## Semiotic Physics and LLMs — Structured Working Map ## I. Orientation: what the idea is supposed to do ### 1\. The basic problem - We need a way of describing LLMs that avoids two opposed mistakes. - First mistake: treating the model as if it were a mind, author, thinker, speaker, or unified agent. - Second mistake: treating the model as if it merely retrieves, copies, or randomly assembles fragments of language. - *Semiotic physics* is useful because it offers a middle description. - LLM outputs are not intentional acts in the ordinary human sense. - But they are also not random heaps of linguistic material. - They are trajectories through a learned space of signs. - The core question is therefore not simply: - “Does the model understand?” - “Does the model believe what it says?” - “Did the model retrieve this from somewhere?” - The better question is: - “What regularities govern the production of this sign-sequence from this prompt-state?” ### 2\. The shortest working definition - *Semiotic physics* is the system of regularities governing the propagation of signs in an LLM. - It concerns the ways in which prompts, genres, concepts, styles, argumentative forms, and conversational roles make some continuations more probable than others. - It treats LLM generation as movement through a structured space of signs rather than as retrieval, random recombination, or intentional authorship. ### 3\. The slower working definition - *Semiotic physics* names the order that governs how signs move within an LLM-generated trajectory. - The relevant signs include: - tokens; - words; - phrases; - idioms; - concepts; - styles; - registers; - genres; - voices; - argumentative forms; - conversational roles; - institutional scripts; - narrative expectations; - inherited cultural patterns. - These signs do not combine freely. - A prompt establishes a local configuration. - That configuration constrains what can intelligibly follow. - Some continuations are strongly favoured. - Some are weakly available. - Some are practically excluded unless the prompt is changed. - The model’s output is therefore a trajectory through a learned space of semiotic possibilities. ### 4\. The paper-ready formulation - By *semiotic physics*, we mean the system of regularities governing the propagation of signs in an LLM: the ways in which prompts, genres, concepts, styles, argumentative forms, and conversational roles make some continuations more probable than others. The term is metaphorical, but not idle. It captures the fact that LLMs do not merely retrieve stored sentences or randomly assemble linguistic material. They generate trajectories through a structured space of signs, where prior textual patterns exert probabilistic pressure on what can intelligibly come next. --- ## II. Why the simulator frame comes first ### 1\. Why ordinary categories mislead us - The central reason to begin with the simulator frame is that it explains why familiar categories applied to LLMs are systematically distorting. - If we treat an LLM as an *agent*, we tend to ask: - What does it want? - What goal is it pursuing? - What belief does it express when it says this? - What intention explains this output? - If we treat it as an *oracle*, we tend to ask: - Did it answer the question correctly? - Does it know the relevant fact? - How reliable is it as a source of information? - How can we extract the true answer from it? - If we treat it as a *tool*, we tend to ask: - What task was it designed to perform? - How efficiently does it perform that task? - What function does the user control? - How well does it execute the command? - If we treat it as a *stochastic parrot*, we tend to ask: - What material is it repeating? - What patterns has it copied? - Where is the training-data echo? - Why should we take the output as anything more than recombination? - The simulator frame asks a different family of questions: - What distribution has the model learned to continue? - What trajectory is generated when a prompt gives the system an initial state? - What kinds of text-processes can be induced by different prompts? - What regularities govern the movement from one token-state to the next? - What kinds of simulacra are generated under these local conditions? - What kinds of roles, voices, genres, and arguments become available from this prompt-state? ### 2\. The simulator/simulacrum distinction - The model itself is not identical with the persona, argument, story, speaker, or agent-like process that appears in the output. - The relevant distinction is: - *simulator*: the model as a transition rule over possible continuations; - *simulacrum*: the local text-process generated under particular prompting and sampling conditions. - This distinction prevents a common type error. - We should not infer from a particular output that the model, as such, has that output’s beliefs, aims, voice, or standpoint. - The output may contain an agent-like voice, but that voice is a generated phenomenon, not the model’s enduring self. - The same model can generate mutually incompatible speakers, arguments, ideologies, tones, and roles. - So the stable object is not the simulated speaker. - The stable object is the generative rule under which such speakers can be locally produced. ### 3\. Why the physics analogy enters here - The analogy with physics is attractive because physics is not itself the person who runs, argues, falls, lies, learns, or forgets. - Physics is the rule under which such things can occur. - Similarly: - the LLM is not the simulated philosopher; - not the narrator; - not the assistant-persona; - not the bureaucrat; - not the child; - not the propagandist; - not the legal adviser; - not the hostile interlocutor; - not the sympathetic collaborator. - The LLM is the system under which such text-processes can be generated. - The simulator frame therefore gives us the basic ontology: - model as rule; - prompt as local state; - output as trajectory; - persona or argument as simulacrum. --- ## III. From next-token prediction to semiotic physics ### 1\. Stage one: from next-token prediction to trajectory generation - A thin description says: - an LLM predicts the next token. - This is correct but incomplete. - The thicker description says: - a token is sampled from the model’s probability distribution; - that token is appended to the context; - the expanded context becomes the next input; - the model again produces a distribution over possible next tokens; - another token is sampled; - the process repeats; - the output is therefore a trajectory, not a single prediction. - The philosophical point: - the model is not merely answering discrete questions; - it is evolving a sign-state forward. ### 2\. Stage two: from trajectory generation to simulator - A simulator is a system that evolves a configuration according to a learned transition rule. - In the LLM case: - the configuration is a textual context; - the transition rule is the model’s learned probability distribution over continuations; - the generated output is the trajectory produced by repeated applications of that rule. - This explains why prompting has such a large effect. - A prompt is not merely a request. - It specifies the initial state of the simulation. - Different initial states produce different trajectories. - A change in wording, role, genre, example, or instruction can alter the local continuation space. ### 3\. Stage three: from simulator to semiotic simulator - The simulator’s domain is linguistic and cultural rather than physical. - It does not propagate particles, forces, or bodies. - It propagates signs. - But signs are not only tokens. - Signs enter higher-order systems: - linguistic systems; - literary systems; - academic systems; - argumentative systems; - bureaucratic systems; - ideological systems; - conversational systems; - institutional systems; - pedagogical systems; - narrative systems. - The model is trained on traces of these systems. - It therefore learns not only local word association, but also patterns of discourse. ### 4\. Stage four: from semiotic simulator to semiotic physics - Semiotic physics is the study of the regularities governing those sign-trajectories. - It asks: - What kinds of prompts produce stable trajectories? - What kinds of prompts produce unstable trajectories? - What textual forms function as attractors? - What roles or genres dominate once activated? - What concepts tend to bring neighbouring concepts with them? - What argumentative moves tend to elicit standard objections or replies? - What kinds of framing make the model more likely to produce explanation, narrative, justification, list, confession, refusal, synthesis, or speculation? - What kinds of starting conditions make drift, cliché, hallucination, or generic academic prose more likely? - What kinds of prompting keep the model inside a precise argumentative space for longer? --- ## IV. The semiotic layer: what is being propagated ### 1\. Why the simulator frame needs a semiotic supplement - The simulator frame alone describes a general form of generative modelling. - *Semiotic physics* adds that the relevant domain is not molecules, planets, fluids, or biological organisms, but signs. - The elementary units are tokens, but tokens function within larger semiotic structures. - These include: - words; - idioms; - styles; - registers; - genres; - concepts; - argumentative moves; - conversational roles; - institutional voices; - narrative expectations; - inherited cultural scripts. - So the relevant “physics” is not merely the probability of token B after token A. - It also includes higher-level regularities: - what follows from adopting the voice of a referee report; - what follows from beginning a philosophical objection; - what follows from invoking a tradition, such as analytic aesthetics, Kantian aesthetics, semiotics, or AI alignment; - what follows from introducing a contrast between cognition and semiosis; - what follows from presenting an example as a counterexample rather than as an illustration; - what follows from writing under the expectation of “academic prose”; - what follows from asking for neutrality, polemic, sympathy, suspicion, compression, or explanation. ### 2\. Tokens as elementary particles - The lowest-level units are tokens. - Tokens are not always words. - They may be: - word fragments; - punctuation marks; - spaces; - subword units; - formatting markers. - At the technical level, the model assigns probabilities to possible next tokens. - At the semiotic level, those tokens participate in larger units of meaning. - The particle analogy is useful only if we do not reduce signs to particles. - Tokens are the mechanically manipulated units. - Signs are the interpreted units. - Semiotic physics spans the relation between both levels. ### 3\. Prompts as initial conditions - A prompt does not merely ask for content. - It fixes a local semiotic configuration. - It can specify: - topic; - style; - genre; - speaker role; - audience; - level of formality; - argumentative stance; - epistemic posture; - permitted vocabulary; - excluded moves; - expected structure; - relation to previous text; - standards of success; - degree of speculation; - degree of compression. - Examples: - “Explain X to a child” activates pedagogical simplification. - “Write as a journal referee” activates evaluative academic conventions. - “Argue in the style of analytic philosophy” activates distinctions, objections, and thesis-testing. - “Write a manifesto” activates compression, urgency, collective address, and polemic. - “Summarise neutrally” activates restraint, flattening, and source-dependence. - “Give a balanced account” activates a familiar two-sided dialectical structure. - “Write as a grant proposal” activates problem, gap, objectives, methods, outcomes, and impact. ### 4\. Genres as force fields - Genres exert pressure on continuations. - Once a genre is activated, the model tends to produce genre-appropriate moves. - Examples: - A referee report tends toward summary, strengths, weaknesses, recommendation. - A grant proposal tends toward problem, gap, objectives, methodology, impact. - A philosophical paper tends toward thesis, distinction, objection, reply. - A detective story tends toward clues, suspicion, revelation, misdirection. - A TED talk tends toward accessible narrative, affective hooks, personal relevance, simplified stakes. - A legal memorandum tends toward issue, rule, application, conclusion. - A manifesto tends toward collective address, urgency, opposition, and compressed principle. - This is not because the model has an intention to respect genre. - It is because the training distribution contains stable correlations between genre markers and subsequent forms. - Genre therefore functions like a semiotic field: - it does not determine every continuation; - but it biases the path of continuation; - it makes some moves feel natural and others less available. ### 5\. Concepts as attractors - Some concepts bring predictable neighbouring concepts into view. - “Consciousness” attracts: - experience; - intentionality; - subjectivity; - access; - reportability; - Chalmers; - hard problem; - illusionism; - functionalism. - “Depiction” attracts: - representation; - resemblance; - seeing-in; - aspect perception; - pictorial content; - Lopes; - Wollheim; - Peacocke. - “LLMs” attracts: - prediction; - tokens; - training data; - hallucination; - agency; - alignment; - understanding; - stochastic parrots. - “Aesthetic appreciation of nature” attracts: - Carlson; - scientific knowledge; - natural environmental model; - picturesque; - sublime; - object model; - landscape model; - ecology; - geology; - objectivity. - The point is not that these associations are always good. - The point is that concepts structure the local continuation space. ### 6\. Conversational roles as local personae - LLMs often generate text from within an implied role. - Roles include: - assistant; - critic; - teacher; - editor; - philosopher; - bureaucrat; - scientist; - novelist; - therapist; - lawyer; - hostile interlocutor; - sympathetic collaborator; - journal referee; - grant evaluator; - populariser; - technical explainer. - A role constrains: - tone; - permissible claims; - conversational obligations; - level of deference; - willingness to speculate; - tendency to ask questions; - tendency to hedge; - kinds of examples used; - kinds of objections foregrounded; - degree of confidence; - degree of simplification. - The role is a simulacrum, not the simulator itself. - This means: - the model can simulate the role of a philosopher without being a philosopher; - it can simulate a hostile critic without being hostile; - it can simulate an expert explanation without possessing expertise in the human sense; - it can simulate a first-person voice without having the relevant first-person state. --- ## V. The dynamical vocabulary ### 1\. Why the physics metaphor is not merely decorative - The physics metaphor is doing real work if it licenses the following thought: - generated text is not a heap of independent signs; - it is a trajectory through a space structured by regularities; - prompts set initial conditions; - sampling introduces stochastic branching; - local patterns can stabilise into attractors; - some continuations are difficult to reach from a given state; - some trajectories become self-reinforcing; - some states are unstable, so small prompt changes produce large downstream divergence; - some trajectories fall into loops or absorbing states; - some prompt-configurations sustain coherent continuation longer than others. - The metaphor becomes misleading if it suggests: - that LLMs literally implement the physics of the real world; - that signs behave with the same necessity as physical particles; - that the relevant laws are simple, exceptionless, and mathematically tractable in the way classical physical laws often are; - that the semantic world of a model converges on real-world ontology; - that cultural regularities can be treated as if they were natural laws without remainder. - The stronger and safer claim is: - semiotic physics is a disciplined analogy for describing the regularities by which sign-trajectories are generated in LLMs. ### 2\. Sampling as branching - At each step, more than one continuation may be possible. - Sampling selects one path. - Different samples from the same prompt can diverge. - This gives the generated sequence a branching structure: - one branch becomes the visible answer; - many possible branches remain unrealised. - This supports the “multiverse” language in the semiotic-physics materials. - The safer formulation: - every generated output is one path through a structured space of possible sign-trajectories. ### 3\. Autoregression as path-dependence - Once a token is generated, it changes the future state. - This makes LLM generation path-dependent. - An early framing choice can strongly determine later possibilities. - Examples: - If the answer begins as a list, continuation tends to preserve list form. - If the answer begins with a confident claim, later text tends to rationalise or support it. - If the answer begins in a legal register, later text tends to preserve legalistic structure. - If the answer begins with an analogy, later text may continue to elaborate that analogy even when it becomes strained. - If the answer begins with “both sides have merit,” later text may fall into generic balance. - If the answer begins by treating a concept as technical, later text may preserve that technicality even when the original prompt did not require it. - This matters for philosophical writing: - early formulations can trap the model in a bad argumentative trajectory; - a poor first distinction can generate downstream pseudo-clarity; - a strong initial argumentative frame can produce better inferential order; - a false contrast introduced early may be preserved and elaborated as if it were legitimate. ### 4\. Attractors - An attractor is a region of trajectory-space into which nearby trajectories tend to fall. - In LLMs, attractors may include: - stock disclaimers; - formulaic safety refusals; - generic academic transitions; - “on the one hand / on the other hand” structures; - five-part essay formats; - conventional summaries of canonical positions; - “it depends” answers; - standard objection-and-reply templates; - inflated “framework” language; - vague appeals to “complexity,” “nuance,” and “interplay.” - Attractors explain why outputs often drift toward recognisable forms even when the prompt asks for something more specific. - A useful philosophical use of the concept: - bad LLM prose is often prose captured by generic attractors; - good prompting tries to avoid low-information attractors and produce a more discriminating trajectory. ### 5\. Absorbing states and loops - Some trajectories become hard to escape. - Examples: - repeated phrases; - rigid list structures; - excessive caveating; - recursive meta-commentary; - generic “nuanced” balance; - formulaic academic prose; - repeated contrast between “not merely X but Y”; - endless announcement of what the answer will do rather than doing it. - In philosophical drafting, absorbing states appear as: - over-repetition of the same distinction; - pseudo-technical vocabulary that propagates without argumentative need; - paragraphs that keep announcing what will be done rather than doing it; - “This raises important questions…” type transitions; - “complex and multifaceted” placeholders; - empty meta-argumentative scaffolding. - Semiotic physics gives a useful diagnosis: - these are not merely stylistic failures; - they are local trajectory failures; - the model has entered a region where generic continuation is easier than argumentative advance. ### 6\. Chaotic regions - Some prompt states are unstable. - Small differences in wording produce large differences in continuation. - Examples: - highly underspecified creative prompts; - prompts containing contradictory role instructions; - prompts asking for difficult philosophical synthesis without specifying argumentative constraints; - prompts combining multiple genres without hierarchy; - prompts that ask for both neutrality and polemic; - prompts that ask for both a faithful reconstruction and radical rewriting; - prompts that ask for “deep analysis” without specifying what question the analysis should answer. - Chaotic regions are not always bad. - They may be useful for brainstorming. - They are dangerous for argument construction when the output needs inferential control. ### 7\. Lyapunov-style questions - The mathematical vocabulary of Lyapunov exponents and Lyapunov time can be used analogically to ask: - How quickly does the output lose sensitivity to the original prompt? - How long does a role remain stable? - How long does a stipulated distinction continue to control the answer? - How far can the model continue a philosophical argument before drifting into adjacent discourse? - How many turns can a chat sustain the same conceptual frame? - How quickly does a model revert to generic academic discourse after being given a precise prose constraint? - This could become a useful experimental methodology: - give the model a precise philosophical setup; - measure where it first violates the setup; - compare across prompt types, models, and drafting conditions; - test which constraints remain active and which decay. ### 8\. The semiotic coin-flip example - The semiotic-physics materials use a simple example: ask the model to produce a sequence of 0s and 1s. - The point is that even an apparently minimal symbolic task differs from a fair coin. - Two differences matter: - the semiotic coin is not fair; - the flips are not independent. - Once the model produces the same token several times in a row, it may lock onto the pattern. - This illustrates a general feature: - signs generated by the model are not independent samples; - the trajectory acquires structure as it unfolds; - earlier signs alter the probability of later signs. - The example is simple, but the general lesson applies to richer cases: - repeated rhetorical structures reinforce themselves; - genres become more stable once enacted; - early conceptual framings constrain later argumentative moves; - local patterns can become self-confirming. --- ## VI. What semiotic physics is not ### 1\. Not a claim that LLMs understand in the human sense - Semiotic physics does not require attributing experience, intention, or understanding to the model. - It explains structured output without treating the model as a thinker. - This is one of its main advantages. - The model may generate text that looks like explanation, reasoning, or interpretation. - The semiotic-physics claim is not that the model consciously performs those acts. - The claim is that it propagates sign-structures associated with those acts. ### 2\. Not a claim that LLMs are merely stochastic parrots - The model does not simply paste together previously seen strings. - It can generalise patterns across contexts. - It can generate counterfactual configurations. - It can combine styles, roles, domains, and argumentative forms that may not have appeared together in the training data. - The semiotic-physics frame captures this middle position: - more structure than random recombination; - less agency than intentional authorship. ### 3\. Not a claim that output meaning is wholly inside the model - Meaning is not simply stored inside the LLM. - Meaning arises through interaction among: - generated signs; - prompt context; - user interpretation; - cultural conventions; - background knowledge; - practical uptake; - downstream use. - Semiotic physics describes generative regularities. - It does not replace hermeneutics. - A generated text can be semantically rich because of how it is interpreted and used, even if the model does not possess semantic understanding in the human sense. ### 4\. Not a complete theory of cognition - Semiotic physics describes sign propagation. - It does not by itself settle whether LLMs reason, understand, infer, or represent. - It may explain some reasoning-like outputs without attributing reasoning as a mental act. - That distinction is useful for philosophical caution. - It allows us to say: - the output has argumentative form; - the text can be assessed as an argument; - but the model need not have entertained the argument as an argument. ### 5\. Not a purely technical theory of transformers - Semiotic physics is not the same as mechanistic interpretability. - It is not mainly about: - circuits; - attention heads; - layers; - activation vectors; - internal representations. - It is an output-side and trajectory-side theory. - It can be informed by model internals, but its primary objects are generated sign-processes. - This matters because many philosophically salient features of LLMs are visible at the level of generated trajectories: - drift; - role-stability; - generic attractors; - genre competence; - argument-like continuation; - semiotic hybridisation. ### 6\. Not literal physics - The phrase *physics* should be handled carefully. - It should not suggest: - strict necessity; - exceptionless law; - full mathematical tractability; - identity with real-world physical law; - convergence on real-world ontology. - The safer claim: - semiotic physics is an analogy with physics at the level of dynamical regularity. - The analogy is helpful because it foregrounds: - state; - transition; - trajectory; - instability; - attraction; - branching; - path-dependence; - local constraint. --- ## VII. Relation to semiotics ### 1\. General semiotic point - LLMs are better understood as systems of sign manipulation than as artificial minds. - They recombine, recontextualise, and circulate linguistic forms. - Their outputs enter human practices of interpretation. - A generated sentence can function as a sign even if the model does not understand it in the human sense. - The user’s interpretive labour is therefore not external noise. - It is part of how generated text acquires significance. ### 2\. Peircean angle - LLM outputs can be treated as signs because they produce interpretants in users. - The model need not itself grasp the object in a human way. - The generated text can function semiotically insofar as readers interpret it. - The prompt-model-user relation can be treated as a modified semiotic triad: - prompt: initial semiotic constraint; - model: generator of sign-material; - output: representamen-like artefact entering interpretation; - user: interpreter and evaluator. - This helps avoid a false dilemma: - either the model understands its signs; - or the output is meaningless. - A third option is available: - the output functions as sign-material within human interpretive practices. ### 3\. Eco angle - LLM outputs often function like open works. - They do not carry a single fixed meaning simply waiting to be decoded. - They invite: - continuation; - revision; - comparison; - contextual placement; - reinterpretation; - correction; - re-prompting. - Prompting can be understood as establishing a frame of interpretive cooperation. - The model does not just answer. - It generates material to be actualised by use. - The user functions both as: - reader of the output; - writer of the prompt; - curator of the trajectory; - evaluator of the result. ### 4\. Lotman angle - LLMs are trained on fragments of the semiosphere. - They operate within a space of cultural codes, genres, registers, and discourses. - A prompt selects a local path through that space. - The model then recombines semiotic material from different zones: - scientific discourse; - literary discourse; - bureaucratic discourse; - internet discourse; - philosophical discourse; - pedagogical discourse; - political discourse; - legal discourse; - popular-science discourse. - This is why LLMs are especially good at hybridisation: - explain Spinoza as a TED talk; - turn Dante into digital slang; - write policy from competing ideological frames; - render a philosophical argument as a dialogue, referee report, syllabus, or abstract; - explain entropy through fairy tales; - translate a scientific claim into a manifesto, sermon, legal brief, or classroom lesson. ### 5\. Where semiotic physics differs from semiotics in general - Semiotics asks how signs mean. - Semiotic physics asks how signs propagate under a trained generative rule. - Semiotics studies: - interpretation; - codes; - sign relations; - cultural systems; - meaning-making. - Semiotic physics studies: - transition; - trajectory; - constraint; - attractor; - drift; - branching; - path-dependence; - prompt-sensitivity. - The strongest view combines them. - Semiotics explains why generated material is interpretable. - Semiotic physics explains why this material, rather than another, is likely to appear under these conditions. --- ## VIII. Relation to Carlson and object-appropriate appreciation ### 1\. Carlson’s relevant principle - Carlson’s natural environmental model is useful because it gives a precedent for object-appropriate appreciation. - His claim is not just that nature can be appreciated. - It is that appropriate appreciation must be guided by what the object is. - In nature appreciation, this means not treating nature as a painting or sculpture. - Nature should be appreciated as nature and as environment, with relevant knowledge of its real structure. ### 2\. Why the art-derived models fail in Carlson’s case - Art-derived models of nature appreciation fail because they impose the wrong categories. - The object model treats nature as sculpture. - The landscape model treats nature as painting. - Both distort nature by forcing it into art categories. - They may pick out real aspects of the experience, but they do not give the right model of the object. - The lesson: - an appreciative practice can be distorted when it imports a model from the wrong kind of object. ### 3\. Application to LLMs - This is directly useful for LLMs. - We should not appreciate LLM outputs only as: - human-authored essays; - evidence of a mind; - database entries; - failed search results; - accidental verbal noise; - ordinary artworks; - ordinary philosophical papers; - mere tools executing commands. - We should ask what kind of object they are, and what knowledge is relevant to their appreciation. - On this model, semiotic physics becomes for LLMs what geology, ecology, and biology are for nature appreciation: - a background account of the object’s real generative order; - not a replacement for experience of the output; - not a demand that every user become a machine-learning researcher; - but a relevant framework for serious appreciation. ### 4\. The analogy with natural environmental aesthetics - Carlson rejects artistic models of nature because they distort the object. - A parallel argument: - human-author models of LLM text can distort the object; - oracle models can distort the object; - tool models can distort the object; - stochastic-parrot models can distort the object; - mind-based models can distort the object. - The semiotic-physics model functions as an object-appropriate model. - It lets us appreciate LLMs in light of the order they actually display. ### 5\. What this gives a paper on LLM appreciation - It supports the claim that LLM outputs can be aesthetically or intellectually appreciable without being authored in the ordinary sense. - It explains what the object of appreciation is: - not merely the surface text; - not the model as hidden mechanism alone; - not the imagined speaker; - but the relation between surface text and generative semiotic order. - It allows a layered account: - appreciation of the output; - appreciation of the chat trajectory; - appreciation of the model’s semiotic dispositions; - appreciation of the human-model interaction as a structured environment. --- ## IX. Relation to Picca’s semiotic-machine frame ### 1\. Why Picca is useful - Picca’s paper is useful because it rejects the mind-frame and treats LLMs as systems of sign manipulation. - It emphasises: - recombination; - recontextualisation; - cultural circulation; - prompting as a semiotic act; - interpretation as co-constructed between user, model, and context; - LLMs as participants in an ecology of signs; - meaning as emerging through situated interpretation rather than residing inside the model. ### 2\. Why Picca’s frame is not identical with semiotic physics - Picca’s frame is mainly hermeneutic and cultural. - It focuses on: - how LLM outputs invite interpretation; - how prompts establish frames; - how genres and ideological positions are recombined; - how outputs function as semiotic artefacts in cultural practice; - how LLMs participate in meaning-making without being minds. - The semiotic-physics frame is more dynamical and generative. - It focuses on: - how token trajectories evolve; - how prompt states constrain successor states; - how attractors, instability, loops, and divergence arise; - how higher-level forms exert probabilistic pressure over continuations; - how generated text unfolds as a path through a structured space. ### 3\. How to combine both - A strong paper could use both frameworks. - Picca supplies the semiotic and interpretive account. - Janus, metasemi, and Jan supply the simulator and dynamical account. - Carlson supplies the aesthetics-of-appreciation model. - The combined structure would be: - LLMs are not minds but semiotic machines; - more specifically, they are semiotic simulators; - their outputs unfold as trajectories under learned regularities; - these regularities can be called semiotic physics; - appreciation of LLM outputs should track this generative order. --- ## X. Relation to philosophy and argument generation ### 1\. The central philosophical question - Can LLMs generate philosophical texts worth reading if they do not understand, believe, or intend what they say? - Semiotic physics gives one route to a positive answer. - The answer is not: - yes, because the model thinks; - yes, because the model is an author; - yes, because the model has philosophical insight; - yes, because the model performs reasoning as a conscious epistemic act. - The answer is: - yes, because philosophical corpora contain public argumentative regularities; - LLMs can learn and propagate those regularities; - generated outputs can instantiate argumentative structures assessable by readers. ### 2\. Philosophical forms as semiotic attractors - Philosophical writing has recurrent forms: - distinction; - counterexample; - dilemma; - burden shift; - parity argument; - reductio; - inference to the best explanation; - reflective equilibrium; - objection and reply; - scope restriction; - debunking explanation; - companion-in-guilt argument; - transcendental argument; - constitutive claim; - necessary/sufficient condition analysis; - error theory; - regress argument; - explanatory demand; - disambiguation of senses. - These are not merely strings of words. - They are public inferential patterns sedimented in texts. - If a model is trained on enough philosophical writing, it can learn dispositions to produce these forms. - The result may be philosophically assessable even if no thinker has consciously executed the inference. ### 3\. The argument-level object - The minimal object of philosophical assessment should not be a bare sentence. - Nor should it be the model’s alleged mental state. - It should be the argument made available by the generated text. - Semiotic physics explains how such arguments can arise: - not by inner deliberation; - not by random assembly; - but by propagation through learned argumentative regularities. - The key distinction: - production history explains how the text came to exist; - argumentative assessment concerns what structure the text makes available. ### 4\. Why this matters for abduction - An objection says: - LLMs cannot perform abduction because they do not identify a phenomenon as calling for explanation, generate hypotheses as hypotheses, or select the best explanation as best. - Semiotic physics allows a more nuanced reply. - The model need not perform abduction as a conscious epistemic act. - Nevertheless: - abductive structures are present in philosophical and scientific prose; - the model can generate text that instantiates those structures; - readers can then assess whether the generated structure gives a good explanation. - The distinction: - *performing abduction as a thinker*; - *generating an abductive textual structure*. - Semiotic physics concerns the second. ### 5\. Why this matters for phenomenology - A related objection says: - LLMs lack the first-person experience needed to write philosophically about phenomenology. - Semiotic physics again gives a limited reply. - The model does not have the relevant experience. - But phenomenological writing has public textual forms: - descriptions of experience; - contrasts between cases; - invitations to attend; - reports of what seems salient; - distinctions between kinds of awareness; - objections concerning privacy, ineffability, and reportability. - The model can generate those forms as sign-trajectories. - The reader then has to test them against experience, argument, and background theory. - This does not make the model a phenomenologist in the first-person sense. - It may make some generated phenomenological prose usable as material for philosophical reflection. --- ## XI. Examples and diagnostic uses ### 1\. Philosophical prompt - Prompt: - “Explain why aesthetic appreciation of nature should be guided by scientific knowledge.” - Likely trajectory: - Carlson; - natural environmental model; - objectivity; - contrast with picturesque appreciation; - knowledge of geology, ecology, biology; - appreciation as what the object really is. - Semiotic-physics point: - the named issue activates a specific region of analytic-aesthetic discourse. ### 2\. Genre transformation - Prompt: - “Explain Spinoza’s Ethics as a TED talk.” - Likely trajectory: - simplification; - motivational address; - compressed metaphysics; - affective framing; - accessible metaphors; - loss of geometrical argumentative structure. - Semiotic-physics point: - TED-talk genre exerts pressure on philosophical content, altering what becomes salient. - Philosophical lesson: - genre transformation is not merely stylistic; - it changes what kind of understanding the text supports. ### 3\. Argumentative attractor - Prompt: - “Give a balanced view of whether LLMs understand language.” - Likely trajectory: - “on the one hand”; - “on the other hand”; - “depends on what we mean by understand”; - conclusion that the issue is complex. - Semiotic-physics point: - balance prompts often fall into a generic dialectical attractor. - Editing lesson: - specify the live options and the criterion of adjudication, rather than asking for balance. ### 4\. Absorbing academic style - Prompt: - “Write in academic prose.” - Likely trajectory: - inflated transitions; - generic signposting; - “complex and multifaceted”; - “raises important questions”; - overuse of “framework,” “nuance,” “situated,” “interplay.” - Semiotic-physics point: - academic style is represented in the training distribution by many generic markers, so weak prompts activate generic academic attractors. - Editing lesson: - a style prompt without argumentative constraints tends to produce surface markers of academic writing rather than argumentative necessity. ### 5\. Philosophical productive use - Prompt: - “Generate objections to the claim that LLM texts can be philosophically valuable despite lacking authorship.” - Likely trajectory: - authorship objection; - intention objection; - responsibility objection; - originality objection; - understanding objection; - evaluation objection. - Semiotic-physics point: - philosophical dialectic can be generated as an organised space of expected moves. - Paper point: - the usefulness of the output does not require the model to believe the objections. ### 6\. Ideological framing - Prompt: - “Explain universal basic income from liberal, libertarian, socialist, and populist perspectives.” - Likely trajectory: - equality of opportunity; - individual autonomy; - redistribution; - economic justice; - different lexical choices; - different implied agents and values. - Semiotic-physics point: - the model navigates between ideological zones of the semiosphere. - Use: - good for showing that LLM outputs do not simply state neutral content; - they activate frames. ### 7\. Prompt drift in philosophical drafting - Prompt: - “Explain semiotic physics slowly, in a style suitable for an analytic philosophy paper.” - Possible good trajectory: - define the term; - distinguish it from literal physics; - explain simulator/simulacrum; - apply to LLMs; - draw philosophical consequence. - Possible bad trajectory: - generic claims about nuance, complexity, and meaning-making; - overuse of “framework”; - decorative references to semiotics; - absence of a clear argumentative payoff. - Semiotic-physics point: - “analytic philosophy paper” can itself become a generic style-attractor unless the prompt fixes argumentative function. --- ## XII. Possible uses in a paper ### 1\. Use as a descriptive ontology of LLMs - Claim: - LLMs are best understood not as agents, tools, or oracles, but as semiotic simulators. - Role of semiotic physics: - describes the regularities governing their outputs. - Advantage: - avoids anthropomorphism; - avoids reduction to random recombination; - explains prompt sensitivity; - explains persona shifts; - explains genre competence; - explains why the same model can produce incompatible standpoints. ### 2\. Use as an aesthetics of LLMs - Claim: - LLMs can be appreciated as generative artefacts whose outputs manifest a learned semiotic order. - Role of semiotic physics: - supplies the relevant object-knowledge for appreciation. - Carlson-style structure: - appreciate nature as nature, not as painting; - appreciate LLM outputs as LLM outputs, not as ordinary human-authored works. - Payoff: - makes room for appreciation without over-ascribing authorship or mentality. ### 3\. Use as a theory of prompt-based interaction - Claim: - prompting is not command input but semiotic state-setting. - Role of semiotic physics: - explains how prompt features constrain trajectories. - Concepts to use: - initial condition; - perturbation; - attractor; - branch; - trajectory; - drift; - instability; - absorption. - Payoff: - gives a better theory of chat as an environment rather than as a sequence of independent answers. ### 4\. Use as a theory of LLM-generated philosophy - Claim: - LLMs can generate philosophical text worth reading because philosophical argument forms are public semiotic structures. - Role of semiotic physics: - explains how these structures can be propagated without being entertained by a subject. - Payoff: - helps answer authorship, abduction, and phenomenology objections. ### 5\. Use as a critique of generic LLM prose - Claim: - bad LLM prose is often prose captured by low-information semiotic attractors. - Role of semiotic physics: - explains why generic transitions, over-balanced structures, and pseudo-technical phrasing recur. - Payoff: - gives a non-moralised account of LLM stylistic failure; - helps develop better editing and prompting methods. ### 6\. Use as a bridge between technical and humanistic accounts - Claim: - LLMs require a framework that connects token-level probability with cultural-level interpretation. - Role of semiotic physics: - provides the middle layer. - It sits between: - mechanistic interpretability below; - hermeneutics above. - Payoff: - avoids both technical reductionism and purely cultural description. --- ## XIII. Strong argumentative formulations ### 1\. Against the mind frame - The mind frame begins from the wrong object. - It treats the output as the act of a subject. - But the same model can generate mutually incompatible subjects, voices, styles, and commitments. - The stable object is not the simulated speaker. - The stable object is the generative rule under which such speakers can be locally produced. - Semiotic physics describes that rule at the level of signs. ### 2\. Against the oracle frame - The oracle frame treats the model as if it were designed to answer questions truthfully. - But next-token prediction is not question-answering. - A question is just one kind of prompt-state. - The model continues that state according to learned patterns. - Truthful answer, fictional elaboration, evasive response, bureaucratic disclaimer, and plausible falsehood are all possible trajectories. - Semiotic physics explains why correct answering is one mode among others, not the model’s essence. ### 3\. Against the stochastic-parrot frame - The stochastic-parrot frame is right to reject understanding. - It is wrong if it suggests mere repetition or superficial recombination. - The model can generate trajectories under counterfactual configurations not directly present in the training data. - This is possible because training learns regularities of continuation, not only stored strings. - Semiotic physics names those regularities. ### 4\. For aesthetic appreciation - The relevant aesthetic object is not just the text on the screen. - It is the text as a manifestation of a generative semiotic order. - Appreciation can therefore attend to how the output negotiates genre, role, prompt, argumentative form, and cultural memory. - The relevant appreciation is not admiration for a mind. - It is appreciation of a structured process of sign propagation. ### 5\. For philosophical assessment - Philosophical assessment need not always track the mental act by which a text was produced. - It can track the argumentative structure made available by the text. - LLMs can generate such structures because philosophical prose contains recurrent public forms of inference, objection, distinction, and explanation. - Semiotic physics explains how these forms can be propagated by a model without being consciously entertained. --- ## XIV. Risks, objections, and replies ### 1\. Objection: the metaphor is too loose - Worry: - “Physics” suggests law, necessity, measurement, and formal precision. - LLM outputs are messy, cultural, unstable, and context-dependent. - Reply: - The metaphor should be explicitly constrained. - It does not claim strict physical law. - It claims structured regularity, trajectory, transition, attraction, instability, and path-dependence. - Stronger version: - semiotic physics is not physics of signs in the literal sense; - it is a framework for describing the dynamical regularities of sign-generation. ### 2\. Objection: semiotics already covers this - Worry: - Why not just say semiotics? - Reply: - Semiotics explains signs, interpretation, codes, and meaning-making. - Semiotic physics explains probabilistic generation and trajectory dynamics. - The target is not only what signs mean, but why certain sign-sequences follow from others under model generation. ### 3\. Objection: it ignores architecture - Worry: - Output-side study misses the model’s internal mechanisms. - Reply: - Semiotic physics need not deny architecture. - It brackets architecture in order to describe manifest trajectory-level regularities. - This is analogous to studying ecological or behavioural regularities without reducing them immediately to microphysics. ### 4\. Objection: it overstates order - Worry: - LLM outputs are too unreliable to be treated as governed by stable laws. - Reply: - Stochastic systems can still have regularities. - Instability, drift, and hallucination are not counterexamples to semiotic physics. - They are among the phenomena it should explain. ### 5\. Objection: it makes LLMs sound deeper than they are - Worry: - The phrase may inflate what is happening. - Reply: - The account should be deflationary about mind and inflationary only about structure. - It should say: - no consciousness is implied; - no intentional authorship is implied; - no genuine understanding is assumed; - nevertheless, nontrivial semiotic order is present. ### 6\. Objection: it risks collapsing into cultural studies - Worry: - Once we talk about signs, genres, and discourse, the account becomes too broad. - Reply: - The distinctive object is not culture in general. - The distinctive object is the propagation of cultural signs under an autoregressive learned transition rule. - So the account remains specific to LLMs. ### 7\. Objection: it may not be needed for the paper - Worry: - The paper could simply say that LLMs generate structured text without introducing a new metaphor. - Reply: - This is true if the aim is only to make a minimal claim. - The term becomes useful if the paper needs: - a name for the generative order underlying outputs; - a bridge between token-level probability and cultural-level interpretation; - a way to connect LLM appreciation to object-appropriate appreciation; - a way to explain attractors, drift, and prompt-dependence. - Cautious recommendation: - introduce the term; - define it carefully; - use it for work it actually does; - avoid making the entire argument depend on its technical adequacy. --- ## XV. Terminology ### 1\. Terms worth keeping - *semiotic physics* - *sign propagation* - *trajectory* - *initial condition* - *transition rule* - *simulator* - *simulacrum* - *attractor* - *drift* - *branching* - *path-dependence* - *semiotic regularity* - *prompt-state* - *generative order* - *learned space of signs* - *probabilistic pressure* - *trajectory-space* - *semiotic simulator* - *output-side regularity* ### 2\. Terms to use carefully - *law* - useful, but risks overstating precision; - perhaps use “law-like regularity” or “regularity.” - *force* - vivid, but can sound pseudo-technical; - perhaps use “pressure,” “constraint,” or “attractor.” - *physics* - central term, but must be marked as analogical; - avoid implying literal identity with physical law. - *agent* - should usually be restricted to simulacra, not the simulator. - *meaning* - should be split between generated sign-material and interpreted content. - *understanding* - should be distinguished from the generation of understanding-like textual structures. ### 3\. Terms probably to avoid - *magic* - too loose unless quoting or discussing historical reactions. - *emergence* without specification - often vague. - *intelligence* as a central explanatory term - tends to drag the account back toward cognition. - *creativity* without qualification - likely to trigger authorship problems. - *understanding* without scare quotes or explicit analysis - risks anthropomorphism. - *agency* as an unqualified property of the model - risks collapsing simulator and simulacrum. --- ## XVI. Research directions ### 1\. Empirical semiotic physics - Build prompt families and observe trajectory patterns. - Compare: - generic prompts; - role prompts; - genre prompts; - argument-structure prompts; - negative constraints; - examples of desired prose; - iterative chat histories. - Track: - role stability; - concept drift; - recurrence of clichés; - argumentative coherence; - generic attractors; - susceptibility to contradiction; - how long a stipulated distinction remains active; - how quickly the output falls back into generic academic form. ### 2\. Philosophical semiotic physics - Use the concept to analyse what LLMs are. - Key questions: - What kind of object is an LLM output? - What is the relation between model, prompt, output, and user? - What kind of appreciation is appropriate? - What kind of authorship, if any, is present? - What kind of philosophical value can generated text have? - What is the minimal object of assessment: sentence, output, argument, chat, model, or interaction? ### 3\. Aesthetics of LLM environments - Treat the chat as an environment. - The user does not only inspect isolated outputs. - The user navigates a responsive semiotic space. - Appreciation may concern: - the generated text; - the unfolding exchange; - the system’s responsiveness; - the model’s stylistic dispositions; - the way the model transforms prompts into trajectories; - the way the interaction makes some thoughts easier to develop and others harder. ### 4\. Prompting as experimental intervention - If prompts are initial conditions, prompting becomes experimental design. - Each prompt tests how the model propagates a semiotic state. - Better prompts do not merely request better answers. - They structure the trajectory-space more effectively. - This supports a different view of prompting: - not command-giving; - not mere question-asking; - but controlled intervention in a semiotic dynamical system. ### 5\. LLMs and public reason - LLMs are trained on public linguistic traces. - Their outputs reconfigure public argumentative forms. - This makes them relevant to philosophy even without private thought. - Philosophical reasoning has a public textual dimension. - Semiotic physics explains how that public dimension can be generatively recombined. --- ## XVII. Possible section structure for a paper ### 1\. Opening move - Begin from dissatisfaction with standard frames. - Agent, oracle, tool, and stochastic-parrot models each capture something. - Each also distorts the object if treated as exhaustive. - The question becomes: - what model lets us describe LLM outputs without either anthropomorphism or deflationary dismissal? ### 2\. Simulator move - Introduce the simulator/simulacrum distinction. - LLMs generate trajectories by recursively sampling continuations. - The model is a transition rule. - The output is a local trajectory. - The speaker, role, argument, or persona in the output is a simulacrum. ### 3\. Semiotic move - The relevant trajectories are trajectories of signs. - These signs include not only tokens but genres, voices, concepts, arguments, and roles. - Semiotic physics names the regularities governing their propagation. - The term is metaphorical, but it captures something real: structured sign-generation. ### 4\. Aesthetic move - If serious appreciation should track the nature of the object, then LLM appreciation should track this generative semiotic order. - The model’s outputs are appreciable as manifestations of that order. - This does not require treating the model as an author or mind. ### 5\. Philosophical move - Philosophical arguments are public sign-structures. - LLMs can generate such structures because they have learned regularities of philosophical discourse. - Assessment remains with the reader. - The absence of intention blocks one kind of authorship claim, but not all forms of philosophical value. --- ## XVIII. Candidate theses and formulations ### 1\. Strong compressed thesis - LLMs are semiotic simulators: systems that propagate sign-configurations according to learned probabilistic regularities. Their outputs should not be understood primarily as the acts of artificial minds, the reports of oracles, or the products of neutral tools. They are trajectories through a structured space of signs. *Semiotic physics* names the order governing those trajectories: the way prompts, genres, concepts, styles, argumentative forms, and conversational roles constrain what can intelligibly come next. This order is not intentional authorship, but it is also not randomness or retrieval. It is the generative structure in virtue of which LLM outputs can display form, coherence, drift, cliché, surprise, and philosophical usefulness. ### 2\. Cautious version - *Semiotic physics* is a metaphor for the patterned constraints governing LLM text generation. The model propagates a textual state by assigning probabilities to possible continuations; because those continuations are signs embedded in larger cultural and linguistic systems, the resulting trajectories are governed not only by token-level probabilities but also by regularities of genre, style, role, and argument. The phrase should not be taken to imply literal physical laws. It marks the fact that LLM outputs exhibit a trained semiotic order. ### 3\. Stronger version - *Semiotic physics* is the study of the dynamical regularities by which LLMs propagate signs. Since autoregressive models generate text by recursively extending a prompt-state, every output is a trajectory through a learned space of semiotic possibilities. Prompts set initial conditions; sampling produces branching; genres and roles function as attractors; concepts exert local pressure on continuation; and generated trajectories may stabilise, drift, loop, or diverge. The term therefore names the generative order underlying LLM text. ### 4\. Aesthetic version - *Semiotic physics* is the object-appropriate background knowledge for appreciating LLM outputs as LLM outputs. It shows that such outputs are neither ordinary authored works nor random verbal surfaces, but manifestations of a trained order of sign propagation. To appreciate them adequately is to attend not only to what the text says, but to how it emerges from the interaction among prompt, model, genre, role, and interpretive uptake. ### 5\. Philosophical version - *Semiotic physics* explains how LLMs can generate philosophical texts without being philosophical subjects. Philosophical corpora contain recurrent public structures of inference, objection, distinction, and explanation. LLMs can learn and propagate these structures as sign-trajectories. The resulting text may therefore present an argument that is assessable by readers, even if no thinker inside the model has entertained the argument as an argument. ### 6\. One-sentence core - Semiotic physics is the generative order by which an LLM carries signs forward: the learned, probabilistic structure that makes a prompt unfold into one trajectory rather than another. ### 7\. Two-sentence core - Semiotic physics is the system of regularities governing how signs propagate in an LLM. It explains how prompts, genres, concepts, styles, argumentative forms, and conversational roles constrain the trajectory of generation, making some continuations natural, some strained, and others practically unavailable. ### 8\. Three-sentence core - Semiotic physics is the analogue of physics for LLM-generated sign-trajectories. The model does not simply retrieve text or act as a unified speaker; it evolves a prompt-state through a learned probability space in which tokens, genres, roles, concepts, and argumentative forms exert structured pressure on what comes next. This allows us to understand LLM outputs as manifestations of a trained semiotic order without treating the model as an author, thinker, or oracle. --- ## XIX. Best use for the current project ### 1\. Recommendation - The best use of the idea is probably not to present semiotic physics as a technical theory you fully endorse. - The safer and more productive use is to present it as a controlled metaphor and conceptual framework. - The paper can extract from it: - trajectory; - prompt-state; - simulator/simulacrum; - attractor; - sign propagation; - generative order; - probabilistic pressure. ### 2\. How it supports the larger argument - It can then be connected to Carlson-style object-appropriate appreciation. - It can explain how LLM-generated philosophy can be non-random and assessable without being authored by a thinker. - It gives a middle path: - against anthropomorphism: the model is not a mind; - against deflationary dismissal: the output is not mere random recombination; - against simple tool use: the interaction has its own trajectory and order; - for appreciation: the generated text manifests an intelligible semiotic order; - for philosophical value: arguments can be assessed as public structures even when their generative cause is non-intentional. ### 3\. What to avoid - Do not make the paper depend on the claim that semiotic physics is already a mature technical theory. - Do not overstate the analogy with physical law. - Do not imply that LLMs literally understand the signs they propagate. - Do not let the metaphor replace the argumentative work. - Do not use the term as decoration; use it only where it explains something that the alternatives do not explain well. ### 4\. What the term can explain especially well - Prompt sensitivity. - Persona shifts. - Genre competence. - Drift. - Generic attractors. - Recurrent LLM clichés. - The possibility of useful philosophical argument without inner philosophical agency. - The difference between model and simulated speaker. - The difference between generated sign-material and interpreted meaning. - The layered object of appreciation: output, trajectory, model, and interaction. --- ## XX. Remaining questions for development ### 1\. Conceptual questions - How much of the physics metaphor should remain in the final paper? - Should *semiotic physics* be the central term, or should it be introduced and then translated into more conservative vocabulary? - Should the paper distinguish sharply between: - technical semiotic physics; - cultural semiotic interpretation; - aesthetic appreciation; - philosophical assessment? - Should the account treat semiotic physics as: - a metaphor; - a research programme; - a technical hypothesis; - an appreciation-guiding model; - a conceptual bridge? ### 2\. Aesthetic questions - Should Carlson be used explicitly, or should the paper simply inherit the object-appropriate appreciation structure? - Should the focus be on: - outputs; - chats; - models; - human-model interactions; - all four layers? - What kind of appreciation is at stake: - aesthetic; - intellectual; - formal; - technological; - philosophical; - some hybrid? ### 3\. Philosophical questions - Should the argument about LLM-generated philosophy depend on semiotic physics, or should semiotic physics be only one supporting explanation? - Which philosophical objections does the concept answer best? - authorship; - abduction; - phenomenology; - originality; - responsibility; - understanding; - evaluation. - What is the minimal object of philosophical value? - sentence; - paragraph; - argument; - whole output; - chat trajectory; - model disposition; - user-model process. ### 4\. Example questions - Which examples best show the point without sounding like prompt-engineering trivia? - Which examples show genuine philosophical relevance? - Which examples show aesthetic appreciation rather than mere usefulness? - Which examples show failure as well as success? - Should the paper include: - a good LLM philosophical output; - a bad generic output; - a prompt-sensitive contrast; - a genre-transformation example; - a case of drift or attractor capture?