Funny enough, with LLMs this trade off may well have flipped. For simple tasks like given a string like X, format it like Y, they work amazingly well.

Absolutely. Even back then, trying to script it would sometimes be fine. I've had plenty of situations where I'd try to write a script, give it 30 minutes then say "fuck it, I'm hand-fixing it." The same probably applies to LLMs, you just need to set a cut-off point before it stops being worthwhile.

Sure, but I have to obtain my dopamine somehow.

I had a similar one where a billing mistake had happened in a system and the other devs were trying to find out how to determine all of the bad values. I asked if they had simply looked at the top N that could fit on their screen sorted by value. The incorrect values were the obvious outliers. Only about 4 of them. They had been blocked on identifying them for about an hour.