What's wrong with training in full precision? There's a huge market for efficient inference in robotics, devices, and autonomous vehicles.