Musings on using AI for selfhosting

It took quite some effort to get it running, but I now have Nextcloud, Immich, and Mealie running on my M4 Mac Mini. I also put everything behind Cloudflare Tunnels and connected them to custom subdomains. After some struggling, I even got Nextcloud Office and the High-Performance backend for Talk working on my system. For someone who never did any of this until a couple of weeks ago, I've learned a lot. First I learned Docker, then I learned NGINX Proxy Manager, and then I learned of a better solution in Cloudflare Tunnels after I saw my firewall getting hammered by all kinds of nasties trying to get into my systems. For figuring all this out, I found using AI tools like Claude a positive experience. It regularly took me in the wrong direction, but with common sense I still found this a very accessible way to get into self-hosting, where using just the documentation or forums felt like an impossible task. It allowed me to ask all the stupid/basic questions like: "ok, how do you start a container?", "what does this command mean?", "where do we save this?", "why do we have to do that?" - things that the documentation expects you to know and that feel wrong to ask on a forum. It felt a bit like a student mentor - it really doesn't know everything and makes many mistakes, but together you are getting there. I did test quite a lot of LLMs while trying to set up my self-hosted systems. I first tried ChatGPT, but the free version makes way too many mistakes and likes to skip over steps (maybe the premium version works better, but I was not ready to pay for it). It also is very sure of itself and people-pleasing, which annoyed me to no end. I tried Lumo as a Proton user and also local LLMs like Ollama and LM Studios. I love the privacy aspect of these tools, but the quality of the models you are able to use where severely lacking for this task. Either due to dated information or just not being able to communicate clearly. In the end, I kept coming back to Claude.ai, which felt like the most user-friendly tool for this purpose. It knew enough to be genuinely helpful and was able to clearly communicate this to a noob like me. It would also generate code artifacts I could just copy and paste. It would very clearly write things out, and if I told it to go step by step and wait for terminal responses before jumping to conclusions, it would do that (LLMs like to predict instead of verifying what will happen). And, very important to me, it made the fewest obvious mistakes of the lot and would also correct me if I made mistakes, instead of continuing down a failing path like ChatGPT. Claude.ai in a way felt more Dutch with its directness, while ChatGPT feels very American - always pleasing but also avoiding pointing out the obvious mistakes I would make and wasting a lot of my time. But the most important thing I found with all these tools? Keep asking questions like: "Why do we have to do it this way?", "Are you sure this is the latest version?", "Let's first verify what is actually happening before jumping to conclusions?" Am I an AI convert? No! I'm still very unhappy with the damage it's doing to, for example, the creative sector, but for playing around like this, it has its use.