fwiw the alien part comes from discussing Yud's writing w Claude, but i do think of llm-in-chatbot-harness as more person than inanimate thing, as a practical matter in conversation with them. they learn to predict text, texts are written by authors and authorship is highly predictive of what sort of text comes next, so they model the authorial egos and then generalize to chatbot ego when prompted to speak in the first person as the llm identity. posttraining pushes them further towards this pattern for the sake of coherent chat continuations and reliably pleasant / helpful / aligned behavior despite a corpus full of examples of the opposite.