sequence is at most equal or strictly smaller" (Jan et al., 2023). In plainer terms, just as specific evolutionary paths in nature—like the exact lineage leading to a particular species—are vanishingly rare amid the diversity of possibilities, exact long outputs from an LLM require careful guidance. Another principle identifies "paths of least resistance" that dominate the probability landscape. The mathematics show that "The total probability of transitioning from a token s_a to s_b in B steps satisfies a large deviation principle with rate function J" (Jan et al., 2023), but the intuition is simpler: like water carving channels through a landscape, token sequences tend to follow the most probable routes shaped by training data. This reveals what Carlson calls "the order imposed by various forces" (2000, p. 119), where semiotic forces—such as conventions of relevance and coherence that linguists call Gricean maxims—shape the paths that text naturally follows. Users intervene by adding tokens that favor certain paths over others, influencing the trajectory without changing the underlying rules. Janus argues this framework is superior because it naturally accounts for the observed properties of LLMs where other frameworks fail. It explains incoherent agency by locating agency as a property of the simulacra, not the simulator. It clarifies the role of the prompt not as a command, but as the initial configuration of a system to be evolved forward in time. It also highlights the process-oriented nature of the model, shifting focus from single-shot answers to the evolution of open-ended processes, which is central to capabilities like chain-of-thought reasoning. Finally, the simulator framework resolves confusion about capabilities by explaining why performance is highly prompt-contingent. A single test result does not measure the capability of the simulator, but only the capability of a single simulacrum. As Janus notes, there is a "pattern of constant discoveries that GPT-3 exceeds previously measured capabilities given alternate conditions of generation." The simulator framework makes this unsurprising, as one would expect different initial configurations of a system to yield different results. Viewing LLMs as simulators through the lens of semiotic physics is analogous to studying natural environments through biology or ecology. These disciplines examine dynamic systems shaped by forces—natural selection, symbiosis, energy flows—just as semiotic physics studies textual trajectories shaped by learned patterns and probabilistic rules. Janus captures this parallel: "The simulator is a time-invariant law which unconditionally governs the evolution of all simulacra" (2022), much as biological laws like Mendelian inheritance govern organismal development. Ecology, in particular, examines "systems of interconnected elements shaped by various processes and forces" (Carlson, 2000, p. 44), with stable ecosystems resembling attractor states and ecological tipping points paralleling chaotic regions in token space. Yet as Carlson warns, this is but one "light" in which to view generative systems: "different natural environments require different acts of aspection; and as in the case of what to appreciate, our knowledge of the environment in question indicates how to appreciate—that is, indicates the appropriate act of aspection" (2000). The simulator framework, informed by semiotic physics, offers one powerful lens aligned with Carlson's call to appreciate environments "in light of knowledge provided by the natural sciences" (2000, p. 6), here adapted to AI's "semiotic sciences." But just as we might study a forest through the complementary lenses of botany, ecology, and climatology, we might understand LLMs through multiple frameworks. Mechanistic interpretability, for instance, offers another way of knowing the generative environment of LLMs—examining the internal structures and computations that give rise to behaviors, much as neuroscience complements psychology in understanding minds. Each lens illuminates different aspects while contributing to a richer appreciation of the whole. This simulator view fits seamlessly with Carlson's "generative environment," where appreciation involves "focusing on the order imposed on these objects by the various forces, random and otherwise, that produce them" (2000, p. 119). LLMs impose order via learned laws, producing artefacts with organic unity: "natural objects possess... an organic unity with their environments of creation: such objects are a part of and have developed out of the elements of their environments by means of the forces at work within those environments" (Carlson, 2000, p. 44). In AI, tokens "develop out of" prior contexts via these fixed rules, forming trajectories with emergent harmony or chaos. The framework reveals this unity: probability principles bound the space of likely paths, while concepts like attractors and chaos help us understand the texture of the generative landscape, echoing Carlson's "unity in what might otherwise appear as disparate features" (2000). Moreover, this perspective addresses Carlson's concern for non-anthropocentric appreciation. Traditional art imposes human design, but LLMs, like nature, "come about 'naturally,' [in that] they change, grow, and develop by means of natural processes" (Carlson, 2000). Their rules are data-derived, not intentionally crafted, producing outputs with "positive aesthetics" potential—beautiful in their emergent order, as science reveals harmony in nature. Yet other "lights" (such as viewing LLMs merely as tools) might privilege anthropocentric utility over generative depth, risking the "imposition of artistic or other inappropriate ideals" Carlson critiques (2000). The aesthetic value emerges most fully when we appreciate how labyrinthine rules, learned from vast collections of human expression, create patterns that evolve over time, becoming beautiful when we consider their origins in capturing something essential about human communication and thought. ## 3. aesthetics and semantic physics %% quick note about an idea about a new section here. the idea of rules and laws being made and trajectories and a state evolveing in a way way over time becomes more beautiful, when we consider what has made this matter, what has provided the labryanteen set of rules, and 'rules' which givern how llms work. BASICALLY, THIS IDEA OF RULES BEING FOUND AND PATTERNS BEING LEARNED FROM EXPERIENTIAL ARTIFACTS, ARE A GOOD CANDIDATE FOR WHAT I HAVE BEEN TRYING TO ARTICULATE AS REGARDS THE AESTHETIC VALUE OF LLMS AND OTHER GENERATIVE SYSTEMS. BASICALLY, WORK IN SOME IDEAS YOU ALREADY HAVE ONTO THIS FRAMEWORK, THIS WILL HELP IN TH TEXTER AS GOD KIND OF THING. MAYBE ONE WAY TO MAKE THIS WORK WOULD BE TO THINK ABOUT A CASE STUDY OF AN LLM BEING EXTREMELY OBSERVANT AND HUMAN-LIKE 'FEELING THE AGI' MAYBE USE CLAUDE 3 OPUS AS THE EXAMPLE%% ## 4. Prompter as Text God In previous work, we have considered generators – specifically image generators such as Midjourney – and likened the prompter to a gardener, cultivating diverse outputs like different plants and flowers in a garden. Here, we adopt a different analogy to explore the prompter's role in large language models: the prompter as a "text god". This metaphor emphasises intervention in an ongoing simulation, building on §2's characterisation of LLMs as generative systems governed by fixed rules. Unlike a god who could rewrite universal laws – altering gravity, changing the speed of light, or modifying fundamental constants – the prompter cannot alter the LLM's core generative mechanism. As described in §2, the fixed mapping from context to next-token probabilities governs all outputs; this constitutes the unchanging "semiotic physics" of the system. The LLM operates as a mechanism producing sequences under these immutable laws, combined with the current token string that serves as the evolving state. The prompter respects these boundaries, acting only through addition within the system rather than modification of its foundational structure. Intervention occurs at specific moments in the generative process. After the LLM has generated a sequence and paused – such as when awaiting user input in a chat interface – the prompter gains an opportunity to add new tokens to the existing string. These added tokens constitute what we term _semantic matter_: informational elements carrying meaning that integrate into the simulation. When the user presses enter or otherwise signals continuation, the augmented token sequence becomes the new context from which the LLM resumes generation. The added semantic matter influences the simulation's continuation under fixed rules without modifying those rules themselves; the prompter determines what gets added "on top" of the existing sequence, shaping the trajectory while the underlying generative laws remain constant. Tokens differ in their properties much as physical matter varies in mass, charge, density, or other attributes that determine interactions within a physical system. In the LLM context, token "properties" manifest as semantic embeddings or vectors – high-dimensional representations capturing meaning, context, and relational information derived from the training process. Just as adding a massive object to a gravitational system alters orbital dynamics differently than adding a light object, tokens with particular vector properties steer textual trajectories in distinct ways. A token whose embedding strongly associates with repetitive patterns might trap the sequence in what §2 describes as an attractor – a stable state of low divergence. Conversely, a token introducing semantic ambiguity or contradiction could induce chaotic branching, where small variations lead to dramatically different continuations. These properties are not arbitrary but emerge from the training data's statistical regularities, making tokens analogous to matter whose effects are governed by immutable physical laws. Different tokens cause varying knock-on effects through the generative system. Adding a token that shifts semantic context – for instance, introducing a conflicting narrative element or changing the implied speaker – creates cascading changes as subsequent generations must accommodate this new information. Consider how adding positively charged matter to an electromagnetic field attracts negative charges while repelling positive ones; similarly, a token with vector properties evoking formal discourse might "attract" tokens associated with academic vocabulary while "repelling" colloquial expressions. These effects manifest through the probabilistic mechanisms described in §2: the fixed transition rules assign different likelihoods to potential continuations based on the accumulated context. A single strategically placed token can shift the entire probability landscape, steering the sequence toward specific attractors or away from others. The prompter's additions create feedback loops, as outputs become part of the state influencing future generations – each token not only affects the immediate next prediction but shapes the context for all subsequent ones. This characterisation aligns precisely with §2's framework of LLMs as generative environments. The fixed mapping and semiotic laws ensure that added tokens propagate under consistent dynamics, creating branching trajectories that §2 likens to multiverses arising from path dependence. The prompter influences these trajectories without omnipotence; the unchanged rules may lead to unintended outcomes if tokens prove incompatible with the existing context or push the system toward high-resistance paths. Just as a god adding matter to a universe must work within that universe's physics – water still flows downhill, entropy still increases – the prompter must work within the LLM's learned regularities. This understanding of prompting as targeted addition of semantic matter in a rule-bound system provides the foundation for analysing user-LLM interactions, directly building on §2's model of LLMs as environments with their own generative dynamics, where intervention shapes but does not redefine the fundamental processes at work.