Comment by buildbot

This paper's claim holds - for 4 layer models. Models improve on out of context examples dramatically at larger scales.