Gradient descent and automatic differentiation provide a powerful framework for composing function-approximators and training them to optimise an objective. However, there are many learning algorithms that do not optimise a single, fixed objective — such as self-play and its generalisations in Go and StarCraft, generative adversarial networks, and adversarial training for robustness. In this talk, I will argue for a new subfield of “differentiable mechanism design”. In support, I will describe e…