At this stage, we're all learning. (not that we ever stop), Vibe coding can be great for learning if the dev is taking time to understand what's being done. But I fear too many are taking the easy path and just writing a prompt and shipping. That's not safe for any environment, let alone production. Some time in the future it may be better. We'll have the proper guide rails for AI, the proper testing paths, and overall reviews. Right now isn't that time. If you want stable, efficient code, which you definitely do for Production and kernel maintenance, AI isn't ready. It's not any more ready than self-driving cars. Can they do it? Sure, would you trust them in every situation? probably not. That "Probably, not" is what gets you killed. Or Panics the kernel, or crashes the DB....etc.
Linus is a smart guy, I might not agree with everything he implements, but for the Linux kernel, I can say I never felt like he went in the wrong direction.
I'm sure this discussion will continue. It's not a once-and-done. But as newer and better coding systems come online, they'll need to be tested and verified, and eventually we may get something that passes the test. It's already miles ahead of where we were only 10 years ago. I can't predict how we'll be next year.