# Scratch Pad # 4. Semiotic Physics Section 2 identified LLMs as trained continuation systems: made systems whose responses develop from dispositions acquired in training and operating on context. It also showed why design knowledge cannot, by itself, guide their appreciation. Design knowledge reaches the scaffold, the objective, and the conditions under which the system is trained and deployed. It does not by itself make visible the order acquired by a particular continuation as context is extended. Section 3 then showed why person-directed knowledge does not reach this object either. It reaches, at most, the person-like figure or profile made available through interaction. These bodies of knowledge therefore leave the same task unsettled. Carlson’s account gives that task a determinate form: the relevant knowledge must make the learned order of generated text visible as the order of a trained continuation system. Several sub-disciplines of computer science might be candidates. One field that has emerged in connection with neural networks is mechanistic interpretability, which investigates the internal workings of these systems by identifying which circuits, attention heads, and internal representations handle different linguistic tasks (Olah et al. 2020; Elhage et al. 2021). This work yields knowledge of how LLMs operate. But mechanistic interpretability functions at a level that requires specialist tools to observe. Its objects of study – weight matrices, activation patterns, circuit-level features – are not available to readers encountering generated text. Consider the difference between chemical physics and geology when appreciating a cliff face. Chemical physics provides knowledge of molecular bonds within rock, but it operates at a scale invisible to the naked eye. Geology, by contrast, offers concepts – strata, faults, erosion channels – that connect to what can be seen. One can perceive strata without specialist equipment, and knowing how sedimentation works makes the visible layering intelligible. Mechanistic interpretability faces a parallel limitation: while it reveals internal mechanisms, its objects of study are hidden from the user reading generated text. For an aesthetics of LLM outputs that is accessible to ordinary users, we need a framework whose concepts describe perceivable features and render them intelligible as products of the system’s learned regularities. The candidate we propose is semiotic physics. Semiotic physics, as we use the term, is an output-side account of how trained systems develop linguistic forms through iterated continuation from context. Section 2 described generation as iterated continuation. At any point in a run, the system receives the context so far and computes a distribution over possible next tokens. Once one token is selected, the context changes, and the next step is produced from that changed context. Training gives the system a graded sensitivity to the regularities of text — to what tends to follow what under what conditions. When the model is run, those regularities operate through a context that changes as the text develops. The prompt fixes the starting point; post-training and deployment affect which continuations are likely to be produced from it; sampling makes the text one realisation among other possible paths. Metasemi formulates this shift from isolated prediction to generated path: “It’s more illuminating to consider what happens when GPT . . . is run repeatedly to produce a multi-token forward trajectory, as in the familiar scenario of generating a text completion in response to a prompt” (metasemi 2023). The path generated by the model is computed token by token; the order at issue is encountered by readers as language. Semiotic physics has to stay tied to both features of the case. The trained system has acquired patterns governing what tends to follow what under given conditions; this is the structure Janus describes when he treats GPT-style models as simulators of a learned distribution (Janus 2022). The continuation also appears as linguistic form, which Picca captures when he describes LLMs as systems that “recombine, recontextualize, and circulate linguistic forms based on probabilistic associations” (Picca 2025, 1). Wolfram’s image of a trajectory in linguistic feature space connects generation and linguistic form: continuation traces a path through a learned space whose structure bears on whether the resulting text is meaningful (Wolfram 2023). The order made visible by semiotic physics is therefore the order of linguistic forms carried forward and transformed through continuation. Generated text develops by carrying forward what it has already produced. A prompt may set the task; the later shape of the continuation is also conditioned by material that has appeared in the output itself. This is why a response can gather a direction as it proceeds, or lose the direction it seemed to have. Semiotic physics directs aspection toward this path-dependence. The reader attends to the developing relation between earlier and later parts of the generated text, and understands that relation as a product of iterated continuation. The same account explains why person-like profiles remain aesthetically salient after the person-directed route has been rejected. A generated text can sustain a recognisable response profile across a continuation. Section 3 argued that such a profile is not the character of a subject. Semiotic physics treats it as a pattern in text propagation. One might object that speaking of forces in relation to LLMs is metaphorical in the same way that speaking of agents or intentions is metaphorical. The question is whether force-talk repeats the mistake of projecting the wrong kind of structure onto a trained continuation system. Agent-talk attributes a subject with beliefs, intentions, a life, or character. Force-talk, in the restricted sense needed here, identifies factors that condition the continuation of text. The analogy with physics goes no further than regularity, constraint, and dependence on prior state. Its use requires regularities stable enough to guide attention to how generated text develops. Knowing that a text is LLM-generated rather than human-written changes how we aspect it. A human-written text is normally read as the product of authorial selection. An LLM-generated text can be read as a continuation shaped by context, learned regularities, and post-training. The same words on the page can therefore become appreciable under a different aspect. Semiotic physics works with ordinary linguistic competence. A reader already has a tacit sense of register, direction, and coherence. Semiotic physics gives that sensitivity a causal articulation by relating these patterns to training and generation. Familiar talk of model vibe can then be understood as a way of registering stable differences in how models propagate text. This order is encountered at more than one scale. A single output is one bounded continuation from a context. An extended chat is a longer process in which earlier turns condition later ones. A model is the trained system whose tendencies become visible across many such outputs and chats. These are different scales at which the same kind of semiotic order can be appreciated. The next section considers each in turn. # 3. LLMs and Person-Directed Knowledge Person-directed knowledge is a natural candidate for what design knowledge leaves out. This temptation arises because the way we interact with LLMs is so similar to interacting with actual human interlocutors. We sometimes appreciate persons aesthetically. A person’s warmth may be aesthetically appreciable as a feature of character, rather than as a feature of bodily appearance. Work on beauty of character treats such appreciation as directed at the traits and dispositions through which a person’s life is intelligible (Gaut 2007; Paris 2018). Something similar is possible with fictional characters: we can aesthetically appreciate a character as a person within a fiction, without believing that the character exists outside it. The account of LLMs in Section 2 puts pressure on this comparison. The system producing the text is a trained continuation system: it generates responses by applying dispositions acquired in training to the context it is given. The person-directed thought can be preserved in two ways. The first grants that the LLM is not literally a person but reads its outputs as fictional speech, so that what is appreciable is a fictional character rather than the system itself. The second argues that the LLM is an intentional system in a thin sense, and that this is enough to license some person-directed appreciation. Mallory (2023) takes engagement with a chatbot to be a game of prop-oriented make-believe. The chatbot provides text that functions as a prop for imagining an interlocutor. Its outputs are "literally meaningless but fictionally meaningful" (Mallory 2023, p. 1082). On Mallory’s account, generated strings can function as props: they make it fictional that a character has said something, while the system that generates them remains distinct from that character. Mallory develops this as a metasemantic and epistemic account rather than an aesthetic one. Adapted to the present question, it suggests that the person-like object of appreciation is the fictional interlocutor made available by the game. The warmth or wit a user finds in Claude or in ChatGPT is the warmth or wit of a fictional interlocutor that the prop, played with, brings into the game. Mallory’s fictionalism depends on keeping the prop apart from the character it makes available. Mallory is explicit: "the character is not this technological infrastructure any more than a character in a play is a human body or a costume" (Mallory 2023, p. 1091). The fictional interlocutor enters the game; the LLM is what makes the game possible. Person-directed appreciation can therefore attach to the interlocutor, but the interlocutor is not the LLM. The trained continuation system Section 2 identified remains to be appreciated as what it is, and Mallory's account leaves that task untouched. Frankish offers a more direct route, because he keeps the intentional description attached to the system itself. Drawing on Dennett’s intentional stance, he argues that we can ascribe beliefs and desires to a system when doing so yields a simple and fruitful account of its behaviour. On the shallow view he adopts, these attitudes need not be inner episodes or conscious states. They can be dispositional patterns in the behaviour of a whole system. This makes the view unusually hospitable to LLMs: if treating a model as having beliefs and desires helps predict its textual behaviour, then the intentional description is not merely a fiction imposed from outside. Frankish’s own account also marks the limits of this move. LLMs are static systems with no needs, no social life, and no updating of their inner architecture through interaction. For that reason, he does not credit them with the range of communicative desires we ascribe to human speakers. The thin agency he allows them is organised around a single goal: to play the chat game, producing textual responses that are cooperative by human conversational standards, given the context. The chat-game agent, however, falls short of what person-appreciation needs. A person's linguistic behaviour, as Frankish himself notes, is embedded in a wider web of non-linguistic behaviour, and tracking the predictive patterns of that whole web involves ascribing a much wider range of desires than the chat game alone calls for (Frankish 2024, p. 15). The chat-game agent has nothing answering to that web. The intentional structure Frankish licenses is restricted to making moves in the chat game; it does not place the text within a life in which earlier conduct informs later conduct. The categories person-appreciation works with depend on this kind of life. Steadiness, for instance, can only be shown across the situations that make up someone's life. The chat-game agent does not have one. Post-training shapes generated text into stable patterns, and this might seem enough for person-appreciation. Users do say that one model feels friendlier than another, and they are picking up on something — a regularity in how chat-optimised systems tend to respond across many prompts. The regularity belongs to the way post-training disposes the system to continue prompts in assistant-like ways. It is not a trait of a temporally extended subject whose responses today are continuous with their responses last week. The categories of person-directed appreciation, applied here, lack the kind of subject they normally require. Neither person-directed route reaches the object isolated in Section 2: the trained continuation system whose responses develop from acquired dispositions operating on context. The remaining question is what kind of knowledge makes such generated text appreciable as the product of a trained continuation system. # What's Happening *Active threads and today's activity — updated by /harvest* ## Active ## Sessions ## Actions ---