Before reading the article I was expecting using 1-bit instead of bfloats, and using logical operators instead of arithmetic.