True, I had the same feeling. The article does go off 256K elements in a bloom filter of 2M. After 1M elements, using 2 bits actually increases false positive rate, but at that point the false positive rate is higher than 50% already.