“Softmax's technical strategy for achieving AI alignment is to use large-scale, multi-agent reinforcement learning simulations to create a surrogate model for cooperation and theory of mind.”