This paper presents a series of analyses to disentangle the role of the locations of weights from their values. By training networks to mirror locations, but not values, the paper shows that networks can be trained to good performance. All the reviewers found the work to be interesting and novel, though there were some concerns about clarity/focus. While I agree that there are some clear ways to improve the clarity, I think the paper has enough strengths to merit acceptance. However, I do encourage the authors to take the feedback regarding clarity to heart and focus on improving this for the camera-ready version. The paper will benefit in the long run from some additional work in this area.