I'm sorry, but you still have to think

Sad to think that it's possible you put more human effort into this analysis and write up than was put into the subject itself.

If you want a reliable system you should know when to fail

I think the main problem is that most people pushing for heavy AI delegation either

1. Don’t really care about system reliability or
2. Don’t understand the difference between code and system, and assume that a system must be reliable if some automated checks pass

In my experience reliability is not something achieved by a repeatable method but by keeping the system simple so it stays easy to reason about… and then reasoning a lot about it. Both things the clankers actively conspire against.

Most of the common failures modes of llms include:

1. Swallowing errors