I read that not as 'step algorithms' but as 'algorithms requiring N^(O(1)) steps'; in other words - polynomial time algorithms.
One way to think of this is in terms of galactic algorithms, which are algorithms that incrementally improve the exponent in a polynomial time algorithm, but have constant terms so huge that they're effectively useless, practically. He's suggesting that there could exist a polynomial time algorithm for an NP problem which is similarly useless - but with the additional caveat that we won't even be able to prove theorems about it!
One way to think of this is in terms of galactic algorithms, which are algorithms that incrementally improve the exponent in a polynomial time algorithm, but have constant terms so huge that they're effectively useless, practically. He's suggesting that there could exist a polynomial time algorithm for an NP problem which is similarly useless - but with the additional caveat that we won't even be able to prove theorems about it!