Summary: Guy uses AI to vibe code an app, and it deletes the prod database. It is only at that point that our protagonist learns that there was only one database shared between preview and prod.
This is the first time I’ve seen these kind of “you don’t know what you don’t know” issues in coding to a degree that I’ve seen them in law, and it’s the same structural issue — when you get down to it, you have to be the kind of expert that the AI promises to replace to be able to use it safely
Replies
-
vibe coding as freeman-of-the-land analogy?
-
AI needs to do more damage to large companies more broadly and quickly for the fucking shareholders that control the boards to even blink.
-
Yes exactly. My own experience, is that these tools are:
Useful and worthwhile for the careful expert who has reliable ways to ensure quality outside the AI tool
Dangerous for the beginner who doesn’t yet have a full understanding
Catastrophic for the layperson who sees it as a shortcut
-
IANAL but I know a thing or two about coding - I imagine it's the same in both fields. You can absolutely have an LLM knock something together, but if it's something that really matters, you want to triple check it and you're absolutely going to find some stuff.
-
And using the genAI will not let you build those skills, either. It’s the deadest of dead ends.
-
This is 100% the issue
it is extremely mentally tiring to be repeatedly told at work “you should use copilot to brainstorm”
-
Somebody really needs to sit these people down in front of the Sorcerer's Apprentice segment of Fantasia and make them watch it on repeat until they get it.
-
If one is that much of an expert, and some of us are, there's little point to AI coding (above existing IDE acceleration). For net new, we can write a deterministic tool that has zero risk in less time than it would take to debug junk, or take off the shelf patterns we've built up over a lifetime.
-
exactly. There’s nothing extant that can replace solid discipline and process.
-
It's not even like having a junior dev working on a thing. It's like having an improv comedian playing a junior dev yes-anding their way through your codebase.
-
The great philosopher Donald Rumsfeld strikes again!
-
Johnny tables dot ai.
-
That’s absolutely the way it works. That’s why it’s useful for researchers looking for new combinations of enzymes or what have you. It’s a limited use case with highly curated training data and the results are still just to aid the human experts, who would have enough sense to toss the bad ideas
