“Training models heavily on synthetic data makes them proficient at academic, benchmark-style problems but terrible at real-world use cases.”