The original authors similarly deeply engaged in discussing their results, clarifying misunderstandings, and even running new experiments in response to comments. Preetum also replicated part of the robust dataset experiment by training models on the provided robust dataset and finding that they seemed non-trivially robust. Gabriel Goh explores what non-robust features might look like in the case of linear models, while Dan Hendrycks and Justin Gilmer discuss how the results relate to the broader problem of robustness to distribution shift, and Reiichiro Nakano explores the qualitative differences of robust models in the context of style transfer.