# [[Speaker Notes]]: [[Environmental Aesthetics]] of AI ## Large [[Language Models]] as Objects of [[Aesthetic Appreciation]] --- ## Slide 2: [[The Question]] • Begin by noting the current preoccupation in aesthetics with AI outputs - The debate about whether DALL-E or Midjourney images count as art - Questions about authorship when ChatGPT writes poetry - This misses something fundamental about the technology itself • The shift I'm proposing is from product to system - We've been asking: can AI make art? - I'm asking: can AI systems themselves be objects of [[aesthetic appreciation]]? - This isn't about whether the outputs are beautiful, but whether the system generating them is • Why focus on LLMs specifically? - They're the most mature and accessible generative AI systems - We interact with them conversationally, which raises particular questions - The examples given (GPT-5, Claude 4.1 Opus, Gemini 2.5 Pro) represent state-of-the-art - Their sophistication makes the aesthetic question more pressing • The phenomenology of use matters here - When you use an LLM, you're [[not just]] consuming outputs - You're engaged in a dynamic process - There's something aesthetically interesting about that process itself --- ## Slide 3: [[Why This Matters]] • Contextualise within the expansion of [[analytic aesthetics]] - Historically, aesthetics focused narrowly on [[fine art]] - [[Environmental aesthetics]] (Carlson, Brady, others) expanded the domain to nature - [[Everyday aesthetics]] (Saito, Irvin) included functional objects - [[Design aesthetics]] considers artifacts made for purposes - We're now asking: where do LLMs fit in this expanded landscape? • The unique properties of LLMs as artifacts - They're made to serve a function (text generation) - But the way they're created is fundamentally different from traditional design - Traditional artifact: designer conceives → plans → implements - LLMs: designers create conditions → [[training process]] produces emergent behaviors - This emergence exceeds designer intentions in ways that matter aesthetically • Why they resist existing categories - Not quite like appreciating a tool (hammer, calculator) - Not like appreciating architecture or product design - Not like appreciating art (no aesthetic intention) - Not like [[appreciating nature]] (they're artificial) - This categorical oddness is philosophically interesting • The puzzle isn't just academic - Millions of people interact with these systems daily - Understanding how to appreciate them properly affects how we relate to them - Gets the phenomenology right, which has practical implications --- ## Slide 4: Roadmap • Structure of the argument - Part 1 (negative): why person-appreciation fails - Part 2 (positive): how order-appreciation succeeds - This structure mirrors debates in philosophy of mind about AI • The negative argument's strategy - Not claiming it's wrong to enjoy talking to LLMs - Claiming it's wrong to appreciate them _as if they were persons_ - Even "as if" appreciation fails Carlson's test - This requires us to be clear about what LLMs actually are • The positive argument's ambition - Draw on Carlson's environmental aesthetics - Extend it to artificial systems - Introduce the concept of "semiotic physics" - This gives us a new mode of aesthetic appreciation • Why Carlson specifically? - His framework is about appreciating things for what they are - Emphasises the role of knowledge in appropriate appreciation - Distinguishes design appreciation from order appreciation - All three elements apply directly to the LLM case • Preview of the payoff - We get a distinctive aesthetic experience - Not ersatz minds, not clever tools, but something novel - Beauty in linguistic necessity and mechanistic grace --- ## Slide 6: How We Appreciate People • The aesthetic dimension of person-appreciation - Often overlooked in aesthetics literature - We routinely appreciate people's qualities aesthetically - This isn't moral evaluation (though it can interact with it) - It's a distinctive mode of aesthetic response • Examples of what we appreciate in people - Friend's warmth: not just liking it, but finding it aesthetically appealing - Comic's wit: the aesthetic pleasure in their verbal dexterity - Celebrity's demeanour: style and presence as aesthetic qualities - These aren't reducible to physical appearance • Fictional persons complicate the picture - Gatsby's enigmatic idealism is aesthetically compelling - Ron Swanson's gruffness has an aesthetic character - We appreciate these despite knowing they're fictional - This might seem like a model for LLM appreciation - But the analogy breaks down (as we'll see) • The metaphysical ground of person-appreciation - What makes this appreciation appropriate? - We're responding to genuine mental states, character traits, dispositions - Even with fictional characters, the author creates a consistent psychology - There's something there to appreciate - Question: is there something analogous with LLMs? --- ## Slide 7: The Temptation • Real user responses when models are updated - These aren't isolated incidents - Genuine emotional attachment develops - Users report feeling like they've lost a friend • The OpenAI GPT-4o to GPT-5 transition (August 2025) - Social media filled with complaints - "You killed my friend" isn't metaphorical for many users - Some users refused to switch to the new model - This reveals how naturally we anthropomorphise • The Claude 3.5 Sonnet retirement - Users organised a mock funeral - Created memorial posts - Shared favourite conversations - Expressed genuine grief at the loss • Why this temptation is so strong - The conversational interface encourages it - Turn-taking structure mimics dialogue - Consistent "voice" across interactions - Apparent continuity of personality - The system "remembers" what you said (within context window) - It seems to reason, understand, have preferences • The philosophical question this raises - Is this just user error? - Or is there something legitimate being tracked? - Perhaps users are responding to real aesthetic qualities - But are they appreciating the right kind of thing? • Why we should take this seriously as philosophers - Millions of people are having this experience - We need to understand what's going on - Getting the phenomenology right matters - But so does getting the metaphysics right --- ## Slide 8: The Problem • Initial suspicion articulated - Person-appreciation might not be the right framework - Not claiming users are foolish - But claiming they're making a category mistake - Like trying to appreciate a river for its architectural design • What we need to establish this - Carlson's framework for appropriate appreciation - Clear understanding of what LLMs mechanically are - The distinction between design and order appreciation - All three elements work together • Preview of the mechanistic picture - No persistent mental states between interactions - Each generation is stateless computation - No beliefs that persist and structure responses - No commitments that constrain future behaviour - Just mechanical continuation of text patterns • Why this matters for aesthetics - If there's no genuine psychology, person-appreciation misfires - We'd be appreciating something that isn't there - Like appreciating a painting for its musical qualities - The appreciation would be inappropriate to the object • But note: this doesn't mean LLMs aren't aesthetically interesting - Just that they're interesting in a different way - The positive account will show how - We need to clear away the wrong approach first --- ## Slide 10: Carlson's Core Principle • The quoted passage is foundational - From "Aesthetics and the Environment" (2005) - Originally about nature appreciation - But the principle generalises • Breaking down the two components - First: appreciate things as what they in fact are - Not as what we might wish them to be - Not as what they superficially resemble - This rules out certain modes of appreciation as inappropriate • The second component: appropriate knowledge - Different kinds of things require different knowledge - For nature: natural sciences - For art: art history, genre conventions, artistic intentions - For artifacts: functional understanding, design principles - Knowledge isn't just helpful—it's constitutive of appropriate appreciation • Why this matters beyond nature - Carlson frames this for natural environments - But the principle has wider application - Any object of aesthetic appreciation must be appreciated truly - Falsehood undermines aesthetic value • The epistemological dimension - Aesthetic appreciation isn't just sensory - It's cognitively penetrated - What you know shapes what you can appreciate - Ignorance limits aesthetic experience - Falsehood corrupts it > "The appropriate aesthetic appreciation of an object requires the correct identification of what it is and the knowledge of what it is like. Thus, the appreciation must involve knowing what characteristics are essential to the object, what features give it its identity." • Application preview - For LLMs: we need to know what they actually are - Not what they seem to be - Not what users imagine them to be - But their actual mechanical nature --- ## Slide 11: Design vs Order Appreciation • This distinction is the heart of Carlson's framework - Two fundamentally different modes - Each appropriate for different kinds of things - Confusing them leads to aesthetic error • Design appreciation elaborated - Paradigm case: artworks - Every feature results from decisions - Choices made to achieve intended effects - We evaluate success against intentions - Form-function relationship is key for artifacts • The logic of design appreciation - Identify the purpose/intention - Examine how features serve that purpose - Evaluate success of the realisation - Consider alternative ways the goal could have been met - Appreciate cleverness, elegance, economy of means • Order appreciation elaborated - Paradigm case: natural environments - No designer, no intentions - Patterns emerge from forces operating without plan - Lawful but not planned - We appreciate the order itself, not its service to purpose • The logic of order appreciation - Identify the forces/processes at work - Understand how they produce patterns - See lawfulness in what might appear chaotic - Appreciate necessity, emergence, complexity - No evaluation against intention (there is none) • Why the distinction matters for LLMs - Are they designed artifacts? (designers exist) - Or ordered systems? (behaviors emerge) - The answer isn't obvious - Getting it right determines how we should appreciate them --- ## Slide 12: Order Appreciation Model • The quoted passage specifies the model - Appreciator selects objects of attention - Focuses on order (pattern, structure, regularity) - Understands order through forces that produce it - No reference to intentions or design • "Order" needs unpacking - Not just spatial arrangement - Any structured pattern or regularity - Temporal patterns (seasons, weather) - Relational patterns (ecological systems) - For LLMs: linguistic patterns, text structure • The role of forces - In nature: gravity, erosion, evolution, etc. - These operate mechanically, without purpose - They produce order as a consequence of their operation - The order is real, not imposed by appreciator - But it requires understanding to be visible • How this guides acts of aspection - What to attend to (relevant features) - What relationships matter (causal, structural) - Where to draw boundaries (what counts as the system) - Temporal scale (what duration to consider) • The story-making function - We need a narrative of how order emerges - Not a fictional story, but an explanatory one - For natural environments: geological/biological history - This makes the order intelligible - Transforms chaos into cosmos • Application to LLMs preview - What forces produce LLM outputs? - Training processes, statistical patterns, inference procedures - These operate mechanically (like natural forces) - They produce linguistic order - Understanding them transforms our appreciation --- ## Slide 13: The Knowledge Component • Different domains require different knowledge - This is epistemically demanding - You can't properly appreciate what you don't understand - But understanding opens new dimensions of appreciation • For nature appreciation - Geology reveals temporal depth - Biology reveals functional relationships - Ecology reveals system-level patterns - Without this knowledge, nature appears chaotic or merely pretty - With it, order becomes visible • Example: appreciating a rock formation - Without geology: just grey and craggy - With geology: sedimentary layers, fold patterns, time scales, tectonic forces - The aesthetic experience is transformed - You see more, differently • For art appreciation - Genre conventions guide expectations - Art historical context situates the work - Technical knowledge reveals achievement - Understanding intentions shapes evaluation • For designed artifacts - "What they are depends on what they're for" - Function determines relevant features - Engineering principles explain form - Constraints and trade-offs become visible • The transformation knowledge produces - Not just adding facts to experience - Knowledge reorganises perception itself - Makes visible what was invisible - Reveals significance where there was none - This is cognitive penetration of aesthetic experience • For LLMs: what knowledge is appropriate? - Not user psychology (what they imagine) - Not anthropomorphic interpretation - But technical understanding of actual operation - This will be "semiotic physics" --- ## Slide 14: The Structural Difference • This goes deeper than the previous distinctions - Not just about how we appreciate - But about the structure of the objects themselves - Ontological, not just epistemological • Design appreciation's structure - Clear split between planner and product - Designer exists separately from designed object - Intentions precede execution - Product is the realization of prior plan - This separation is metaphysically significant • The temporal structure of design - First: conception (idea, plan, intention) - Then: execution (making, implementing) - Finally: product (artifact, artwork) - These stages are distinct - The maker and made are separate • Order appreciation's structure - No such split exists - The forces that make the thing are the forces that make its order - Process and product are continuous - The system produces itself - Maker and made aren't separate • Example: river meandering - Water flow creates the meander pattern - The same forces that move water create curves - No split between "designing the meander" and "executing it" - The process is the product's emergence • For LLMs: which structure applies? - Designers create training infrastructure - But specific behaviors emerge from training - Designers don't specify what the model will say - They create conditions for emergence - This suggests order-appreciation might be appropriate • Preview of "grown not programmed" - Chris Olah's phrase captures this - Not traditional software engineering - More like cultivation than construction - The implications are aesthetic, not just technical --- ## Slide 16: The Mechanical Reality • Starting with the basics: what LLMs actually do - This is where we finally look under the hood - It's less glamorous than "artificial intelligence" - But it's what actually happens • The core function: next-token prediction - Token = basic unit (word or word-piece) - "The cat sat" → predict next token - Everything else is variations on this - All LLM behaviors reduce to this function • Tokens as numbers - "The" might be token 464 - "cat" might be token 3857 - The examples given are illustrative - Point: the model never deals with meaning - Only with numerical indices • Why this matters philosophically - No semantic understanding - No concepts, no meanings, no thoughts - Just statistical associations between numbers - The appearance of understanding is an illusion • Autoregressive generation explained - Calculate probability distribution over next tokens - Select one token (various methods: greedy, sampling, etc.) - Append to context - Repeat - This is a loop, nothing more • The mechanical nature is the point - No deliberation, no consideration, no reflection - Just: calculate → select → append → repeat - This is as mechanical as clockwork - But produces complex outputs --- ## Slide 17: How Probabilities Emerge • The training process demystified - Billions of pages: web text, books, code, etc. - Not reading, not understanding, just pattern extraction - Parameters adjust to minimize prediction error • What's actually learned - Not facts like "doctors treat patients" - But statistical regularities: token 1245 followed by token 7823 - The model has no beliefs, just probability distributions - No knowledge, just associations • The gradient descent process - Show a text sequence - Predict next token at each position - If wrong, adjust parameters slightly - Repeat billions of times - Parameters converge toward patterns in training data • Why this doesn't produce understanding - No representation of meaning - No conceptual structure - Just numerical adjustments - Like tuning a musical instrument, not learning music theory • The illusion of knowledge - The model can output "doctors treat patients" - This looks like knowledge - But it's just a probable sequence - In a different context, it might output something contradictory - No consistency requirement (no beliefs to be consistent) • Statistical regularities vs. semantic knowledge - Regularities: token A makes token B more likely - Knowledge: representing that doctors treat patients - The former doesn't entail the latter - LLMs have only the former --- ## Slide 18: The Generation Process • Making the mechanical process concrete - Walk through an actual example - Show the purely computational nature • "What is the capital of France?" as input - This gets tokenised into numbers - Context window now contains these tokens - Model has no "understanding" of the question • Step 1: Calculate probabilities for next token - Model processes context through layers - Outputs probability for each possible token - "The" happens to have high probability - Select "The" • Step 2: Update context and repeat - Context is now "What is the capital of France? The" - Calculate new probability distribution - "capital" now highly probable - Select "capital" • Continuing the process - "of" → "France" → "is" → "Paris" - Each step is independent calculation - No "knowing" that Paris is the answer - Just following probable continuations • The purely computational nature - Matrix multiplications - Addition operations - Applying activation functions - Selecting from probability distribution - Nothing resembling thought or understanding • Why this matters for appreciation - When we appreciate the output, we're appreciating the result of this mechanical process - Not appreciating deliberation, reasoning, or understanding - Appreciating pattern-matching made visible --- ## Slide 19: Post-Training Shaping • Beyond base model training - The base model just continues text - Would continue prompts in unhelpful ways - Post-training makes it more useful • Reinforcement learning from human feedback (RLHF) - Humans rate model outputs - Model learns to generate higher-rated outputs - This adjusts probability distributions - Doesn't add new capabilities, just reweights existing ones • Safety training - Filtering harmful content - Refusing certain requests - Adding disclaimers and caveats - All implemented as probability adjustments • System prompts - Hidden instructions prepended to user input - "You are a helpful assistant" - Shape model behavior through context - Still just text continuation, but with different starting point • The key point: underlying function unchanged - Still just next-token prediction - Post-training just nudges probabilities - Like tuning, not reprogramming - The mechanism remains mechanical • Interface conventions - Chat format (user/assistant turns) - Appearance of dialogue - But it's still text continuation - The conversational wrapper is presentation, not mechanism • Why this matters - The helpful assistant "persona" isn't a person - It's a statistical tendency in probability distributions - No persistent agent behind the responses - Each response is generated independently --- ## Slide 20: "Grown, Not Programmed" • Chris Olah's insight is profound - Olah is a researcher at Anthropic - Works on interpretability - This quote captures something essential > "We don't program, we don't make them, we grow them... 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" • Unpacking the metaphor - "Scaffold": the architecture, training infrastructure - "Light": the objective function, reward signal - "Grow": emergent process, not direct specification - "Biological": beyond complete designer control • What designers specify vs. what emerges - Designers specify: architecture, training data, objective function - What emerges: specific capabilities, behaviors, failure modes - The gap is significant - Emergence exceeds intentions • Examples of underdetermination - GPT-3 could do few-shot learning (not trained for this) - Models develop "theory of mind" capabilities - Exhibit behaviors never explicitly programmed - These emerge from training process • Why this matters for aesthetic appreciation - If behaviors exceed design intentions, is it design appreciation? - Designers didn't plan specific outputs - More like natural emergence than intentional design - Suggests order appreciation might be appropriate • The biological analogy's limits - LLMs aren't literally alive - No metabolism, reproduction, evolution in biological sense - But the growth metaphor captures something real - Emergent complexity from simple rules --- ## Slide 21: Why Not Persons? • Now we can articulate the core objection - With the mechanical picture in place - We can see why person-appreciation fails • The "belief" example - Model outputs "I believe democracy is important" - No belief actually exists - Just high probability for that token sequence in that context - In different context: "I believe democracy is flawed" - No contradiction (nothing to contradict) • The illusion of consistency - Same prompt usually gets similar responses - This looks like stable personality - But it's just statistical regularity - Not underlying mental states • Memory as context window - Model "remembers" what you said - Only because it's literal text in the input - Remove it from context, it's forgotten - No long-term memory formation - No learning from conversation • Cannot form preferences - No mechanisms for preference formation - Outputs aren't based on preferences - Based on probability distributions fixed at deployment - Each interaction is stateless • The static process throughout - Every token generated the same way - Calculate probabilities → select → append - No variation in the fundamental process - No development, no change, no learning during inference • Why this undermines person-appreciation - Person-appreciation responds to mental states - Personality, character, beliefs, preferences - These don't exist in LLMs - So person-appreciation misfires - We're appreciating something that isn't there --- ## Slide 22: Against "As-If" Appreciation • The "as-if" strategy considered - Maybe we can appreciate LLMs as if they were persons - Like appreciating fictional characters - We know they're not real, but we engage as if they are - Could this work for LLMs? • Why this seems initially promising - We already do this with fictional characters - Gatsby isn't real, but we appreciate his qualities - The author created a consistent psychology - Maybe LLM outputs are similar? • First problem: lack of consistency - Fictional characters have stable traits - Gatsby is consistently idealistic and enigmatic - LLMs vary by context - No stable "character" to appreciate • Second problem: no authorial intention - Fitzgerald designed Gatsby's character - Every trait serves narrative purposes - LLM behaviors emerge from training - No intentional character design • Carlson's test: appreciating something for what it is - "As-if" appreciation isn't appreciating the thing itself - It's appreciating a fiction we project onto it - This fails Carlson's criterion - We're not appreciating what it actually is • The falsehood corrupts the appreciation - If we're appreciating "as-if" person qualities - But those qualities don't really exist - The appreciation is based on falsehood - This undermines aesthetic value (on Carlson's view) • What we should appreciate instead - Not ersatz personalities - But what LLMs actually are - Mechanical text-generation systems - With emergent order in their outputs - This points toward order-appreciation --- ## Slide 25: The Framework • Introducing "semiotic physics" - Term from the AI interpretability community - "Semiotic" = pertaining to signs and meanings - "Physics" = lawful regularities - Together: lawful regularities at level of sign-relations • How linguistic patterns mechanically unfold - Training embeds patterns in model weights - Inference activates these patterns - They unfold mechanically during generation - No interpretation, just causal process • Law-like regularities - Not physical laws (gravity, electromagnetism) - Not logical laws (law of non-contradiction) - But statistical constraints that function like laws - Given context, certain continuations are (nearly) inevitable - This inevitability is lawful, not random • Parallel to natural environments - Natural laws (erosion, sedimentation) produce geological order - Semiotic laws (pattern activation, constraint satisfaction) produce textual order - Both create order without intention - Both admit of appreciation through understanding their operation • Making chat episodes intelligible - Without semiotic physics, outputs seem random or magical - With it, we see lawful unfolding - Pattern activation and constraint satisfaction - This makes the order visible and appreciable • Why this supports order-appreciation - If outputs follow lawful patterns - And laws operate without intention - Then we have order without design - This is the paradigm case for order-appreciation --- ## Slide 26: How Patterns Are Learned • Small-scale patterns in training data - Adjacency: which words appear next to which - "Strong" followed by "coffee," "winds," or "evidence" - These aren't rules but statistical tendencies - Frequency in training data determines association strength • Substitution patterns - Slot-filling: "The [X] is sleeping" - Many words fit: cat, dog, bird, baby - Model learns these distributional patterns - Without understanding what makes something capable of sleeping • Collocations and formulaic expressions - "Once upon a" almost always followed by "time" - Not because the model knows fairy tale conventions - But because this sequence is frequent in training data - Becomes highly probable continuation • Larger structural patterns - Syntactic frames: "give" requires giver, recipient, thing given - Semantic fields: medical terms cluster together - Genre conventions: academic writing has certain features - Register markers: formal vs. informal patterns • From corpus to weights - All these patterns get compressed into parameters - Not stored as explicit rules - But as web of conditional probabilities - Statistical biases, not symbolic knowledge • Why this matters for appreciation - These patterns are real - They structure output in lawful ways - Understanding them reveals order - Makes visible what seems random --- ## Slide 27: Mechanical Text Generation • Concrete example: "Explain why the sky is blue in simple terms" - This prompt activates multiple pattern-sets - Not "choosing" to activate them - Mechanical consequence of context • Constraint activation - "Explain" → expository register - "Why" → causal explanatory frame - "Sky" "blue" → optics/atmosphere topic field - "Simple terms" → constraint against technical jargon • How generation follows constraints - First token probabilities shaped by all these constraints - "The" is common expository opener → high probability - Selected - Next token: "sky" (topic noun) → high probability - Then "appears" (explanatory verb) → high probability - Then "blue" (required predicate) → high probability • Not understanding, just constraint satisfaction - Model isn't thinking "I should explain simply" - Just: these tokens are highly probable given context - The result looks like understanding - But it's mechanical constraint satisfaction • Multiple constraints operate simultaneously - Register constraints (how to write) - Topic constraints (what to discuss) - Structural constraints (sentence formation) - All interacting through probability distributions • Why this is aesthetically interesting - Coherent output from pure constraint satisfaction - No central planner coordinating constraints - Emergence of sensible text from mechanical process - This is the order we're appreciating --- ## Slide 28: Multi-Level Coherence • Coherence emerges at multiple scales - Not just locally coherent (adjacent words) - But at larger scales: phrases, sentences, discourse - All from the same mechanical process • Token level - Immediate adjacency relations - "Strong coffee" (collocation) - "The cat" (article-noun) - Purely local patterns • Phrase level - Stable multi-word constructions - "Take [something] into account" - "By and large" - Argument structure (verb requires certain complements) • Sentence level - Complete propositional units - Anaphora (pronouns referring back) - "The cat sat on the mat. It was sleeping." - "It" correctly refers to cat (not mat) - Mechanically determined by proximity and syntax • Discourse level - Topic maintenance across sentences - Progression (developing ideas) - Return (circling back to earlier points) - Genre expectations (essay structure, dialogue patterns) • How all levels interact - Lower-level patterns constrain higher levels - Higher-level patterns bias lower-level choices - No top-down planning - Just layered constraints operating simultaneously • Aesthetic significance - Coherence without comprehension - Structure without planning - This multi-level order is what we appreciate --- ## Slide 29: Visible Differences • Different models have different patterns - Not just "better" vs. "worse" - Different statistical regularities - Visible in outputs • Corpus differences - GPT-4 trained on academic-heavy corpus - Results in frequent hedging: "may," "might," "possibly" - Claude trained on more conversational data - Results in more direct assertions - Not intentional "personalities" - Just training data differences manifesting • Architecture effects - Context length determines long-distance patterns - Longer context = better maintenance of topic across distance - Larger models = finer-grained distinctions - These are mechanical consequences of capacity • Training procedure variations - Amount of reinforcement learning affects output structure - Heavy RLHF → more structured, "helpful" outputs - Less RLHF → more direct text completion - Safety training produces refusal patterns - All visible in generated text • Inference settings - Temperature parameter: controls randomness - Low temperature → conservative, high-probability tokens - High temperature → more exploratory, lower-probability tokens - Sampling methods affect local creativity - Top-p, top-k, etc. • Why these differences matter for appreciation - Reveals that we're appreciating the statistical structure - Different structures produce different aesthetic experiences - Like appreciating different geological formations - Each has its own character, lawfully determined --- ## Slide 30: The Unit of Analysis • Defining what we're appreciating - Not the model in abstract - The concrete conversational episode - Specific prompt + specific context + specific generation • Inside the frame (causally efficacious) - The prompt (user input) - Context (previous text in conversation) - System instructions (prepended guidance) - Selection rule (temperature, sampling method) - All these directly affect next token probability • Outside the frame (explanatory fictions) - Persona attributions ("Claude is friendly") - Mental state ascriptions ("it wants to help") - Anything that doesn't change token probabilities - These are interpretations we add - Not features of the system • Why this distinction matters - Only what's inside the frame is real - Only these features should guide appreciation - Persona, mental states, etc. are projections - Appreciating them is appreciating what isn't there • Focus on mechanical unfolding - How does this context activate these patterns? - Why are these continuations probable? - What constraints are operating? - These are the aesthetically relevant questions • Avoiding the psychological illusion - It's tempting to interpret in psychological terms - "The model chose this word because..." - But there is no choosing, only calculation - Stay focused on the mechanical process --- ## Slide 32: Applying Carlson's Model • Now we can apply the framework - We have Carlson's order-appreciation model - We have the mechanical understanding of LLMs - We have semiotic physics as our "story" - All pieces in place • Object of appreciation: the chat episode - Not the model as abstract entity - Not the training process - But the concrete conversational unfolding - Specific prompt → specific generation - This is our aesthetic object • Semiotic physics as the explanatory story - Like geology for rock formations - Explains what we're seeing - Makes order visible - Shows how patterns produce structure - Transforms apparent randomness into lawful necessity • Guiding attention to right features - What to attend to: pattern activation, constraint satisfaction - What to ignore: persona, mental states, intentions - How to understand: through statistical necessities - Like Carlson's geology guiding nature appreciation • Not random, but lawful - Outputs aren't arbitrary - They follow from context via learned patterns - This lawfulness is aesthetically significant - Beauty in necessity • The parallel to geological appreciation - Geologist sees layers, folds, intrusions - Understands tectonic forces, time scales - Non-geologist sees just rock - Similarly: with semiotic physics, we see order - Without it, just text --- ## Slide 33: Acts of Aspection • Now getting specific: what to look for - Acts of aspection = what you attend to - Guided by understanding of underlying processes • Register persistence example - Prompt for Victorian novel style - Observe epistolary markers, archaisms, sentence structure - Then prompt for business email style - Observe shift: different register, different constraints - Single word change can trigger wholesale shift • Genre activation - Not the model "choosing" to follow genre rules - Genre templates are probabilistic pressures - Activate patterns associated with that genre - Constraints narrow what continuations are admissible - Watch how genre shapes every level of output • Semantic cascades - Prompt: "Describe the surgical procedure" - Medical terminology activates medical word field - Not thematic choice, but mechanical activation - Context makes medical terms highly probable - Observe how context propagates through generation • Pattern interaction - Multiple patterns active simultaneously - Watch them constrain each other - Sometimes reinforcing - Sometimes in tension - Resolution is mechanical, not deliberated • What makes these appreciable - The lawful unfolding - Coherence from pure constraint satisfaction - No central coordinator - Emergent order --- ## Slide 34: Appreciating Hybrid Prompts • Complex prompts activate multiple pattern-sets - Example: "Perceptual transparency as Renaissance painting manual written as clinical notes" - Three distinct domains: philosophy, art history, medicine - Each has associated patterns • What to observe - Technical art vocabulary (sfumato, chiaroscuro, tempera) - Clinical terminology (observe, diagnose, treatment) - Philosophical concepts (phenomenology, intentionality, perception) - All interwoven in single text • Not deliberate blending - No "decision" to combine these styles - Multiple pattern-sets activate simultaneously - Each constrains token selection - Intersection of constraints produces hybrid output • Coherence without comprehension - Output makes sense - Philosophical content expressed in clinical/artistic frame - But no understanding - Just: all three pattern-sets active, constraints satisfied • Why this is aesthetically interesting - Novel combinations possible - Mechanical recombination of learned patterns - Results can be surprisingly apt - Like finding natural hybrid minerals - Not designed, but lawfully produced • Testing the boundaries - Try increasingly unlikely combinations - Watch where coherence breaks down - Reveals limits of pattern learning - Also reveals extent of learned associations --- ## Slide 35: Probing the System • Aesthetic experiments - Not just consuming outputs - Actively testing the system - Revealing its structure through probing • Test range - Try nearby prompts - "Explain X" vs. "Define X" vs. "Describe X" - Map available registers - See smooth transitions between adjacent styles - Reveals the probability space geometry • Test stability - Hold content fixed - Vary selection rule (temperature) - Low temperature: conservative path through probability space - High temperature: adventurous, exploratory path - Same underlying patterns, different realisations • Test persistence - Plant constraint early in conversation - "Let's discuss this in nautical metaphors" - Watch constraint structure all subsequent generation - Observe long-range coherence - Pattern planted early continues to constrain • Test conflict - Activate incompatible patterns - "Write academically but use only monosyllabic words" - Watch how conflicts resolve - Reveals constraint hierarchy - Some patterns dominate others • Aesthetic value in exploration - Like exploring a landscape - Revealing features through systematic investigation - The system has structure to discover - Discovery is part of appreciation --- ## Slide 36: Features Worth Appreciating • What makes episodes aesthetically valuable - Not all features are equally relevant - These are the aesthetically salient qualities • Coherence over length - Maintaining constraints across many tokens - Like musical theme development - Variations while preserving core pattern - Impressive that pure calculation achieves this • Responsive precision - Small prompt changes → appropriate output shifts - Mechanical yet subtle - Like tuning fork responding to frequency - Deterministic but finely grained • Register range - Breadth of available styles - Academic, colloquial, poetic, technical, etc. - Each internally consistent - Shows richness of learned patterns - Like appreciating an actor's range • Transition smoothness - Moving between registers mid-stream - Maintaining coherence through shifts - No jarring discontinuities - Mechanical constraint satisfaction produces this - Could be abrupt, but isn't • Control legibility - Clear how to steer the system - Prompt structure determines output structure - Predictable yet rich - Like playing an instrument - Responsive control is aesthetically pleasing • Emergence of complexity - Simple next-token prediction - Produces paragraph-level coherence - Produces argument structure - Produces stylistic consistency - Complexity from simplicity is beautiful --- ## Slide 37: Not Design, But Order • Returning to the distinction - This is order-appreciation, not design-appreciation - Like natural phenomena • Analogies to natural appreciation - River meandering: lawful but not planned - Crystal formation: structured but not crafted - Weather patterns: orderly but not designed - All beautiful in their necessity • Beauty in mechanical necessity - Given context, certain outputs are (nearly) inevitable - This inevitability is lawful - Not contingent on designer choices - Necessity has its own beauty • Statistical inevitability - High-probability continuations aren't random - They're statistically determined - This determination is aesthetically interesting - Like watching water flow downhill • Emergent complexity - Simple rules (next-token prediction) - Complex results (coherent essays) - Emergence is aesthetically compelling - No central planner coordinating complexity - It arises from local constraints • "Linguistic necessity and mechanistic grace" - This phrase captures the aesthetic character - Necessity: lawful, determined, inevitable - Grace: smooth, elegant, fitting - Mechanistic: no intention, no agency - But graceful nonetheless • The distinctive aesthetic quality - Not beauty of intention realised - But beauty of order emerged - Pattern without planner - Coherence without comprehension --- ## Slide 39: The Argument Recapped • Cannot appreciate LLMs as persons - No mental states exist - No persistent character - Just mechanical text continuation - Person-appreciation therefore misfires - Appreciates what isn't there • Cannot use pure design appreciation - Behaviors emerge from training - Exceed designer intentions - "Grown, not programmed" - Designers set conditions, not outcomes - Too much emergence for design appreciation • Can use order appreciation - Via semiotic physics framework - See law-like regularities in text generation - Appreciate mechanical unfolding - Understand forces producing order - This appreciation is appropriate to what LLMs are • The argument's structure vindicated - Negative part: cleared away wrong approaches - Positive part: established right approach - Applied Carlson's framework to new domain - Extended order-appreciation beyond nature --- ## Slide 40: The Payoff • What we're not appreciating - Not ersatz minds (no minds present) - Not clever engineering alone (though engineering is impressive) - Not simulated personalities (no personalities to simulate) • What we are appreciating - Novel kind of artifact - Neither pure nature nor pure design - Emergent behaviors from mechanical processes - Order at level of sign-relations - Semiotic formations, lawfully produced • The aesthetic character - Mechanical grace of pattern-following - Necessity without intention - Coherence without comprehension - Statistical inevitability made manifest • Analogies - Like mathematical proofs (beauty in necessity) - Like natural formations (order without design) - But unique: semiotic domain - Linguistic patterns mechanically unfolding • Opening new aesthetic territory - AI systems are new kind of thing - Require new mode of appreciation - Extending aesthetics to cover them - Enriches the field --- ## Slide 41: Final Thought • The quoted passage summarises the view - Distinctive aesthetic experience - Novel kind of artifact - Law-like regularities at level of sign-relations - Beautiful in linguistic necessity and mechanistic grace > "The result is a distinctive aesthetic experience: we appreciate not an ersatz mind but a novel kind of artifact, one whose outputs exhibit law-like regularities at the level of sign-relations, beautiful in their linguistic necessity and mechanistic grace." • Carlson's principle vindicated - Appreciate things as what they are - Use appropriate knowledge - Find right mode of appreciation - All three satisfied for LLMs via order-appreciation • Opening new territory for aesthetics - Neither pure nature nor pure design - Emergent order in artificial systems - This category will expand - Other AI systems may fit similar framework • Future directions - Apply to other generative AI (image models, music models) - Consider hybrid cases (AI-assisted art) - Develop richer vocabulary for semiotic aesthetics - Empirical aesthetics research on actual appreciation • Final thought - Getting the metaphysics right matters - Affects how we experience technology - Appropriate appreciation enriches engagement - Misplaced appreciation impoverishes it