I hope they're going even farther and considering minifloats. 8-bit integers are easy but limiting and unexciting. 8-bit floats are exactly what some machine le

I hope they're going even farther and considering minifloats. 8-bit integers are easy but limiting and unexciting. 8-bit floats are exactly what some machine learning needs. If you're updating your weights enough, precision can just be wasted computation. A 512-bit-wide computation could be updating 64 very coarse-grained weights at a time.