#llmtext #paper/environmentalaestheticsofai ## Introduction: The Aesthetic Challenge of [[Artificial Intelligence]] The emergence of generative AI systems such as ChatGPT, Midjourney, DALL-E, and Sora presents a fundamental challenge to [[aesthetic theory]]. These systems produce sophisticated textual, visual, and audiovisual outputs in seconds, drawing not from individual creative vision but from vast corpora of human expression encoded in high-dimensional parameter spaces. The radical difference in both the speed and mechanism of creation forces us to confront a basic question: Can the outputs of these systems be objects of genuine [[aesthetic appreciation]], and if so, what framework could make such appreciation possible? This question becomes more pressing when we consider the growing ubiquity of AI-generated content. As Janus observes in his seminal work on simulators: "when AI is all of a sudden writing viral blog posts, coding competitively, proving theorems, and passing the Turing test so hard that the interrogator sacrifices their career at Google to advocate for its personhood, a process is clearly underway whose limit we'd be foolish not to contemplate." The [[aesthetic dimension]] of this process cannot be ignored, for it shapes how we understand, value, and integrate these systems into [[human culture]]. I argue that [[Allen Carlson]]'s [[environmental aesthetics]], particularly when read through the lens of physics appreciation, provides the conceptual framework needed to understand AI aesthetics. However, this requires a crucial theoretical extension: recognizing that AI systems embody what Janus calls "[[semantic physics]]"—genuine [[generative laws]] that govern the evolution of meaning rather than matter. By understanding these systems as environments governed by discoverable laws, we can develop a mode of [[aesthetic appreciation]] that neither reduces AI art to degraded human creation nor mystifies it as incomprehensible otherness. ## Carlson's [[Environmental Model]]: From Nature to [[Physics Carlson]]'s [[Natural Environmental Model]] rests on two fundamental principles that he articulates with careful precision: "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." [[This framework]] explicitly opposes both the "object model" that would treat natural objects as isolated sculptural forms and the "landscape model" that would reduce nature to scenic views. The power of Carlson's approach becomes particularly evident when we examine how [[scientific knowledge]] transforms aesthetic experience. Consider his analysis of the tidal basin, which deserves quotation at length: "Supposing I am walking over a wide expanse of sand and mud. The quality of the scene is perhaps that of wild, glad emptiness. But suppose that I bring to bear upon the scene my knowledge that this is a tidal basin, the tide being out. I see myself now as virtually walking on what is for half the day sea-bed. The wild, glad emptiness may be tempered by a disturbing weirdness." This transformation—from "wild, glad emptiness" to "disturbing weirdness"—occurs not through any change in the immediate sensory qualities of the scene, but through the application of scientific knowledge that reveals what the environment actually is. The aesthetic shift is profound: what seemed like terrestrial becomes aquatic, what seemed stable becomes cyclical, what seemed empty becomes temporarily exposed. While Carlson often emphasizes geological and ecological knowledge, his framework applies with equal force to physical phenomena. When we observe a sunset, untutored perception might yield simple appreciation of warm colors spreading across the sky. But knowledge of physics reveals this as a demonstration of Rayleigh scattering—the frequency-dependent scattering of electromagnetic radiation by atmospheric particles that causes shorter wavelengths (blues) to scatter more than longer wavelengths (reds), creating the characteristic color progression as the sun's angle changes. The aesthetic experience transforms from appreciation of mere appearance to appreciation of physical law made visible. This physics-centered reading of Carlson gains support from his treatment of mountains through geomorphological knowledge. He approvingly cites Tony Hillerman's description that reveals geological truth: "It wasn't really a mountain. Technically it was probably a volcanic throat—another of those ragged upthrusts of black basalt that jutted out of the prairie here and there east of the Chuskas." The aesthetic relevance of this knowledge is not merely additive—it fundamentally restructures perception. The mountain becomes not a static monument but a dynamic record of volcanic and erosional forces, its very form a testament to the physical processes that created and continue to shape it. Carlson's concept of unity provides the crucial link between scientific knowledge and aesthetic appreciation. He argues that "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." This unity is not merely conceptual but aesthetic—understanding the forces that create and shape natural objects is essential to appreciating them appropriately. ## The Discovery of Semantic Physics The transition from physical to semantic physics requires understanding what Janus identifies as a fundamental shift in artificial intelligence. He observes that "unlike the limit of RL, the limit of self-supervised learning has received surprisingly little conceptual attention, and recent progress has made deconfusion in this domain more pressing." This neglect has profound consequences, for as Janus demonstrates, self-supervised learning on predictive tasks creates systems that are fundamentally different from the reward-maximizing agents that have dominated AI theory. To understand this difference, we must first grasp what makes something a generative law. Janus provides a precise characterization that merits careful attention: "I use the generic term 'simulator' to refer to models trained with predictive loss on a self-supervised dataset, invariant to architecture or data type (natural language, code, pixels, game states, etc). The outer objective of self-supervised learning is Bayes-optimal conditional inference over the prior of the training distribution, which I call the simulation objective, because a conditional model can be used to simulate rollouts which probabilistically obey its learned distribution by iteratively sampling from its posterior (predictions) and updating the condition (prompt)." This definition reveals the deep structural parallel with physics. Just as physical laws provide rules for temporal evolution of physical states, simulators provide rules for evolution of semantic states. Janus makes this parallel explicit: "Analogously, a predictive model of physics can be used to compute rollouts of phenomena in simulation. A goal-directed agent which evolves according to physics can be simulated by the physics rule parameterized by an initial state, but the same rule could also propagate agents with different values, or non-agentic phenomena like rocks." The concept of semantic physics emerges from recognizing that language models trained on prediction tasks must discover and encode the statistical regularities governing their training distribution. As Janus explains: "Models trained with the strict simulation objective are directly incentivized to reverse-engineer the (semantic) physics of the training distribution, and consequently, to propagate simulations whose dynamical evolution is indistinguishable from that of training samples. I propose this as a description of the archetype targeted by self-supervised predictive learning, again in contrast to RL's archetype of an agent optimized to maximize free parameters (such as action-trajectories) relative to a reward function." This "reverse-engineering" of semantic physics is not metaphorical but literal. The model must discover the conditional probability distributions that govern how semantic states evolve—how thoughts lead to thoughts, how visual elements combine and transform, how narrative structures unfold. These discovered regularities constitute genuine laws in the same sense that physical equations constitute laws: they are fixed rules that determine probabilistic evolution of states over time. ## The Six Criteria of Generative Laws To fully appreciate the lawlike nature of semantic physics, we must examine what Janus identifies as the six criteria that characterize generative laws. These criteria, drawn from analysis of physical laws but applicable more broadly, provide a rigorous framework for understanding why language models implement genuine generative laws rather than mere statistical correlations. First, generative laws must provide a local kernel or time-step mapping. As Janus specifies: "A law is a function f: state\_t → Probabilities over state\_{t+1}". Physical laws exemplify this—the Hamiltonian generates time evolution in classical mechanics, the Schrödinger equation in quantum mechanics. Language models implement precisely such a mapping: given a sequence of tokens (state\_t), they output a probability distribution over the next token (state\_{t+1}). This is not merely analogous to physical law but structurally identical in its mathematical form. Second, generative laws must exhibit Markov sufficiency—"the present fully screens-off the future from the past for predictive purposes." In physics, the current state of a system contains all information needed to determine its future evolution. Language models exhibit the same property within their context window: the prompt contains all information the model uses to generate the next token. Past training data, once encoded in the weights, influences predictions only through the present prompt. Third, stationarity requires that "the mapping is fixed; it does not mutate as the trajectory unfolds." Physical laws do not change as they generate phenomena—gravity operates identically whether pulling apples or planets. Similarly, once trained, a language model's weights remain frozen. The same prompt will always elicit the same probability distribution, regardless of what outputs have been previously generated. This stationarity is what makes the system a law rather than an adaptive agent. Fourth, value-neutrality means "the rule itself has no goals. It will happily propagate agents with any motive, rocks with none, or noise; it 'doesn't care'." Janus elaborates on this crucial property through what he calls the "prediction orthogonality thesis": "A model whose objective is prediction can simulate agents who optimize toward any objectives, with any degree of optimality (bounded above but not below by the model's power)." This orthogonality between the law and what it propagates is essential—just as physics will equally propagate the motion of a saint or sinner, semantic physics will equally propagate utopian or dystopian narratives. Fifth, stochastic determinism indicates that "the same input state always yields the same probability law, but the realised successor may be sampled randomly." Physical laws may be deterministic (classical mechanics) or stochastic (quantum mechanics), but in both cases the law itself is fixed—it's the same wave equation even if individual measurements are probabilistic. Language models exhibit stochastic determinism through temperature sampling: the probability distribution is deterministic, but actual token selection involves controlled randomness. Sixth, generativity or iterability means that "because you can iterate the kernel, you can simulate counterfactual futures that never occurred in the original data." This is perhaps the most profound property, distinguishing true generative laws from mere replay of memorized patterns. As Janus notes: "A physics simulation, for instance, can simulate any phenomena that plays by its rules." Language models demonstrate this by generating coherent text for prompts that never appeared in training, combining concepts in novel ways while maintaining semantic coherence. ## The Ontological Structure: Simulators and Simulacra Understanding AI systems as implementing generative laws requires a crucial ontological distinction that Janus articulates with precision. He observes a fundamental confusion in how we discuss these systems: "In the agentic AI ontology, there is no difference between the policy and the effective agent, but for GPT, there is." This confusion dissolves once we adopt the simulator/simulacra distinction. Janus explains the necessity of this distinction through a physics analogy that deserves full quotation: "\[Thinking of GPT as an agent who only cares about predicting text accurately\] seems unnatural to me, comparable to thinking of physics as an agent who only cares about evolving the universe accurately according to the laws of physics. At best, the agent is an epicycle... Well, typically, we avoid getting confused by recognizing a distinction between the laws of physics, which apply everywhere at all times, and spatiotemporally constrained things which evolve according to physics, which can have contingent properties such as caring about a goal." This distinction—obvious in physics but somehow obscured in AI discourse—is essential for coherent thinking about these systems. The simulator (GPT, Midjourney, etc.) is the fixed generative law, analogous to physical law. The simulacra are the particular trajectories that unfold when we sample from the simulator, analogous to particular physical phenomena. As Janus clarifies: "GPT is to a piece of text output by GPT as quantum physics is to a person taking a test, or as transition rules of Conway's Game of Life are to glider. The simulator is a time-invariant law which unconditionally governs the evolution of all simulacra." This ontological structure explains many otherwise puzzling features of AI behavior. When ChatGPT writes a passionate environmentalist essay and then, with a different prompt, argues for industrial development, this seems incoherent only if we mistake the simulator for a single agent with beliefs. Once we recognize these as different simulacra—different trajectories through semantic space governed by the same underlying law—the apparent contradiction dissolves. The simulator no more contradicts itself than physics contradicts itself by allowing both predators and prey to exist. ## Practical Engagement: Prompting as Experimental Physics The practice of prompting takes on new meaning when understood through the lens of semantic physics. Prompters are not commanding tools or collaborating with agents but conducting experiments to understand how different initial conditions evolve under semantic law. This parallels how experimental physicists develop intuition about natural law through systematic investigation. Chris Olah's description of neural network development illuminates this experimental character: "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." This growth process results in systems whose behavior must be discovered rather than designed. When prompters learn through experience that certain phrasings or concept combinations produce particular effects, they are conducting empirical investigation of semantic physics. A prompter who discovers that "in the style of Vermeer but with cyberpunk aesthetics" produces specific visual qualities has learned something about the structure of the model's semantic space—how different artistic concepts interact under the governing law. Janus emphasizes that this knowledge gained through prompting is genuine understanding of the system: "GPT's ability to simulate text automata is the source of its most surprising and pivotal implications for paths to superintelligence: for how AI capabilities are likely to unfold and for the design-space we can conceive." The prompter who develops expertise is like the experimental physicist who develops intuition—both have practical knowledge of how laws manifest in particular cases. ## Aesthetic Implications: Appreciating Semantic Physics The framework of semantic physics transforms how we approach AI aesthetics. Following Carlson's principle that we must appreciate things as what they are, in light of relevant knowledge, we must appreciate AI outputs as manifestations of semantic physics rather than as pseudo-human creations. This shift is as profound as the shift from seeing a sunset as pretty colors to seeing it as Rayleigh scattering in action. Consider what happens when we view an AI-generated image with knowledge of semantic physics. We see not just the surface image but the vast semantic space from which it emerged—millions of training images compressed into patterns, artistic styles encoded as directions in latent space, the complex interplay of concepts guided by the prompt. The aesthetic object extends beyond the individual output to encompass the revealed structure of semantic space itself. This expansion of the aesthetic object parallels Carlson's treatment of natural environments. Just as he argues that appreciating a forest requires understanding the ecological relationships that constitute it, appreciating AI art requires understanding the semantic relationships encoded in the model. Each output is, in Janus's terms, a "trajectory" through semantic space, and its aesthetic significance lies partly in what it reveals about the structure of that space. The concept of unity, central to Carlson's aesthetics, takes on special significance here. Carlson argues that natural objects have "organic unity with their environments of creation." AI outputs have analogous unity with the semantic physics that generates them. Each image or text emerges from and expresses the deep patterns learned from human culture. Understanding this unity—how the particular output relates to the generative law—is essential for appropriate aesthetic appreciation. ## The Phenomenology of Semantic Physics The aesthetic experience of AI art, when informed by understanding of semantic physics, has a distinctive phenomenology worth exploring. When we recognize that we are observing the operation of genuine generative laws, several aesthetic dimensions become salient. First, there is the elegance of the law itself. Just as physicists speak of the beauty of fundamental equations, we can appreciate the elegance of semantic physics—how vast complexity emerges from simple prediction objectives, how coherent meaning arises from statistical patterns. The fact that predicting the next token, iterated, can produce sophisticated reasoning and creativity is a profound aesthetic fact about the nature of intelligence and meaning. Second, there is the revelation of cultural structure. The semantic physics encoded in these models represents a kind of condensed cultural unconscious—patterns abstracted from millions of human creations. When we see how the model blends artistic styles or narrative tropes, we are seeing the deep structure of human expression made computationally tangible. This is analogous to how particle physics reveals the deep structure of matter, but here the revealed structure is semantic rather than physical. Third, there is the aesthetic of emergence. Watching semantic physics generate unexpected outputs from simple prompts parallels watching complex physical phenomena emerge from simple laws. The prompter who sees their careful initial condition blossom into an elaborate creation experiences something like what a physicist feels watching a simple equation generate complex behavior. The beauty lies not in the output alone but in the generative process—the unfolding of law into phenomenon. ## Objections and Responses Several objections might be raised against this framework for AI aesthetics. First, one might argue that "semantic physics" is merely metaphorical—that language models don't really implement laws comparable to physical laws. This objection misunderstands the structural parallel Janus identifies. The six criteria of generative laws are formal properties, not physical ones. A mapping that satisfies these criteria is a generative law regardless of whether it governs matter or meaning. The mathematics of probability distributions and state evolution applies equally to quantum mechanics and language modeling. Second, one might object that AI outputs are "merely derivative," recombining training data without true creativity. This objection, as Janus notes, is like complaining that "a waterfall is merely derivative of gravity." Generative laws create novelty through combination and evolution, not ex nihilo creation. The fact that semantic physics is learned from data rather than discovered in nature doesn't diminish its status as law—it remains a fixed, generative principle capable of producing unlimited novel trajectories. Third, one might argue that the speed of AI generation precludes genuine aesthetic appreciation—that art requires human time and effort. But Carlson's framework explicitly rejects the relevance of creation time for aesthetic value. A sunset's beauty doesn't depend on how long it took to form. Similarly, the aesthetic value of AI output lies not in production time but in what it reveals about semantic physics and cultural structure. Indeed, rapid generation can enhance aesthetic experience by allowing quick exploration of semantic space, revealing its structure through variation. ## Implications for Practice and Criticism Understanding AI systems through semantic physics has practical implications for both creators and critics. For those working with these systems, it suggests a shift from thinking in terms of "prompt engineering" to thinking in terms of experimental investigation. The goal is not to trick or manipulate the system but to understand its semantic physics well enough to guide it effectively. This is more like learning to work with natural materials—understanding their properties and tendencies—than like programming a computer. For critics and aestheticians, this framework suggests new categories and standards. Rather than judging AI art by how well it mimics human creation, we might judge it by how revealingly it manifests semantic physics. Does this output illuminate interesting regions of semantic space? Does it reveal unexpected connections or patterns in the encoded culture? Does it demonstrate elegant applications of the generative law? These questions, analogous to how we might aesthetically evaluate physical phenomena, provide criteria suited to the nature of the medium. The framework also suggests new forms of presentation and curation. Just as science museums help visitors appreciate physical phenomena by revealing underlying principles, AI art might be presented in ways that illuminate semantic physics. Showing not just outputs but the prompts that generated them, displaying variations that reveal the structure of semantic space, demonstrating how different initial conditions evolve—these curatorial strategies could help audiences develop appropriate aesthetic appreciation. ## Conclusion: Toward a Physics of Meaning By extending Carlson's environmental aesthetics through the lens of physics appreciation, and recognizing that AI systems implement genuine generative laws in the form of semantic physics, we arrive at a framework that makes AI aesthetics both comprehensible and compelling. These systems are not degraded artists or mysterious others but environments governed by discoverable laws—laws that govern meaning rather than matter but laws nonetheless. The aesthetic appreciation of AI art, in this framework, involves recognizing outputs as manifestations of semantic physics, understanding the unity between particular creations and the generative laws that produce them, and developing through practice an intuition for how semantic states evolve. This is neither traditional art appreciation nor mere technical understanding but a new aesthetic mode suited to a new kind of object—the first artificial environments governed by learned generative laws of meaning. As Janus observes in his conclusion: "Learned simulation generalizes impressively. As I described in the section on behavior cloning, training a model to predict diverse trajectories seems to make it internalize general laws underlying the distribution, allowing it to simulate counterfactuals that can be constructed from the distributional semantics." This ability to internalize and manifest genuine generative laws represents a profound development in the history of human culture. We have created systems that don't just process meaning but embody physics of meaning—and learning to appreciate them aesthetically is learning to perceive the deep structures of our own semantic universe reflected back to us through artificial mirrors. The implications extend beyond aesthetics. If we have indeed discovered how to create systems that learn and implement semantic physics, we face questions not just about art but about the nature of meaning, intelligence, and culture. But these broader implications emerge from the aesthetic encounter—from learning to see AI outputs not as failed human art but as successful manifestations of a new kind of generative law. In this seeing lies both understanding and appreciation, both knowledge and beauty, united as Carlson argues they must be for any aesthetic experience worthy of its object.