You've probably seen it, but a more recent, related paper (that I think has some of the same authors) about inherent features of modern ML that make models so f

You've probably seen it, but a more recent, related paper (that I think has some of the same authors) about inherent features of modern ML that make models so fragile, even if they test OK: Underspecification Presents Challenges for Credibility in Modern Machine Learning, D'Amour et al., https://arxiv.org/abs/2011.03395 https://arxiv.org/abs/2011.03395