> Confidence in AI output is inversely proportional to one's ability to verify it

I like this / generally agree. The only wrinkle is that - for some tasks - the verification _is_ "run the script, see if it worked, don't care how... just that it did" which is distinctly different from "not only did it do it correctly, it did so in the most direct and performant way possible".

For a _lot_ of what I use LLMs to build, the former is all I need.

And for as long that that runs on your computer, I don't care.

But the problem is that for many people they now believe it's ok to present a 10k line vibe-coded PR that only has been verified against external behavior, and some Senior Engineer needs to review it, in time, under pressure, without too much push-back, and lastly, it's the Senior Engineer that gets paged at 2am because something has fallen over.

Also, those scripts tend to start a life of their own, and because it looks good enough, people don't look at them again.

I recall a bug of someone vibe-coding a cleanup script for folders older than $x (on Windows).

Get the CreationDate, and sort. Delete older than $x. Except CreationDate can be null and null is always smaller than $x.

Oops.