The first thing we should never, ever vibe code: cryptography. Building good crypto requires not only advanced understanding of mathematics, but also a deep knowledge of the history of cryptography and security, and you can not be wrong. You can't let a bug slip in, or hallucinate something. It has to stand up to years of scrutiny by world-class experts before it can be considered secure. And it may still fall to a technique nobody predicted. Not only can vibe-coding not guarantee any of that, it's so flawed that failure is virtually certain.
And there's other things that simply can't be allowed to fail. Flight control systems. Vehicle ECUs. Weapons systems. Nuclear power. Industrial controls for municipal water, power utilities, etc. Even if vibe-coding could approximate these things, nobody ever would (or should) trust it.
So while we're out here finding really fun uses of AI coding, we should also consider that there will be very real barriers to its use. The "old world" of coding will have to stick around. We're going to end up with two very different universes of software, with different companies, different staff, different rules. We may even have to update APIs/ABIs to mark when an input/output comes from an LLM, to prevent one component's hallucinations from "infecting" another system that needs to remain safe and reliable.