According to the externalist view of referential semantic competence in the philosophy of language, language users learn references not by directly interacting with referents (see also: Augustine’s “semantic” version of Meno’s paradox), but by gleaning referential meaning from “linguistic division of labor” or historical-causal chains of usage.
By appealing to externalism, @2024millierePhilosophical reframe the question of how large language models may come to have referential semantic competence as the question of whether LLMs can belong to a linguistic community: “If reference can be determined by a word’s history of use within a linguistic community, then LLMs may inherit referential abilities by being appropriately linked to the causal chain of meaningful word use reflected in their training data.”
Reading notes
- “When I believe that water is wet and my twin believes that twin water is wet, the external features responsible for the difference in our beliefs are distal and historical, at the other end of a lengthy causal chain. Features of the present are not relevant: if I happen to be surrounded by XYZ right now (maybe I have teleported to Twin Earth), my beliefs still concern standard water, because of my history. In these cases, the relevant external features are passive. Because of their distal nature, they play no role in driving the cognitive process in the here-and-now. This is reflected by the fact that the actions performed by me and my twin are physically indistinguishable, despite our external differ.” (9)