**Links to**: [[Imitation and adaptation]], [[Private language argument]], [[Kripke]], [[Wittgenstein]], [[Rule]], [[Rule induction]], [[Rule-following paradox]]. _Kripkenstein is covered in [[Imitation and adaptation]]. Below some notes for further thinking/writing about this_. #todo ### [[Prompt engineering]]: how to think how the machine thinks to prompt it for our purposes? #### GPT3, Floridi and Chiriatti, CJ Blunt >“GPT-3 works in terms of statistical patterns. So, when prompted with a request such as “solve for x: x+4=10” GPT-3 produces the correct output “6”, but if one adds a few zeros, e.g., “solve for x: x+40000=100000”, the outcome is a disappointing “50000” (see Fig. 3). Confused people who may misuse GPT-3 to do their maths would be better of relying on the free app on their mobile phone.” > > Floridi & Chiriatti, 2020, p. 688. >They are correct. GPT-3 is pretty good at simple maths with small numbers (e.g. “What is 43+68?” or “50-21=?“) but struggles once the numbers get larger. OpenAI studied the mathematical abilities of GPT-3 in detail, and reported a far more detailed analysis of its mathematics (Brown et al. 2020). The model performs almost perfectly in two-digit addition.” >From https://cjblunt.com/imitating-imitation/. #### Murray Shanahan >“Could a user-provided image stand in for an external reality against which the truth or false- hood of a proposition can be assessed? Could it be legitimate to speak of a VLM’s beliefs, in the full sense of the term? We can indeed imagine a VLM that uses an LLM to generate hypotheses about an image, then verifies their truth with respect to that image (perhaps by consulting a human), and then fine-tunes the LLM not to make statements that turn out to be false. Talk of belief here would perhaps be less problematic. >However, most contemporary VLM-based systems don’t work this way. Rather, they depend on frozen models of the joint distribution of text and images. In this respect, the relationship between a user-provided image and the words generated by the VLM is fundamentally different from the relationship between the world shared by humans and the words we use when we talk about that world. Importantly, the former relationship is mere correlation, while the latter is causal.10 >The consequences of the lack of causality are troubling. If the user presents the VLM with a picture of a dog, and the VLM says “This is a picture of a dog”, there is no guarantee that its words are connected with the dog in particular, rather than some other feature of the image that is spuriously correlated with dogs (such as the presence of a kennel). Conversely, if the VLM says there is a dog in an image, there is no guguaranteehat there actually is a dog, rather than just a kennel.” >Murray Shanahan, “Talking about Large Language Models,” arXiv preprint, 2023, p. 7.