All reviewers agreed on the technical quality of this paper and that the ideas are interesting. Opinions diverged on (a) the possible applications and (b) the degree to which these matter. There was extensive discussion amongst all reviewers (and I thank them for this). By the end, I think many reviewers shared a concern that the motivation for this work is a remaining weakness. I strongly recommend this be taken seriously in a final revision if the impact of this paper is to be maximised. Despite this, NeurIPS is a suitable venue for theoretical work and given this article is otherwise well written and interesting, I find it reasonable that the community can decide to what extent they are excited about this after publication.