This paper revisits random pruning at initialization and tests when training a randomly sparsified network is competitive with more elaborate sparse training methods.