One <a href="https://arxiv.org/abs/2402.10475" target="_blank">paper</a> was accepted to the ICLR 2024 Workshop on <a href="https://sites.google.com/view/bgpt-iclr24" target="_blank">Bridging the Gap Between Practice and Theory in Deep Learning</a>. We prove that alternating gradient descent-ascent (GDA) converges faster than simultaneous GDA, and propose an even faster variant!