No I think the original comment is correct. You need to be able to evaluate the result.

The goal of software is not to compile and run, it's to solve a problem for the user when running. If you don't know enough about the problem to be able to assess the output of the program then you can't trust that it's actually doing the right thing.

I've seen code generated by LLMs that runs, does something, and that the something would be plausible to a layperson, but that I know from domain knowledge and knowing the gotchas involved in programming that it's not actually correct.

I think the comparison with translation is a very apt one.