How randomization helps is by making it much easier to design algorithms. E.g. Verifying a solution is cheap, so proving your random choice is in some class of good choices and making an algorithm that turns that into a solution is still an interesting approach and opens up solutions to things we cant yet solve.
Derandomizing as presented apparently means proving for a few random but careful choices you're in the funnel towards the right solution, or (hopefully) can detect otherwise, and a few transformations or reasonably performant steps produce your solution.
I must have put too much emphasis on PRNG or on maybe there's a more intelligent way to sample - you can ignore those comments. I just think it's framed poorly to make it sound more counterintuitive than it is.
Oh, that's cool, do you have a reference for that?
The OP article has such a reference, but theirs is paywalled, and perhaps you missed it, so you may wish to see this no paywall link to the paper:
Hardness vs. Randomness by Noam Nisan & Avi Wigderson https://www.math.ias.edu/~avi/PUBLICATIONS/MYPAPERS/NOAM/HAR...
Instead we could say "we know shortcuts that work for most values in a population, but we don't know how to figure out which values without expensive computation" and that's not very mystical or paradoxical.
And using random sampling is now the obvious way to deal with a situation. But it's not random doing the work, that's an implementation detail of how you do the sampling.
It very well can be that there isn't another obvious way to iterate through the sample space that isn't a disaster. Maybe bad values are clumped together so when you retry on failure you retry a lot. But if that's the case, there might also be a better way to sample than random - if bad values are clumped then there may be an intelligent way to sample such that you hit two bad values in a row less often than random.
My question is if that's what's being referred to as 'derandomizing' - taking one of these algorithms that uses random sampling and sample more intelligently. Or if they instead mean using what they learned from the (so-called) probabilistic algorithm to go back to a more traditional form.