[oxag3n]

What if that AI was trained to hide implementation as much as possible? E.g. by making the client as thin as possible, using OAuth for authentication, following some robust template, would it be better?

Avoiding such low-hanging-fruit disasters is doable and major players have incentives to fix it.

This however wouldn't change the goal - replace all software engineers with DIY nephews with LLM in their hands.

Even worse - people start thinking it's infeasible to learn hard things, because it prevents you from moving fast, from one domain to another.

My area of expertise is cryptography-adjacent. Non-trivial fraction of students think that cryptography and security is a dead-end skill as all companies in this area are being replaced with AI. I asked them to implement web-bot-auth as simple as possible, because I know how AI can read specification and follow it.