Thanks! The corresponding graphs in the paper show that it's OK at two-digit operations (except multiplication), but it doesn't generalize to bigger numbers. This would seem to support LeCun's statement that there's a lot of over-hyping going on.