The `nan` thing is actually because `float('nan')` can (and does) return a new object (well, with a different ID) each time. So it's like going `d[object()] = '

The `nan` thing is actually because `float('nan')` can (and does) return a new object (well, with a different ID) each time. So it's like going `d[object()] = 'a'; d[object()] = 'b'` and not retaining the objects for lookup. > Personally I'd just make any comparison involving different types an error As you point out, this is what Python 3 did for `None`. But it's been idiomatic to have 0 = False when languages didn't have `False`; so `bool(0) == False` makes sense, but also `sum(boolean for boolean in sequence)` works. So you could argue that for `True` and `False`, meaningful coercion is possible, and in reverse, if zero is false-y and everything else is truth-y (like in C), then that also makes sense. Imagine if you had to cast every number to a boolean explicitly! So I guess having `d[0] == d[False]` is the lesser of evils.