Sure, but you have to have the bad attempts to learn things that don't work or we won't end up with the good ones. We fall into a local maximum instead of exploring farther and potentially going higher. Every dead end is another bit of the map explored when it comes to creating expressive languages.

Now, is there an argument we'd be better off doing that exploration in a space designed to make DSLs easier than in building entirely new languages? That I dunno but it is worth considering.