Assume one who watched a few apps be vibe coded in days.. It is impressive to see. Still risky as all get out mind you!
I have never been a fan of AI, but I figured I was doing myself a disservice by not learning to use it, so I set up a couple of local models to play with and learn on. I figure by doing this I avoid the worst problems of privacy, and also don't give any of the big AI companies any money.
I have an instance of Laguna S 2.1 (118B Q4_K_M) running on CPU the server. It's too big to fit in GPU memory, but it is surprisingly capable for coding, so I keep it running slowly on the server CPU, which I can get away with only because of the 8x channels of RAM giving me much greater RAM bandwidth than on a consumer system. (and LLM performance is all about RAM bandwidth)
I have on a few occasions brainstormed ideas and come up with requirements and a project plan on my faster model running on GPU (Gemma4:26B Q8_0) and when I think I get it to a decent enough place then feed the plan to the Laguna model, have it refine the plan, review it one last time and then tell it to write code.
I've only done this for some python scripting for server maintenance thus far, but I have been quite surprised at how well it often "just works", especially since I am not a programmer and don't really know what I am doing.
Occasionally there is some troubleshooting when the code is written, and that's where it gets a bit time consuming.
I need to figure out how to give the models a working folder so I don't have to constantly copy and paste things back and forth between the chat and my running environment. That is the big time suck right now. Because running the model on CPU is slow, the cycle of testing, modifying and testing again takes a lot of time. If I could set up a working folder, I could just let it do it's thing in the background more easily. I played with this a little bit but ran into some troubles getting it to work.
As far as risks go, I think the key is to be cautious about where you use it, and what the risks in those applications are, and not just blindly trust the output, instead being smart about it.
Right now I have it writing a manual sensor-based fan control loop for one of my servers, and it seems to be working pretty well.