“Edwin Chen argues that an over-reliance on synthetic data makes models good at academic, benchmark-style problems but "terrible at real world use cases."”