Comment by applicative

What is most charming about this benchmark is that by now they do train on past results, and thus commit all the same errors over and over.