It might seem obvious what sort of knowledge would be required to ground the appreciation of LLMs. LLMs are artefacts, and design appreciation, as Section 1 set it out, is one way of satisfying Carlson’s demand that appreciation answer to the kind of thing its object is. In the case of technical artefacts, this requires knowledge of the purpose for which the object was made and of how its features serve that purpose. LLMs initially look like straightforward candidates for this mode of appreciation: they are engineered systems, and their responses depend on choices made before any user enters a prompt. We shall argue in this section that knowledge of the creation and functioning of LLMs cannot, on its own, ground their aesthetic appreciation. A car is appreciated through its designed form. The shape of its body is at once what the appreciator looks at and what determines how the car moves through air at speed. This shape is what it is because of what the car is for. To appreciate the car aesthetically is to attend to this fit between form and purpose, and design knowledge — knowledge of the car's purpose and of how its features answer to that purpose — reaches the form the appreciator engages with. In the LLM case, design does not completely settle how the system will respond to a prompt; it determines only the conditions under which the model is trained. A transformer architecture gives text a processable form by dividing it into tokens and relating positions in a sequence through attention. The training objective then gives the system a standard against which its outputs can be adjusted: given a stretch of text, the model assigns probabilities to possible continuations, and its weights are altered when those probabilities diverge from the continuation found in the training data. What the trained model then does with a prompt, however, is not settled by these choices; it is acquired over the course of training. When the system responds to user input, it uses what it has acquired in training to continue the text it has been given. The prompt, together with whatever the system has already produced, gives it a context. From that context the model assigns probabilities to the possible next tokens, one of which is selected and added to the context before the same step runs again, so that what appears at the end as a single answer is built through successive transitions of this kind. Later parts of the continuation depend on earlier parts, and each step is shaped by the dispositions acquired in training. The text the appreciator reads is the trained system going on from the prompt — under conditions design has set, but in a direction it has not. Olah describes the relation between what design fixes and what training produces as a kind of growth. > I think one useful way to think about neural networks is that we don't program, we don't make them, we grow them. We have these neural network architectures that we design and we have these loss objectives that we create. And the neural network architecture, it's kind of like a scaffold that the circuits grow on. It starts off with some random things, and it grows, and it's almost like the objective that we train for is this light. And so we create the scaffold that it grows on, and we create the light that it grows towards. But the thing that we actually create, it's this almost biological entity or organism that we're studying. (Olah 2024) What design reaches is the scaffold and the objective; the form that grows under them is acquired through training. In the car case, what the appreciator engages with is the form the designer built. An LLM's response is produced by trained dispositions operating on the prompt and on the continuation as it develops. Design knowledge can explain the conditions under which the response becomes possible; it does not by itself give us the order of the generated text. Carlson's form-follows-function claim also depends on there being a function determinate enough to guide appreciation. The car has such a function: it is made to be driven, and its form is shaped by what that requires. LLMs do not stand in this relation to any function. Next-token prediction is the training objective, but no one consults an LLM in order to have next tokens predicted. The deployment intention that the system should be helpful is closer to use, though too unspecific to constrain the trained form. A prompt can recruit the same system into quite different activities, and the configuration that supports this openness has not been shaped to any one of them in particular. Form-follows-function reasoning needs a determinate function for the form to follow, and an LLM has none. Artefact is too coarse a description for the appreciation Carlson requires. LLMs are artefacts, and design knowledge bears on how they should be appreciated; ignoring how they were built would distort that appreciation. Carlson's question, however, is what the thing is at the level needed for appropriate appreciation. At that level, an LLM is a trained continuation system: a made system whose responses are generated from dispositions acquired in training and operating on context. Post-training helps explain why the generated continuation can seem like a speaker's reply. It is the further stage of training in which the model's responses are shaped toward the form of an assistant's answer. The output therefore arrives in the form of a turn in an exchange. Person-directed knowledge is the next candidate for what design knowledge cannot supply. Section 3 takes the question up. [^1]: Strictly speaking, generation proceeds token by token. Since token boundaries vary across tokenisation systems, the difference can be left in the background. [^2]: The regularities at issue operate at many scales, from local word co-occurrence to the structuring of extended discourse. Calling them dispositions marks this probabilistic and context-sensitive character; it carries no attribution of beliefs, intentions, or other personal states to the model.