[littlestymaar]

> The best way to drive inference cost down right now is to use TPUs

TPUs are cool, but the best leverage remains to reduce your (active) parameters count.