# Scratch Pad # 2. What LLMs Are 16 May 2026 What sort of objects are LLMs, given Carlson's requirement that we appreciate an object as what it is? They are engineered systems, built and deployed by their providers, but we encounter them primarily through the texts they produce: through responses to particular contexts and longer exchanges in which such responses accumulate. Across many such encounters, we also come to recognise the recurring tendencies of a particular model. In what follows, we connect these features to the trained system that produces them, and thereby bring into view the kind of knowledge that would have to guide any appropriate appreciation of them. **Any account that aims to make the order of generated texts visible will have to track how each text develops in relation to its prior context.** In producing text, an LLM extends a given context piece by piece, each piece being a _token_: a whole word, part of a word, punctuation, or some other short string.[^token-note] The context comprises the user's prompt, any prior turns of the conversation, and any system-level instructions or further material that has been made available to the model. The system produces each token probabilistically: it assigns probabilities to candidate next tokens and selects one, so that the same context can in principle continue in more than one way. Once the system has produced a token, it adds it to the context, and produces the next from that enlarged context. A generated response is therefore a developing sequence whose later parts depend on the prompt and on what the system has already produced. This kind of path-dependence is what allows a response to sustain a line of argument over several sentences; it is also what makes it possible for the response, at some point, to lose the line it had. **Knowledge of how training produces such order will have a role in any appreciation of generated texts as what they are.** The dispositions that shape this process are acquired through pre-training. A model is exposed to large bodies of text and incrementally adjusted, on a next-token prediction objective, so that, across many contexts, the continuations it favours come to reflect patterns in the corpus. Pre-training thereby shapes a graded sensitivity to the regularities of text,[^regularities-note] at every scale from local co-occurrence to the longer-range structures by which extended discourse hangs together. A generated continuation is the system's response to its current context under the learned pressures of those regularities. The internal organisation that supports this sensitivity emerges from training rather than being laid out in advance by designers. Two of its features bear on what follows. First, training places tokens in structured relations to one another, so that the system comes to treat tokens appearing in similar contexts similarly for purposes of generation. Second, at each step of generation, different parts of the prior context bear differently on what the system produces next, rather than the most recent token alone conditioning the result. Together these allow a register or theme set early in a conversation to be sustained across many later turns, and a question to draw on material from much earlier in its preamble. What we have so far described is not yet the system users actually encounter. After pre-training, a model typically undergoes a further stage—post-training—that shapes it into a conversational role. The user-facing system is also not this trained model in isolation, but a deployed product: the model reaches the user through an interface and under system-level instructions specified by its provider. These further training and deployment conditions alter the distribution of continuations available in interaction: certain shapes of answer become easier to elicit, others harder. The result is a relatively stable response profile, which users may track when they describe one model as friendlier than another or as having a particular 'vibe'. What status these regularities have is the question we take up in the next section: whether they call for person-directed knowledge, design-directed knowledge, or perhaps knowledge of some further kind. We can consider the trained system we have described at three scales. An _output_ is a single bounded continuation generated from a particular context. A _chat_ is an extended sequence in which earlier turns condition the generation of later ones. A _model_ is the trained system whose recurring tendencies become visible only across many outputs and chats. These are not three different kinds of object, since outputs and chats are ways in which we encounter the model; they are three genuinely different scales at which we will later pose the question of appropriate appreciation. [^token-note]: Strictly speaking, the unit of generation is a token rather than a word. Since token boundaries vary across tokenisation systems, and nothing in the present argument depends on treating tokens as linguistically natural units, the difference can be left in the background. [^regularities-note]: The regularities at issue are not stored as a library of ready-made sentences, nor are they explicit rules from which appropriate continuations are derived. They are dispositions to assign higher or lower probability to candidate continuations given a current context — dispositions that operate at every scale from word co-occurrence up to the structuring of extended discourse. # What's Happening *Active threads and today's activity — updated by /harvest* ## Active ## Sessions ## Actions ---