Hard to comment without seeing more concrete examples of his approach. But it makes me wonder whether processors are now so fast that, for a large class of numerical problems, the default should be to give up calculational speed in return for eliminating some of the trickiness involved with floating point.