I've been doing this long enough to remember when it was bad. Not "imperfect" bad — confidently, fluently wrong, in a way that sounded more certain than it had any right to. I've watched it go from that to something genuinely useful, then watched the rules around using it change more than once. This is just the long version of that, written down.
Confidently wrong, and still useful
The first real use I got out of it was in college, right when ChatGPT showed up. I used it for physics — explanations for concepts I wasn't getting from the textbook. It was good at that even when it was wrong about other things; it could take something dense and make it click in a way a professor rushing through slides couldn't always manage. I wasn't taking coding classes at the time, so that's mostly what I used it for. Then I stopped thinking about it for a while. The AI-generated-image wave came and went in the background, and I wasn't paying much attention.
From explanations to my own writing
When I came back to it, it started with asking for explanations again, then it became rewording — my own writing, my messages, my emails. It was genuinely good at that, better than I expected. There's a real difference between using something to understand a concept and using it to sound like a better version of yourself in an email, and I went through both phases before I touched code with it at all.
The vibe-coding attempt that didn't work
Eventually I tried actually vibe coding — just asking for code and trusting it. It didn't go well. The code broke. I didn't always know what it was doing. Each round of fixes took a while and tended to introduce something new instead of just fixing the old thing. Rewording my writing worked. Generating code I could trust without checking it did not, not yet, at that point. It got better over time, the more people built with it and fed their own code back into these models, but that gap between "good at language" and "good at code I can actually trust" was real, and I felt it directly.
Learning which model to trust for what
At some point I stopped just using ChatGPT for everything and started learning the actual differences between models. Claude turned out to be better for code. Perplexity turned out to be better for research — actual citations, not just confident prose. I started routing tasks by what each model was actually good at: stock and company research to Perplexity, rewriting emails and messages to ChatGPT, vibe coding and technical questions to Claude. That's its own skill, deciding which tool to trust for which question, and it's one I use every day now without thinking about it.
Watching the rules change
Somewhere in there, the free tiers went from generous to bad to good again. API access that used to feel unlimited started getting metered. I don't think that pattern stops — energy and data-center costs aren't going down, and pricing that looks generous early tends to tighten once a tool proves itself useful. I don't think AI is going anywhere. I do think a lot of what's happening right now has bubble characteristics, and watching access get restricted in real time is one of the signs.
The vibe-coding economy I still go back and forth on
At the same time, I've watched people with zero technical background vibe code an app, ship it, and start making money from people who don't know or care that the database isn't encrypted properly or that the architecture is held together with duct tape. As a developer, the easy reaction is to point at the security gaps and the spaghetti and say it doesn't count. But they're shipping, and making money, and I'm not always sure that reaction is the right one — maybe what they actually have more of than me is deployment, marketing, and the nerve to put something in front of paying customers.
I've also heard senior engineers at the biggest AI companies say vibe coding is going to produce a wave of technical debt nobody wants to deal with later. I don't think they're wrong. I also don't think they're fully right. My own experience tells me it's possible to vibe code something and go back and verify the quality of what came out — those aren't mutually exclusive, even if a lot of people skip the second part. I'm not settled on this one. I go back and forth on it more than most of what's on this page.
What I actually think
None of this is abstract for me. The agents I built for work, this portfolio itself, the SQL scripts I was able to untangle with an agent's help — those are the proof of concept, built over years of actually using these tools, not reading about them. I've watched API access go from unlimited to metered, watched companies compete in directions that have nothing to do with chatbots, and watched layoffs get attributed to the same technology I use every day to do my job better.
AI isn't going anywhere. It's basically what science fiction always said was coming, and it showed up faster and stranger than most of those movies guessed. I think we're lucky and unlucky in roughly equal measure to be the ones living through the part where it's still figuring itself out.