Early speed optimizations aren’t premature

I wouldn't say using vectorized operations as opposed to for loops is such a good example. You're simply using the right tool for the job. Is using a fork to eat soup suboptimal? Is switching to a spoon optimizing the way you eat?

But what about using a dict or a list or tuple? That kind of decision can make a performance difference down the road, and there are trade offs in ease of development.

Anyway, the whole thing feels a little contrived. Who would ever argue that you shouldn't use numpy from the get-go? Who is the author arguing against?

This article would be more convincing if the NumPy version were actually vectorized. Replace the for-loop with

    result = (arr - min_val) / (max_val - min_val)

That should often run faster and removes the allocation and for-loop, so it's easier to read.

Like most comtrarian articles, it is only meaningful if you haven't read the original, which it isn't actually contradicting.

https://softwareengineering.stackexchange.com/questions/8008...

Of course. Just engage into mature speed optimizations from the beginning, as everyone else.