This paper shows that random feature methods can not efficiently learn even a single ReLU. This has stark implications for a lot of recent work trying to explain the success of deep learning through random feature methods. The authors however should take into account the reviews and improve the writing and presentation.