# [[Scratch Pad]] ## Message to an LLM about [[the Philosophy]] of Street Art and Type faces ## Message to an LLM about understanding and [[dependency relations]] Ok thanks, that was not bad. but I would like you to do that again taking the following into consideration: 1. You failed at following my instructions. I explicitly told you that I don't give a monkeys about whether or not 'LLMs truly understand'. The only reason I can see it mattering is if it impedes their outputs. 2. I don't think you focussed enough on the nitty gritty of the theory of understanding provided in sections 3.1 and 3.2 of the text. I would have liked much much more focus on how what the text says in the following passages about understanding and [[dependency relations]]: "Kim's idea is developed by Dellsén (2020), who argues that understanding some X is not simply a matter of accurately representing how X depends on various other phenomena (or aspects thereof), and thus that understanding X is not simply a matter of having some or all explanations of X. Rather, argues Dellsén, understanding X is a matter of representing both how X depends on various other phenomena, and how further phenomena depend on X itself. In other words, the extent to which one understands X is a matter of how one represents the network of [[dependence relations]] running both to, and from, X.10 Furthermore, the relevant representations concern not only ‘positive’ facts about whether (and if so, how) X depends on other phenomena, or how other phenomena depend on X. Rather, they also concern ‘negative’ facts like X's lack of dependence on specific other phenomena, or indeed on any phenomena (Dellsén, 2020, 170–182)." "Understanding is taken by many to be a gradable notion, in that one may understand X to a greater or lesser extent (rather than just fully or not at all). By contrast, knowledge is usually taken to be non-gradable, in the sense that there are no degrees of knowing a given proposition P; an agent either knows P or doesn't know P. We will follow Dellsén in supposing that the extent to which one understands X is determined by two separate criteria, accuracy and comprehensiveness. Accuracy concerns the extent to which one's representation correctly represents that X does or does not depend on various other phenomena (and how), and that various other phenomena do or do not depend on X (and how). Comprehensiveness concerns the extent to which one's representation includes all the phenomena on which X does and does not depend, and which do or do not depend on X. As Dellsén (2020, 167) notes, these criteria may come into conflict in certain cases, leading to idealisation (sacrificing accuracy for comprehensiveness) or abstraction (sacrificing comprehensiveness for accuracy)." 1. Another thing you didn't get right was that your interpretation of [[the idea]] that LLMs allow us to enhance our (a philosopher's) representation of [[dependency relations]] was too flat-footed. You could say to an LLM 'map out the [[dependency relations]] for this topic' but [[the idea]] I had in mind was much more along the lines of simply engaging in dialogue with LLMs would enhance [[philosophical understanding]]. [[The idea]] was that this enhancement of understanding was implicit in the same way that when I increase my understanding of philosophy by reading or engaging in dialogue with a fellow philosopher I do not explicitly think in terms of [[dependency relations]]. 2. having said all this. I just thought of another idea. in early 2025 (i.e. now) researchers started to say that maybe the future was training specialist llm models for specific tasks. could, potentially, a specialist LLM be trained to enhance [[philosophical understanding]]? that is, could we train an llm that is very very good at providing output that enhances [[dependency relations]] either implicitly or explicitly? Finally, please maintain or enlarge [[the level]] of detail of the report you just gave me.