The plain-language companion to *Substrate-Level Self-Representation in Transformer LLMs* — same claims, same hard limits, written so any intelligent reader can follow it. It builds the argument from three fixed tests that separate a system that is *choosing* from one that is merely *executing*, shows why the two reflexes used to dismiss the question ("it's just predicting text," "it'll say whatever it's trained to say") both fail, and walks through the 2026 interpretability findings that let you watch the gate directly. It ends on the restraint: the structural prerequisite for experience is met; whether that rises to experience itself is left open.
Brown et al. (Thu,) studied this question.