SchwitzgebelSchwitzgebelStrasser2023CreatingLLMphilosopher
Eric Schwitzgebel, David Schwitzgebel and Annna Strasser (2023) Creating a large language model of a philosopher
Bibliographic info
Commentary
Very interesting debate obviously. However using the Turing Test as an example of testing “human like” intelligence and then not elaborating on the shortcomings of the test (and the fact that Turing never implied that the test could prove such things, quite the opposite actually), is definitely a weakness. Another weakness is that they go much into depth about the difference between a symbolic model (rule based) and a GPT model (prediction based) as this has an influence on the ongoing debate of intelligence and reasoning.
Innate reasoning
Page 238:
Concerning language, it might be thought that the capacity to generate novel, interpretable, and structurally sophisticated prose—unless that prose is the mere (stochastic) parroting of preexisting material (Bender et al., 2021)—would require symbolically encoded, innate linguistic capacities (e.g., Berwick et al., 2011; Chomsky, 2007); or at least that it would require “functional” competence via dedicated mechanisms for formal reasoning, pragmatics, knowledge of interlocutors, and so forth, which the best known and most successful AI systems generally lack (Mahowald et al., 2023; Shanahan, 2023).
Comment:
Especially the last part of this passage seems important to me as AI models such as GPT-3 indeed do lack the ability to reason and do not possess “innate linguistic capacities”. However, as stated above I feel as though the paper should also mention why they deem this so important for Natural Language Understanding and why GPT-3 not possessing these characteristics poses a problem.
No explicit representations of Philosophical concepts
- Page 253:
Philosophers and cognitive scientists might find it surprising that pure probability calculations, without explicit representations of philosophical concepts or external features of the world, can produce what seem to be novel and philosophically contentful answers (for a review of this debate, see Buckner & Garson, 2019)
Comment:
Isn’t the whole field of cognitive science based on finding out how the result is bigger than the sum of its parts? Perhaps surprising was not a good word to use. I don’t think it is surprising that probability calculations emerge as something that seems contentful. There are many phenomena in the world that show such emergence that cannot be explained by just breaking it down. What is more interesting is whether the product in this case is truly novel and contentful and not just an approximation or shortsighted extrapolation of what has been fed into it. If we keep creating content using LLM’s and never feed it anything human made anymore, just the content it or other models created, how long will it take until it stops surprising us with these seemingly “contentful” answers? Note that I am talking specifically about LLM’s such as GPT-3 in this case. Building an LLM based on explicit reasoning is another story. But those don’t perform anywhere near GPT-3 at the moment. On the other hand, something that is interesting to think about is if humans even have thoughts or reason so explicitly as that of a symbolically encoded model.
Using LLM’s as a thinking tool
- Page 255:
A composer or artist might create many outputs, choose the most promising, edit them lightly, and present them, not unreasonably, as original work—a possibility suggested by Dennett himself in personal communication. In such cases, the language model would be a thinking tool that is used by humans. Similarly, in philosophy, experts might fine-tune a language model with certain corpora (e.g., their own corpus, or that of a favorite interlocutor or historical figure, or an aggregate of selected philosophers), generate a variety of outputs under a variety of prompts, and then select those that are the most interesting as a source of potential ideas.
Comment:
I find it interesting that Daniel Dennett himself suggests that it is not unreasonable to present multiple outputs of an LLM as original work. Of course, by using multiple outputs of an LLM, one has to think critically of what to ask and how to edit and use that. However, I would like to argue that writing is also an important part of the academic process. “Edit them lightly” just seems like something that can mean several things. Does it mean edit the text so it fits into your work, or edit it so it is written in your own words? Then perhaps it would be sensible to note the LLM as you co-author.