Ask HN: Do Large Language Models Have a Decodable “Internal Language”?

Summary

An Ask HN post asks whether large language models translate English and other natural languages into an internal language before predicting the next token. The author describes this internal language as vector representations that might encode relationships such as gender and semantic roles, using analogies like “king” and “queen” as an illustration. The post also suggests that similar representations may connect words across languages, while noting that natural languages contain irregularities. Drawing a comparison with constructed languages such as Loglan, Lojban, and Ithkuil, the author proposes creating a lexicon that maps internal vectors to interpretable terms. If such a mapping could be learned, communication with models might become more direct and less affected by linguistic ambiguity. The post presents these ideas as a question and asks whether anyone is already working on them; it does not provide experimental evidence that LLMs possess a discrete internal language.