# Slide 1 — Aim and scope - Can we aesthetically appreciate AI? - In particular, can we aesthetically appreciate LLMs, such as GPT-5, [[Claude Opus]], [[Google Gemini]] 2.5 Pro? - Whether and how we can appreciate the system is closely connected to appreciating the outputs of LLMs, for example a poem coaxed from an LLM by a prompter. - We’re not going to explore this subquestion today, just flagging it up. - This is interesting in its own right. - These systems are unlike other sorts of objects in some ways. - So it is interesting to think about them as potential targets for [[aesthetic appreciation]] # Slide 2 — Carlson in outline - Carlson’s view has two core commitments and a set of operational consequences. - First, _what it is in fact is_. - Appropriate appreciation is “centred on and driven by the real nature of [[the object]] of appreciation itself.” - This is a constraint on category-fixing: in each domain, kind is determined by the facts that constitute the thing - history of production for artworks - [[natural history]] for nature - function and realisation for [[designed artefacts]] - Misclassify the kind and one’s attention is misdirected. - Second, _in light of the right knowledge_. - For nature, this is the natural and environmental sciences - for art, art-historical knowledge - for artefacts, knowledge of function and of how that function is realised. - The knowledge disciplines “acts of aspection”: what to attend to, where to draw boundaries, which features are aesthetically salient. # Slide 3 — Working parts (1–2): Kinds, blueprint, and objectivity - From these, we get a view of aesthetics with the following features. 1. Kinds and the blueprint Carlson’s “blueprint” pairs kind-fixing with [[right knowledge]]. - In art, frames/categories fix boundaries and foci (e.g., _Guernica_ as a painting [[guides attention]] to pictorial structure). - In nature, [[scientific understanding]] replaces frames with environment-relative boundary tests (the wind across a valley is inside; the cough in a concert hall is not). - In [[designed artefacts]], kind is fixed by function and its mode of realisation (what it is for, and how that purpose is carried out) 2. [[Order appreciation]] and process–product pairing - For environments, appreciation targets [[the order]] imposed by forces producing what is present. - Scientific lenses disclose a duplex register: _natura naturans_ (generative processes) and _natura naturata_ (generated configurations). - Acts of aspection move across the two—e.g., from shoreline forms to tidal cycles and sediment transport that shape them. 3. Boundary tests and foci - Correct kind-knowledge yields current boundaries, foci (what to include/exclude), and acts of aspection (how to attend). - Procedure: fix the kind via domain-constitutive facts; use the right sciences to discipline attention; adopt boundary tests that follow from the kind; derive acts of aspection answering to [[the object]]’s nature. [[This approach]] excludes both formalist reduction and stance-driven projection: attention must track [[the object]]’s nature # Slide 4 — Transition to LLMs - Applying Carlson to LLMs such as GPT-5. - An LLM is a neural network with billions of parameters whose learned weights encode statistical patterns in language, enabling it to predict the next token and generate coherent text. %% - This needs to be a much better definition - It would also be good to distinguish between an LLM itself (e.g. Claude 4.1 Opus) vs a particular chat instance (*this chat* with Claude 4.1 Opus) %% - Carlson’s middle domain is the natural home for LLMs: they are designed artefacts. - The blueprint therefore tells us to (i) fix kind by _function and mode of realisation_; (ii) discipline appreciation by the “right knowledge” for that kind; (iii) set boundaries appropriate to this artefact _now_; (iv) derive acts of aspection accordingly. %%this bullet point needs to reflect structural changes that have been/will be made%% # Slide 8 — Kind-fixing for LLMs - **Kind-fixing.** - An LLM is engineered to learn a conditional continuation rule over token sequences during training (next-token prediction) and to apply that rule at use in an autoregressive loop. - Its _function_ is learned conditional continuation; its _mode of realisation_ is a particular computational architecture (e.g., transformer self-attention with tokenisation, positional encoding, and a decoding policy). - Dialogue, helpfulness, or personality are interface- and alignment-level overlays, not constitutive of the artefact’s kind. - Under Carlson’s test, dialogue cannot fix kind; learned continuation can. # Slide 9 — Right knowledge for LLMs - **Right knowledge.** - The relevant sciences are not folk psychology but the sciences and engineering of trained sequence models: training objective and loss, tokenisation, attention mechanisms, decoding policies, alignment layers. - At a complementary descriptive level, semiotic analysis supplies a non-mentalistic vocabulary for outputs and prompts: tokens as _representamens_, prompts as semiotic acts that perturb constraints, corpora as a sample of the _semiosphere_ of cultural codes. - This reframing avoids anthropomorphism while explaining what the system in fact does—organising and recombining signs under learned statistical constraints. # Slide 10 — Boundary tests for LLM appreciation - **Boundary tests.** - What counts as part of the setting for appreciation _now_ is: the fixed weight–rule (the learned continuation operator), the current context window (state), the decoding policy and safety constraints in effect, and the interface conventions that present trajectories. - External user aims or institutional purposes are not constitutive; they guide use but do not reclassify the object’s kind. - Carlson’s own boundary discipline motivates excluding ad hoc ends that are “imported” rather than discovered in the artefact’s nature. # Slide 11 — Acts of aspection for LLMs - **Acts of aspection.** - Given the kind and boundaries, attention should track (i) trajectories rather than single tokens (the autoregressive evolution is the dynamic object); (ii) sensitivity to perturbations in initial conditions (prompt edits), policy settings, and constraints; (iii) code-activation patterns in outputs (register, genre, rhetoric) as signs of the learned continuation rule’s operation; and (iv) stability and transitions across discourse zones (“register mixing”) as the model navigates the learned semiosphere. - These are the analogue, for this artefact, of surveying a prairie versus scrutinising a forest floor: they follow from what the thing is. # Slide 12 — Generative environment; agentive stances relegated - On this reading, Carlson’s blueprint does double work. - It _excludes_ stance-based, agentive framings as primary lenses—useful heuristics, perhaps, but not lights that answer to the artefact’s nature—and it _licenses_ a non-mentalistic, architecture-aware light (training and inference mechanics, supplemented by semiotic description) under which LLM behaviour becomes aesthetically appreciable as a _generative environment_: the learned rule (_machina naturans_) producing token trajectories (_machina naturata_) in response to local perturbations. - This also clarifies where familiar “as-if participant” talk sits. - Accounts that shift the locus of value to a human–AI interaction often lean on “participation” and “dialogue” metaphors. - As an _as-if_ stance within an artist-structured practice this is intelligible, but Carlson’s method warns against elevating the stance into the primary lens for appreciating the artefact. - The right knowledge and kind-fixing retain primacy. # Slide 13 — Framing plurality: lights without relativism - Framing the plurality of “right knowledge” in Carlson. - Carlson’s formula—appreciate x “as what it in fact is” and “in the light of the right knowledge”—is pluralistic but constrained. - Pluralistic, because multiple sciences can be “right” of the same environment; constrained, because each admissible _light_ must (i) answer to the same kind-fixing facts, (ii) respect boundaries appropriate to that kind here and now, and (iii) yield disciplined acts of aspection rather than free projection. - A helpful way to see this is to treat each light as a scale- and process-sensitive map from kind to attention. - Geomorphology, ecology, and microclimatology all study _the very environment_—not a proxy—because each tracks objective structures and processes that constitute that environment at different scales. - They are not mere descriptions pasted on top; they are ways of latching onto determinants of the present order (e.g., dune migration, trophic interactions, convective cycles). - The lights are complementary: each sets boundaries (what counts as inside the scene), foci (what features matter now), and acts of aspection (how to look) that are answerable to the same object. - Plurality without relativism. # Slide 14 — Constraints on plurality - Two constraints keep the plurality disciplined. - First, _constitutive tie_: a light is “right” only if its generalisations are keyed to processes and structures that in fact determine, maintain, or transform what is present (the naturans–naturata pairing). - Second, _boundary fidelity_: a light must not smuggle in extrinsic ends that reclassify the object (e.g., treating a coastline as a real-estate asset is not a light of the coastline _as environment_). # Slide 15 — Transposing multi-light discipline to LLMs (I) - For designed artefacts the kind is fixed by function and mode of realisation. - With LLMs, the function is learned conditional continuation; the realisation is an implemented operator (e.g., transformer) plus decoding policy and constraint stack. - Within that kind, several lights can be “right” in Carlson’s sense. - Each studies _the artefact itself_ at a distinct level while respecting the same boundaries. - First, the _mechanistic_ light (training and inference). - This fixes the learned operator and its behaviour under context, decoding policy, and constraints. - It licenses acts of aspection that track trajectories rather than isolated tokens, sensitivity to prompt perturbations, and policy-dependent branching. - Second, the _data-regime_ light (corpus composition and weighting). - This is not sociology tacked on; it concerns the frequency, coverage, and co-occurrence structure that shape the learned operator. - It directs attention to distributional artefacts (register priors, genre defaults, long-tail brittleness) that manifest in outputs because they are consequences of the training dynamics. # Slide 16 — Transposing multi-light discipline to LLMs (II) - Third, the _semiotic_ light (non-mentalistic sign structure). - Tokens are treated as sign-forms; prompts as constraint-setting acts; continuations as trajectories through a learned semiosphere. - This light studies the same artefact by making the code-level regularities and transitions (register shifts, genre uptake, idiom blending) available to aspection without positing beliefs or intentions. - Fourth, the _constraint-stack_ light (alignment, safety filters, system prompts, tool routers). - These are part of the realised artefact in use, not mere “context”. - They determine reachable continuations and therefore alter the aesthetic profile of trajectories (e.g., refusal-turn cadence, hedging templates, style normalisation). - These lights satisfy the constitutive tie and boundary fidelity tests. - Mechanism and constraint-stack describe how the function is realised; data-regime explains why the realised operator behaves as it does; semiotics explicates formal regularities in outputs that the operator in fact produces. - None requires adopting an agent stance; each yields concrete prescriptions for where and how to look. # Slide 17 — Cooperation among lights - How the lights cooperate rather than compete. - First, they co-determine boundaries. - Mechanistic and constraint-stack lights tell us what is _inside_ the artefact now (weights, current context window, decoding policy, active filters); data-regime and semiotics tell us which output regularities are artefact-internal rather than user-imported. - Together they exclude extrinsic ends (e.g., the user’s business goal) from the appreciative object. - Second, they triangulate foci. - Suppose the model produces a deft thread of legal argument. - Mechanistic light recommends inspecting temperature and top-p settings; data-regime explains why certain citations and phrasings dominate; semiotic light directs attention to register discipline, argumentative scaffolds, and genre markers; constraint-stack explains hedges and disclaimers. - The result is a single scene viewed under compatible lenses, not four scenes. # Slide 18 — Commensurable acts of aspection; conflict resolution - Third, they yield commensurable acts of aspection. - A practical trio: vary a prompt minimally (mechanistic sensitivity), watch register transitions and code-mixing (semiotics), and interpret shifts through known corpus priors (data-regime). - Each step is object-guided and produces checkable expectations under intervention—Carlson’s hallmark of “right knowledge”. - When lights seem to conflict, Carlson’s constraints adjudicate. - Constitutive levels (operator, constraint-stack) override derivative readings; scale mismatches are corrected by aligning the act of aspection to the operative timescale (token-to-token, turn-to-turn, session-level). - The aim is not reduction to one light but coherence across lights under a single kind. # Slide 19 — Making the quoted passage precise - Making the quoted passage precise. - Your passage already sets up a two-layer pairing: mechanistic sciences and a complementary semiotic level. - Read through Carlson, that pairing is exactly a case of admissible multi-light appreciation. - Mechanistic knowledge anchors kind and realisation; semiotic vocabulary does not float free but tracks the artefact’s objective regularities in output space. - The two lights are co-tethered: semiotic patterns earn their keep by being counterfactually stable under the interventions that the mechanistic light predicts (prompt perturbations, decoding changes, constraint toggles). - That stability is what makes semiotics here a way of studying _the thing itself_ rather than a gloss. # Slide 20 — Short synthesis (for signposting) - Short synthesis for the draft/presentation: _Right knowledge is plural but disciplined_. - For environments, multiple sciences are “right” because each latches onto determinants of the present order at its proper scale; for LLMs, multiple lights—mechanistic, data-regime, semiotic, constraint-stack—are “right” because each studies determinants of token-trajectory production within the artefact’s realised boundaries. - Plurality expands what can be seen without loosening objectivity: attention remains centred on what the thing in fact is, with each light yielding acts of aspection that are mutually constraining rather than merely descriptive. # Slide 21 — Semiotics deep dive (setup) - Third, the _semiotic_ light (non-mentalistic sign structure). - Tokens are treated as sign-forms; prompts as constraint-setting acts; continuations as trajectories through a learned semiosphere. - This light studies the same artefact by making the code-level regularities and transitions (register shifts, genre uptake, idiom blending) available to aspection without positing beliefs or intentions. - At a complementary descriptive level, semiotic analysis supplies a non-mentalistic vocabulary for outputs and prompts: tokens as _representamens_, prompts as semiotic acts that perturb constraints, corpora as a sample of the _semiosphere_ of cultural codes. - This reframing avoids anthropomorphism while explaining what the system in fact does—organising and recombining signs under learned statistical constraints. # Slide 22 — Semiotics deep dive (method as acts of aspection) - Given the kind and boundaries, attention should track trajectories rather than single tokens (the autoregressive evolution is the dynamic object) and sensitivity to perturbations in initial conditions (prompt edits), policy settings, and constraints. - Watch code-level regularities and transitions: register shifts, genre uptake, idiom blending, rhetorical scaffolds. - Interpret these as the artefact’s objective regularities in output space, co-tethered to the mechanistic light by counterfactual stability under interventions (prompt perturbations, decoding changes, constraint toggles). - This is an admissible light because it yields disciplined acts of aspection keyed to determinants of token-trajectory production within the realised boundaries. # Slide 23 — Semiotics deep dive (analogy to Carlson’s natural sciences) - A helpful way to see this is to treat each light as a scale- and process-sensitive map from kind to attention. - Geomorphology, ecology, and microclimatology all study _the very environment_—not a proxy—because each tracks objective structures and processes that constitute that environment at different scales. - They are ways of latching onto determinants of the present order (e.g., dune migration, trophic interactions, convective cycles). - By analogy, semiotics functions as a science-like light for artefactual outputs: it tracks code-level determiners (register priors, genre conventions, idiom repertoires) that shape trajectories. - In Carlson’s terms, this is the same naturans–naturata discipline: learned continuation as _machina naturans_; token configurations as _machina naturata_; prompts as perturbations of process akin to tidal cycles modulating a shoreline’s forms. - Acts of aspection are then guided across the two. # Slide 24 — From semiotics to an aesthetics of LLMs - On this reading, LLM behaviour becomes aesthetically appreciable as a _generative environment_: the learned rule (_machina naturans_) producing token trajectories (_machina naturata_) in response to local perturbations. - Boundary tests fix what counts as part of the setting for appreciation now (weights, context window, decoding policy, active constraints). - Semiotic light then yields foci—register discipline, genre uptake, idiom blending—and commensurable acts of aspection—track trajectories, test perturbations, read code-mixing—just as geomorphology and ecology yield foci and acts of aspection for environments. - The result is plurality without relativism: multiple lights—mechanistic, data-regime, semiotic, constraint-stack—cooperate under a single kind to discipline attention to what the artefact in fact is. # Slide 25 — Closing synthesis - We can state the thesis succinctly. - Right knowledge is plural but disciplined. - For designed artefacts the kind is fixed by function and mode of realisation. - With LLMs, the function is learned conditional continuation; the realisation is an implemented operator plus decoding policy and constraint stack. - Within that kind, mechanistic, data-regime, semiotic, and constraint-stack lights are “right” because each studies determinants of token-trajectory production within the artefact’s realised boundaries. - Plurality expands what can be seen without loosening objectivity; appreciation remains centred on what the thing in fact is, with each light yielding acts of aspection that are mutually constraining rather than merely descriptive. # New Plan on the train ** ## Through a Text box Darkly: The Metaphysics of Large Language Models ## Brainstorm - We need to talk about the difference between recording and representation - We need to talk about the difference between logical, iconic, and distributed. - We need to talk about lossiness - logical – paraphrase? - iconic - blurry - distributed? ### 1. The Metaphysics of Digital Artifacts ### 2. What sort of Digital Artifact is ChatGPT? - Main claim: ChatGPT is both a Distributed Recording and Distributed Representation of the Web - Recording and Representation - Recording is a witless process 2. ### What sort of Artifact is ChatGPT? ** # Notes from train conversation - From the text: >Consider once more a photo of an inscription: in favorable cases, it both represents and records that text; and, meanwhile, the text itself is a (different) representation of something else. - could we say the same about chatgpt? - chatgpt both represents and records (is a recording of) the web. Meanwhile, the web itself represents something else. - # My version ## Recording ## Degradation of Recording - happens for many types of non-digital recording - Not really considered by Kulvicki or Haugeland - But pretty common, especially when the recording is from one format to another - A JPEG photo of a painting taken on a very bad digital camera - A cassette recording oa CD (limited dynamic range, limited frequency response) - Sometimes, degradation is deliberate –we compress files to free up spapce ## LLMs are Degraded Recordings - In short: "all the text of the web" is converted into a different format –weights in neural networks - Analogy: Neural network learning compared to creating a map from a landscape. - Map Creation: Doesn't replicate every feature (tree, rock), but abstracts and encodes key terrain features into symbols. - Neural Network Weights: Don't copy text directly, but abstract and encode linguistic patterns from the data. - Function: These abstractions enable language understanding and generation, akin to how a map guides through terrain. ## Text to Neural Network Transformation - **Syntax and Grammar**: The model encodes rules of language structure, such as word order and sentence construction. - **Semantics**: Meanings of words and phrases are encoded based on context and usage patterns. - **Thematic Elements**: Recurrent themes or subjects are recognised and encoded. Nonetheless, the intricate layers or unique interpretations of these themes might be underrepresented. - **Word Co-occurrence and Associations**: The model learns common pairings and associations of words. While effective for understanding language patterns, it may not capture the rich, associative meanings humans perceive in literary or nuanced texts. ## Is ChatGPT an Analog Recording? ## Recording ## Chiang: ChatGPT as a blurry JPEG - ## ChatGPT is lossy compression - format changing from text to weights in a neural netwoek - ## ChatGPT and Neur ## Enrico's bit - Artifacts are created entities that perform their function in virtue of their structure.  - In hylomorphic terms (Evnine 2016, Passinsky 2021), the structure corresponds to the matter and the form to the function.  - Digital artifacts have bits as ultimate constituents of their matter.  - Evnine: there may be a hierarchy of material constituents.  - The computer is an artifact but is not a digital artifact.  - The computer performs the function of executing computer programs in virtue of its architectural structure which connects devices (Von Neumann, Tenenbaum XXX).  - Computer programs are digital artifacts.  - The structure of a computer program: instructions and data.  - The function of a computer program: turning inputs into outputs.  - The broadest category of digital data are files.  - Two kinds of files: computer programs and data (which can be inputs or outputs of computer program).  - The computer network is an artifact but is not a digital artifact.  - The computer network performs the function of executing networked programs in virtue of its architectural structure which connects computers (Tenenbaum YYY).  - The structure of a networked program: the client-server architecture (Tenenbaum YYY).  - The function of a networked program: turning client inputs into client-server outputs.  - The browser is a typical client component of a client-server architecture.  - The website is a typical server component of a client-server architecture.  - The function of the search engine: keywords as its inputs and lists of websites as its outputs.  - The structure of the search engine: a database of websites and search algorithms.  - The function of the language model: user prompts as its inputs and texts as its outputs.  - The structure of the language model: **a representation of the web and neural networks.**  ## What Sort or Representation is ChatGPT? - Ted Chiang ChatGPT Is a Blurry JPEG of the Web >What I’ve described sounds a lot like ChatGPT, or most any other large-language model. Think of ChatGPT as a **blurry JPEG of all the text on the Web**. It retains much of the information on the Web, in the same way that a JPEG retains much of the information of a higher-resolution image, but, if you’re looking for an exact sequence of bits, you won’t find it; all you will ever get is an approximation. But, because the approximation is presented in the form of grammatical text, which ChatGPT excels at creating, it’s usually acceptable. You’re still looking at a blurry JPEG, but the blurriness occurs in a way that doesn’t make the picture as a whole look less sharp. - How seriously should we take this analogy?  - We suggest, quite seriously: ChatGPT is a degraded copy of the internet - ChatGPT however, is not an image-like representation ## Recording - **Description of what recording is according to Kulvicki and Haugeland** - **Differentiate from a representation?** ## Degradation of Recording - happens for many types of non-digital recording - Not really considered by Kulvicki or Haugeland - But pretty common, especially when the recording is from one format to another - A JPEG photo of a painting taken on a very bad digital camera - A cassette recording on CD (limited dynamic range, limited frequency response) - Degraded copies can still be very useful - Sometimes, degradation is deliberate –we compress digital files to free up space ## How ChatGPT is Trained - Training involves inputting a vast array of text into the neural network. - The network employs algorithms to identify and assimilate patterns in this data, modifying its internal framework to reflect these patterns. - The model's parameters, referred to as weights, are refined throughout this process to effectively represent the **structural/syntactic/linguistic** elements of the input text. ## ChatGPT as a Degraded Copy - Degradation can occur when copying from one format to another (e.g., a JPEG photo of a painting), ChatGPT's training involves a form of 'degradation' or abstraction from the original text sources to neural network weights. - The process, while not specifically addressed by Kulvicki or Haugeland, resembles the loss of fidelity in traditional recording processes. - Analogy: Neural network learning compared to creating a map from a landscape. - Map making can be a 'witless' process: automatic arial photography - Map Creation: Doesn't replicate every feature (tree, rock), but abstracts and encodes key terrain features into symbols. (what was thatBorges short story?) - Neural Network Weights: Don't copy text directly, but abstract and encode linguistic patterns from the data. - **Syntax and Grammar**: The model encodes rules of language structure, such as word order and sentence construction.  - **Semantics**: Meanings of words and phrases are encoded based on context and usage patterns.  - **Thematic Elements**: Recurrent themes or subjects are recognised and encoded. Nonetheless, the intricate layers or unique interpretations of these themes might be underrepresented. - **Word Co-occurrence and Associations**: The model learns common pairings and associations of words. While effective for understanding language patterns, it may not capture the rich, associative meanings humans perceive in literary or nuanced texts. ** It's a change of format which is compressed along the way, and then uncompressed. ![[[[[[ChatGPT and Skeletal Content]]]]]] Photography and Encoding vs Recording ![[[[[[Taking Photos vs Training ChatGPT]]]]]] https://www.theverge.com/24027198/adobe-dana-rao-ai-copyright-fair-use-figma-acquisition-deal-decoder-interview ![[archive everything/screenshot.png]] ![[archive everything/screenshot.png]] Chat about aesthetics of ai 9 Sep 2025 is an inscrutable artifact connect to machina naturans # Chat with ChatGPT about this stuff Certainly! This conversation covered a broad range of topics centred around the ontology of digital artifacts, with a specific focus on ChatGPT and its comparison to various forms of representation through the lens of Kulvicki's structural features. Here's a detailed summary with section headings: ### Introduction to Ontology of Digital Artifacts - **Understanding Digital Artifacts**: We began by exploring the nature of digital artifacts, acknowledging their significance in contemporary technology. Digital artifacts were identified as complex entities whose understanding requires a deep dive into their structure, functionality, and representation. ### ChatGPT as a Linguistic Representation - **ChatGPT's Nature**: ChatGPT was characterized as a sophisticated AI language model that interacts, responds, and synthesizes information based on input, functioning akin to a linguistic representation or an "image of information." - **Comparison to Text Files**: ChatGPT was contrasted with a static text file, highlighting its dynamic, interactive, and contextually adaptive nature, as opposed to the fixed and unchanging nature of a text file. ### Kulvicki's Structural Features in Representational Systems - **Repleteneness, Syntax, and Semantics**: Kulvicki's concepts of repleteness, syntactic and semantic density, transparency, and sensitivity were introduced as a framework to analyze different representational systems. - **Application to Various Forms**: These structural features were applied to compare and contrast various forms of representations, including visual pictures, audio pictures, fMRI images, linguistic representations, graphs, and ChatGPT. ### Philosophical Implications and Aristotle's Hylomorphism - **Philosophical Discussion**: The conversation delved into philosophical aspects, discussing Aristotle's hylomorphic view of artifacts, which considers every entity as a composite of matter (hyle) and form (morphe). - **Relation to Digital Artifacts**: The hylomorphic framework was connected to the ontology of digital artifacts, emphasizing the role of 'matter' as digital data and 'form' as the specific structure and organization that give digital artifacts their identity and functionality. ### ChatGPT as an Image of Information - **Unique Perspective**: ChatGPT was conceptualized as an "image of information," a perspective that aligns with its abstract, synthesized, and dynamic representation of knowledge and information. - **Advantages Over Text Files**: This conceptualization was argued to be more fitting for ChatGPT than for static text files due to its interactive, context-sensitive, and interpretative nature, akin to how an image represents complex ideas or scenes. ### Systematic Taxonomy of Digital Artifacts - **Categorization of Digital Artifacts**: A taxonomy was proposed to categorize digital artifacts into various types, including data storage, software, interactive systems, network and communication forms, embedded systems, data structures, and specialized forms like blockchain and virtual reality. ### Detailed Analysis of Representational Systems - **Comparative Analysis**: Detailed comparisons were drawn between different representational systems using Kulvicki's framework, highlighting the unique features of visual pictures, audio pictures, fMRI images, linguistic representations, graphs, and ChatGPT. - **Insights from the Analysis**: The analysis provided insights into the complexity, interactivity, and representational style of ChatGPT, likening it most closely to linguistic representations due to shared characteristics like abstraction, semantic richness, and low transparency. ### Conclusion - **Key Takeaways**: The conversation underscored the multifaceted nature of digital artifacts and the relevance of philosophical frameworks like Aristotle's hylomorphism and Kulvicki's structural features in understanding AI models like ChatGPT. - **Future Implications**: The discussion highlighted the importance of considering the ontology of digital artifacts in understanding modern technology and its evolution, suggesting that this perspective can offer profound insights into the nature, functionality, and implications of AI and other digital systems. # Table of Kulvicki's ideas applied to different types of image | Feature | Description | Visual Pictures | Audio Pictures | ChatGPT's Model Files (as a single representation) | fMRI Images | Linguistic Representations | Graphs | |-----------------------|-----------------------------------------------------------------------------|---------------------------------------|------------------------------------------|-------------------------------------------------------------|--------------------------------------------------|---------------------------------------------|------------------------------------------------| | **Repleteness** | How many features of a representation are crucial to its identity. | High (color, shape, etc. are crucial) | High (pitch, tone, volume, duration) | High (model parameters crucial for varied output) | High (location, intensity of signals are crucial)| Moderate (depends on language and context) | Low (specific data points, not rich detail) | | **Syntactic Density** | Degree to which small changes in representation's features change its identity. | High (small changes affect identity) | High (small changes in sound waves) | Moderate (parameter adjustments affect output) | Moderate (variations in imaging affect interpretation) | Low (structured by grammar and syntax) | Low (structured by data points and axes) | | **Semantic Density** | Range of meanings or concepts a representation can express. | High (wide range of possible scenes) | High (varied emotions/themes) | High (generates diverse and contextually varied text) | Moderate (represents specific physiological states) | High (capable of expressing a wide range of meanings) | Low (represents specific data or trends) | | **Transparency** | How directly a representation depicts its subject. | High (directly depicts its subject) | High (recording resembles the sound) | Low (abstract representation, not depicting specific data) | Low (represents physiological activity, not direct images) | Low (symbolic, not directly depicting) | Low (represents data abstractly, not directly) | | **Syntactic Sensitivity** | Sensitivity of a representation to changes in syntactic features. | High (sensitive to minor variations) | High (sensitive to sound variations) | High (sensitive to input nuances) | High (sensitive to physiological changes) | Moderate (sensitive to word choice and order) | Low (changes may not significantly alter representation) | | **Semantic Sensitivity** | Sensitivity of a representation to changes in meaning based on context or use. | High (meanings change with content) | High (meanings change with composition) | High (meaning of output varies with context) | Moderate (interpretation dependent on medical context) | High (context and usage greatly affect meaning) | Low (specific interpretations of data) | # ChatGPT and Kulvicki on representation 1. **Repleteness**: - The model's repleteness is substantial but not complete. It embodies a vast array of linguistic structures, patterns, and information derived from its training data, but it does not encapsulate the entirety of the human language or the full scope of the data it was trained on. The model's representation is rich in details regarding language use, syntax, and semantics but is inherently limited by the scope and nature of its training data. 2. **Syntactic Density**: - ChatGPT's model exhibits high syntactic density in its structure. The model's architecture, particularly the patterns and structures it has learned, allows for nuanced understanding and generation of language. Small variations in the syntactic structure of the language it has learned can lead to different representations within the model itself. However, as a representation, it doesn't capture the full syntactic variability of natural language. 3. **Semantic Density**: - The model also demonstrates high semantic density. It contains a diverse array of meanings and contexts derived from its training data. The model's ability to relate different words, phrases, and structures to a vast range of meanings is a testament to its semantic richness. Yet, it's an imperfect representation because it can't fully encompass the depth and breadth of meanings present in natural language and human communication. 4. **Transparency**: - Transparency in the context of ChatGPT's model is moderate. While the model's responses and behavior can provide insights into its training and structure, the internal workings and decision-making processes are not entirely transparent, even to experts. The model's representation of its training data is abstract and encoded in complex, multi-dimensional matrices within neural networks, making direct interpretation challenging. 5. **Syntactic Sensitivity**: - The model possesses syntactic sensitivity, meaning it is responsive to the syntactic structure of the language data it was trained on. However, this sensitivity is within the confines of its training and architecture, making it an imperfect reflection of the full syntactic diversity of natural language. 6. **Semantic Sensitivity**: - Semantic sensitivity is also a characteristic of the model. It is attuned to the meanings and contexts present in its training data and can reflect these in its internal representations. Nevertheless, the model's semantic sensitivity is bounded by the limitations of its training, making it an incomplete representation of the vast and nuanced semantic landscape of natural language. # Relevant Notes ## Hylomorphism ![[[[Hylomorphism]]]] ## Hylomorphism and Digital Artifacts ![[[[Hylomorphism]] and Digital artifacts]] # Relevant Papers ## Evnine Paper ![[evnineMakingObjectsEvents]] # Nice Quote from Kulvicki on Mental Images >"perhaps mental images are merely isomorphic or homomorphic representations. They do not share qualities with their objects, but they share a structure with them". (Images ) # From Bits to Jpegs, based on Kulvicki's structural theory of images ## One version of chatGPTs about the journey from bit to jpeg Certainly, I will provide a detailed journey from bits to JPEGs, integrating Kulvicki's image features throughout the progression to elucidate how image qualities emerge at each stage: ### Bits and Bytes - **Foundation**: Bits, as the most basic units of digital data, represent binary values (0 or 1). They form the foundation of all digital artifacts, including JPEGs. At this stage, Kulvicki’s concepts such as repleteness and syntactic density are not yet applicable, as bits in isolation do not convey complex information or representational qualities. ### Pixel Representation - **Emergence of Repleteness and Syntactic Density**: As we group bits into bytes to represent pixels, the initial emergence of image features occurs. Each pixel’s color information, encoded in bytes, begins to form the basic visual elements of an image. This stage marks the onset of repleteness, where each pixel’s color is integral to the overall image. Additionally, syntactic density starts to become relevant, as changes in pixel color directly affect the image's appearance. ### JPEG File Structure - **Developing Syntactic and Semantic Density**: JPEG images, through their compression algorithms, display a marked increase in syntactic and semantic density. The compression process, which includes the Discrete Cosine Transform (DCT) and quantization, can subtly alter visual details, affecting both the structure and the meaning of the image. This demonstrates Kulvicki’s concept of syntactic density, where minute alterations in the image's data can lead to significant changes in its identity. Simultaneously, semantic density is evidenced as these alterations can modify the perceived meaning or quality of the image. - **Transparency and Sensitivity**: JPEGs strive to maintain transparency by retaining essential visual information, although the lossy nature of JPEG compression means some transparency is sacrificed. This process also introduces increased syntactic sensitivity, as alterations in compression parameters can significantly change the image's appearance. Semantic sensitivity becomes apparent, as the manner of compression influences the viewer's perception and interpretation of the image. ### File Storage and Compression Algorithms - **Repleteness in Stored JPEGs**: A stored JPEG file represents a system that captures a wide array of visual information, aligning with Kulvicki’s concept of repleteness. The image includes a broad range of colors and details, although this is balanced against the need for smaller file sizes. The compression algorithms play a crucial role here, determining how effectively the JPEG format can balance image quality with file size. - **Ongoing Sensitivity**: In the stored JPEG, syntactic and semantic sensitivities are maintained. Any further editing or compression can lead to changes in the image's structure (syntax) and the perceived meaning or quality (semantics). This reflects the dynamic nature of digital images, where even post-creation modifications can have significant impacts. ### Conclusion Throughout this journey from bits to JPEGs, we observe the gradual incorporation of Kulvicki’s image features. Starting from the basic representation of color and light information in bits and bytes, progressing through the organization of this data into pixels, and culminating in the complex, compressed JPEG file, each stage introduces and develops key aspects of repleteness, syntactic and semantic density, transparency, and sensitivity. This progression not only illustrates the layered complexity inherent in digital image creation and storage but also demonstrates how digital processes imbue raw data with representational qualities characteristic of images in Kulvicki’s framework.