It still seems odd to me that there are many different NaN values in floating point arithmetic. Those bits could have improved precision!