From the avrix (sp?) paper that they released. I was skeptical of this, but evidently its completely open-sourced and thus verifiable.

People on twitter have said most of this "breakthru" was common sense tuning optimizations that resulted in less memory use, with a slight uptick in error rate, but the optimizations were scaled so that the increase in error rates didn't spike significantly.

Basically: Anyone with a constrained hardware budget would've eventually taken this approach.