The article's 0.65 vs 0.66 float64 example doesn't indicate much since neither 0.64 nor 0.65 have a terminating representation in base 2...