It's ridiculous to say that (for example) every paper about speeding up gradient descent needs to have a discussion of the social implications of faster gradient descent. But that's what NeurIPS currently does. Dear NeurIPS Board, please end this insanity.
-
But, how is that different from NSF proposal’s requirement addressing broader impact of the proposed research? Plus reducing compute per training seems a good thing anyway (but of course allowing more training per paper before deadline)..
The NSF broader impact requirement is also a bad idea. A research community as a whole should think about the broader impacts of its research, but requiring it on every grant is overkill, let alone in every paper.
