# LLMs Don't Know Anything: Reply to Yildirim and Paul  ## Metadata - Author: [[Mariel K. Goddu]] - Full Title: LLMs Don't Know Anything: Reply to Yildirim and Paul - Category: #articles - Summary: insert summary - My notes: - Document Tags: [[generating philosophy paper]] - Summary: The authors argue that large language models (LLMs) do not possess true knowledge; they are merely tools created by humans to perform specific tasks. They emphasize that LLMs can predict outcomes but lack understanding of the causal relationships behind those predictions. Ultimately, they contend that LLMs should not be viewed as agents with knowledge, as they do not have the ability to know or do anything independently. - URL: https://readwise.io/reader/document_raw_content/254173282 ## LLM Chats ## NotebookLM ## LLM Audio ## Highlights > This is a misleading character- ization of ‘instrumental knowledge’.The term is typically used to refer precisely not to statistical or associative learning, but rather, by contrast, to knowledge achieved by humans and other animals on the basis of appreciating contingen- cies between their actions and outcomes (e.g., operant conditioning) [4]. Whereas instrumental knowledge, properly under- stood, is ‘difference-making’ knowledge that supports causal intervention [5–9], the capacity to guess what comes next, no matter how robust, is no such ability. No amount of prediction, from a set of observations no matter how large, can support the grasp of the notion of ‘making’, ‘generating’,or ‘doing’ that is the basis of our world knowledge. ([View Highlight](https://read.readwise.io/read/01jg6g14s9pmjndctkh63pmezk)) > The map does not know the way home, and the abacus is not clever at arithmetic. ([View Highlight](https://read.readwise.io/read/01jg6frsv69dwtze7711pdee13)) > Imagine seeing a sequence of lights on the wall: green, yellow, red; green, yellow, red. Given these observations, you may suc- cessfully predict the next color (green). The capacity to make this prediction, a form of statistical inference, is distinct from the capacity to understand what generates the sequence, or what would change it (e.g., knowing it is made by the trafficlight outside, without which it would not occur). Evidence from developmental psychology suggests that the human capacity to ap- preciate this type of dependency (i.e., to have a generative concept of ‘cause’) de- velops from our experience of our own and others’ goal-directed actions [5]. Human learning occurs via active participa- tion in world affairs, by making the differ- ences that we need and want to make in our lives as social, biological organisms. ([View Highlight](https://read.readwise.io/read/01jg6g3nfz8jt6s66yeyzeq2t7))