“Large language models do not learn from unexpected outcomes during a conversation; they do not adjust their internal state or parameters based on surprise.”