Interesting paper and thanks for the excellent Twitter summary thread! Regarding the decision boundaries being sensitive to label names and example order, I'd be curious what would happen if you tried inserting K different permutations of the N examples (perhaps with different

→ View original post on X — @jeffdean