(Author’s Note: I drafted this blog post back in June 2025, and even then, I think the topics were slightly out of date. I suspect I didn’t post it because I wanted to edit it and never did.
Well, since then, everything has changed in the capabilities of AI. I have another blog post coming soon on that topic, but I thought I would share this unedited anyways just to contrast my experience.)
Since I setup my private LLM chatbot, I actually have been using LLMs a lot more. Rather than running it off my personal server, I have been using a family of online services instead, and it has been solidly hit or miss. Here’s what has been working for me, with the notable caveat that the technology is rapidly evolving.
Coding in my editor
I can’t say I’m an early adopter of LLMs for coding, but it was certainly the first major use case that I started using extensively, and I really do like it a lot. It started out as just being better auto-complete, and it has evolved to do more and more work.
I’m firmly on the side of keeping my hands on the wheel. More recent coding trends have been to provide deeper access for LLMs like Copilot to have actual access to APIs and execution. I haven’t played around with those as much, but I can see the benefit.
There are two cases where I really like using LLMs.
First, taking care of boilerplate. I actually need to write logic, but LLMs can get most of the scaffolding correct around it. Especially when I’m writing one-off scripts, the quality is actually better than me trying to just get it done.
Second, writing tests, doing refactors, or doing anything that sounds too boring for me to do myself. I’m not guaranteeing that it will do it right, but it is a great way for me to get started on something that I wouldn’t otherwise.
Sysadmin tasks
I got in a bad spot with my personal server i.e. the server that serves this website. I was on an outdated version of Ubuntu that was so old that Trump hadn’t even been elected the first time. The problem was that there wasn’t even an upgrade path available. Instead, I would need to do a clean installation.
To start, I got some bad advice from ChatGPT: it suggested an upgrade path that wouldn’t work in practice, and when I suggested a way to do the upgrade that I was relatively confident would work, it said I should do something else.
However, once I coerced it into my way by insisting on a certain path, it was much more helpful addressing a variety of other tasks along the way of doing the server migration. I ended up going back and forth and copy-pasting commands to do things. There was notably one case where it even suggested additional things I hadn’t considered in doing my MySQL migration.
There was one place where I was getting error messages where I noticed myself leaning on ChatGPT too much. Rather than using my actual knowledge to read the error message and address it, I just kept giving the error message to ChatGPT and letting it try things.
Overall, I considered that attempt a success. I have gotten stumped on many sysadmin tasks in the past, and with ChatGPT, I was constantly making progress and got it done quickly.
However, if I didn’t generally know what I was doing with sysadmin, I suspect that I could have been led astray by listening to it too much. It made some bad, high-level suggestions that I needed to use my prior experience to avoid.
I actually had the flip side of that while trying to do some WordPress customizations. I’m much less familiar with WordPress, and while trying to get some additional functionality, it took me down an entirely wrong path, and it took some analysis from my side to realize how I should have done it completely differently to start.
Internet Search
Early on when Google added the AI overview at the top, I thought the quality of the response was pretty poor. And even recently, I have had some issues where it was misleading. For example, I was having difficulty trying to configure the settings for an app, and the overview definitely hallucinated some options.
Otherwise, the AI overview has become better at getting quick responses to searches. My negativity, however, is that this is only true due to the strange way that SEO has tweaked the content of the internet. People have complained that Google searches have gotten worse over time, and my intuition is that the Google algorithm isn’t the problem: the problem is the content on the internet.
It turns out that a lot of the internet is filled with clickbait and generating ad revenue. A simple question like “where do avocados grow?” ends up being a 1000 word article with the answer somewhere around 2/3s of the way down because one way to improve your ranking is to have sufficient content on the page, so people pad articles for no other reason. And there are ads everywhere.
So in that sense, the AI overview is a solution to the problem that Google exacerbated in the first place.
Final thoughts
I really haven’t picked up multimodal use cases yet. I haven’t used LLMs much for either audio or for pictures or videos.