Current models do great on well-specified, verifiable benchmarks, and not so great on unbounded, underspecified evaluations that depend on the users' taste. But you need both to eventually match or exceed human creativity. So we made a suite of them: https://drops.mts.now/creative