# Tacit_knowledge_LLMs  ## Metadata - Author: [[readwise.io]] - Full Title: Tacit_knowledge_LLMs - Category: #articles - Summary: insert summary - My notes: - Summary: The paper discusses how Large Language Models (LLMs) can acquire tacit knowledge, which refers to implicitly represented rules or structures that affect the system's behavior. By meeting constraints such as semantic descriptions and causal systematicity, LLMs can be said to possess tacit knowledge, allowing for better understanding and intervention in their functioning. - URL: https://readwise.io/reader/document_raw_content/155334732 ## LLM Chats ## NotebookLM ## LLM Audio ## Highlights > In the particular case of LLMs, it has been suggested that they do not just perform > next-word prediction based on surface statistics, but that they learn something akin to symbolic rules (Pavlick 2023), linguistic competence (Mahowald et al. 2023), or even knowledge (Meng et al. 2022; Yildirim and Paul 2023). Specifically, LLMs seem to successfully acquire aspects of syntactic knowledge and formal linguistic competence (Linzen and Baroni 2021; Mahowald et al. 2023). While it remains an open question whether LLMs acquire knowledge of semantics, various studies in the technical literature claim to have identified representations of world models (Li et al. 2023), implicit meaning (Hase et al. 2021), and factual knowledge (De Cao, Aziz, and Titov 2021; Meng et al. 2022). ([View Highlight](https://read.readwise.io/read/01hstntwnj0pmgbr8t2g5ay5re)) - Tags: [[toread]] > That is, I argue that if the constraints as set out by Davies (1990) are met by a particular LLM, that system can be said to have acquired a form of knowledge: tacit knowledge. Tacit knowledge, in this context, refers to implicitly represented rules or structures that causally affect the system’s behavior. As connectionist systems are known not to have explicit knowledge, tacit knowledge thus provides a promising way to characterize and identify meaningful representations in the model internals. ([View Highlight](https://read.readwise.io/read/01hstnwrg33yyk8w4bzyc0pjzx)) > Transformer architecture particularly suitable for language is the addition of an attention mechanism (Bahdanau, Cho, and Bengio 2016), which allows the network to keep track of long-range dependencies and to integrate information about previous tokens (i.e. context) in its predictions. That is, the attention mechanism allows the network to learn what previous words or parts of the input sequence it should focus on when processing a particular word or token. ([View Highlight](https://read.readwise.io/read/01hstp0cdkb5cc26snej8af45a)) > There is fairly robust evidence that LLMs learn and represent grammatical rules, i.e. syntax (Linzen and Baroni 2021; Mahowald et al. 2023). In contrast, technical research on semantic aspects of knowledge, such as world knowledge (Li et al. 2023; Yildirim and Paul 2023) or factual knowledge (Geva et al. 2021; Meng et al. 2022), is still in early stages. ([View Highlight](https://read.readwise.io/read/01hstp53yvggsk51vw3mdgbvk6)) > If there are no explicit rules, it could be argued that we can only say that the behavior of the system seems to conform to these rules, but not that it is actually guided by them ([View Highlight](https://read.readwise.io/read/01hstpa345gmza5trypk4atx5e)) > According to Davies, tacit knowledge can only be attributed to systems whose input and output states have semantic descriptions, i.e. which register or represent something. Davies defends this point by introducing a counterexample. Take a billiard ball with mass m that moves with a certain acceleration a in response to a push of force F. It could be argued that this billiard ball has tacit knowledge of the rule a = F/m. This seems absurd, however: the billiard ball does not have knowledge of the rule, though its behavior may conform to it. ([View Highlight](https://read.readwise.io/read/01hstpdxjtrrmkwghv8xwy0xc7))