03/09/26 02:29
In theory, I'm not a fan of generative AI. I dislike the way it writes things. I dislike that AI companies steal content and scrape websites. I dislike that it has caused people to develop psychological issues and even commit suicide or homicide. I'm fed up with every other article on Hacker News being about it. I could go on.
I say in theory because... I've been using chatGPT heavily the last few days.
Not to publish anything - I'm not going to subject anyone to its output. But I used it for asking questions and figuring things out. I'd previously dabbled with LLMs and image generators, when they were new and exciting. I tried a bit of vibe-coding, and even had that code hosted on this site (though clearly marked as LLM-generated). But until recently, I was convinced they were something of a dead end, and then I started disliking the whole field, and how society at large is using it. Improvements were incremental, I just couldn't ge any serious/sustained use out of them.
Well at some point, those incremental improvements added up. ChatGPT now searches the web pretty aggressively when answering questions, so it's less likely to get things completely wrong straight away. And when I couldn't figure out something with a traditional search engine when revamping this site last week, I figured 'eh, I'll try an LLM, it's not like I'm not subjecting anyone but myself to slop'.
The LLM gave an actually-useful answer, with links to sources so I could double-check its output. And then it was all too easy to start asking follow-up questions. This is where things started going wrong. I felt like I was learning, I felt like I was being productive.
And probably, to a certain extent, I have been. By dumping questions into ChatGPT, I migrated this site, set up a Matrix homeserver, set up my entire logging stack. Ironically, it helped me set up my AI tarpit. It felt easier and faster than reading documentation and figuring things out for myself. I haven't gotten into a flow state like the last few days in a long time. To be clear, I still 'did the work' myself - I'd ask questions on how to do something, or ask about an error message, then I'd use its answers to implement and fix things. But I'm sure I've learned less than I would have if I'd done everything myself.
I think the first thing that contributed to it's addictiveness was that I'd tuned the settings/prompt to reduce sycophancy. This made reading its output much more tolerable. Second, they're now accurate/useful enough - at least for the technical stuff I was asking about - that they get you about 80-90% of the way there when deploying/troubleshooting stuff. And third, an LLM is always available, and never gets tired of you. It will never tell you to stop asking it the same dumb questions in different ways. It will never tell you to do your own research or think for yourself. It is easy, in the way doomscrolling is easier than doing something creative.
I'm quite conflicted. I've gotten actual use out of something I'm not supposed to like. I think maybe my problems with it are just sharing slop, and the ethics of LLMs-as-a-service. In a utopia where datacenters don't cause environmental issues, and everyone just uses LLMs for research and never posted slop, maybe I would have no problems with them.
Although, there is one more problem this utopia would still have. You are what you eat. And after bingeing LLM content the last few days, I've started noticing it's affecting how I write. I had to re-write this blog post because I wanted to start putting in headings and bullet points. I'd like to think I'd never suffer any serious problems from continued use, but y'know, that's exactly what any addict says.
I've also noticed myself jumping straight to LLMs for every question and idea that pops into my head, rather than searching for human-written content, or just thinking for myself for 30 seconds first. That's definitely got to have a detrimental effect. For one, I can't even think of any examples of stuff I've asked it about, since I tend to delete the chat log afterwards. It's like doomscrolling and then not being able to actualy remember one thing you've seen.
So, I need to quit it with the LLMs, and go back to searching and reading for myself, or asking people. I'll update on how that goes in a future post.