Standardizing next-generation narrow precision data formats for AI

Article note: Interesting. I love how after decades of "Doubles required for srs bsns" (even though it's rarely the case that doubles will save you if your operation stack-up breaks on singles) we're getting all these ultra-low-precision formats for approximate bullshitting. Curious that everything except E5M2 has no defined infinities. I think it does cover pretty much all the sane uses of bits in the narrow sizes.