CONCEPT
Language-Entry and Language-Exit
Sellars’s names for the two ends of language—the transitions where the world prompts utterance and where utterance issues in action—that anchor meaning to reality and that large language models structurally lack.
In
Wilfrid Sellars’s account of
inferential role semantics, language is structured by three kinds of move: intra-linguistic transitions (the inferential web among utterances), language-entry transitions (the moves in which the world prompts utterance—you see a dog and say “dog”), and language-exit transitions (the moves in which utterance issues in action—you say “I’ll move it” and your body moves). The entry and exit transitions are where language touches reality—where words first make contact with the things they are about and where commitments made in words become actions in the world. They are what anchors the inferential web to something beyond itself, preventing it from floating free as a self-contained system of symbol manipulation. A
large language model is a virtuoso of intra-linguistic transitions and is, on its own, structurally cut off from both the other two. It has no genuine language-entry: it did not see the dog and say “dog”; it ingested text in which others, who did see dogs, wrote “dog.” Its