I am really a supporter of the guideline "getting far with small steps".
However, there are situations and problems which can not be solved that way.

If you learn how to ride a bike, you will not succeed when being to cautious to steer hard enough.
There is no way to learn this by first steering a little bit only, and then increasing this in small steps.

I did a simulation today of a random chaotic system, trying to find system parameters where this system would do what I wanted. I did not succeed first, only after I substantially increased the mutation size of the parameters the system would snap over from the chaotic behavior to the stable behavior.

Can this be generalized somehow?
Well, looks like making small steps is good to find a solution only in certain environments, in other environments it is necessary to make big steps.
This is not as trivial as it may sound in the first place.