I've been doing this long enough to remember when it was bad. Not imperfect: confidently, fluently wrong, in a way that sounded more certain than it had any right to. I've watched it go from that to useful, and watched the rules around using it change more than once.
Confidently wrong, and still useful
My first real use was in college, right when ChatGPT showed up: physics explanations for concepts the textbook wasn't landing. It was good at that even when it was wrong about other things, taking something dense and making it click in a way a professor rushing through slides couldn't. I wasn't taking coding classes then. Then I stopped thinking about it for a while; the AI-image wave came and went without me paying much attention.
From explanations to my own writing
When I came back, it started with explanations again, then rewording: my writing, my messages, my emails. It was better at that than I expected. Using something to understand a concept and using it to sound like a better version of yourself in an email are different things, and I went through both before I touched code with it.
The vibe-coding attempt that didn't work
Eventually I tried asking for code and trusting it. It didn't go well. The code broke, I didn't always know what it was doing, and each round of fixes introduced something new. Rewording my writing worked; code I could trust without checking did not, not then. It got better as more people built with it and fed their code back into the models, but the gap between "good at language" and "good at code I can trust" was real and I felt it.
Which model to trust for what
At some point I stopped using ChatGPT for everything. Claude was better for code. Perplexity was better for research, with actual citations instead of confident prose. I route by what each is good at: stock and company research to Perplexity, rewriting emails to ChatGPT, vibe coding and technical questions to Claude. Deciding which tool to trust for which question is its own skill, and one I use every day now.
Watching the rules change
The free tiers went from generous to bad to good again. API access that felt unlimited started getting metered, and I don't think that pattern stops: energy and data-center costs aren't going down, and pricing that looks generous early tightens once a tool proves useful. AI isn't going anywhere, but a lot of what's happening now has bubble characteristics, and access getting restricted in real time is one of the signs.
The vibe-coding economy I go back and forth on
I've watched people with no technical background vibe code an app, ship it, and make money from users who don't know or care that the database isn't encrypted properly. The easy developer reaction is to point at the security gaps and say it doesn't count. But they're shipping and making money, and maybe what they have more of than me is deployment, marketing, and the nerve to put something in front of paying customers.
Senior engineers at the big AI companies say vibe coding will produce a wave of technical debt nobody wants to deal with later. I don't think they're wrong, or fully right. You can vibe code something and go back and verify what came out; a lot of people just skip the second part. I go back and forth on this one more than anything else on this page.
What I think
None of this is abstract for me. The agents I built for work, this portfolio, the SQL scripts I untangled with an agent's help: that's the proof, built over years of using these tools rather than reading about them. AI is roughly what science fiction said was coming, and it showed up faster and stranger than the movies guessed. We're lucky and unlucky in about equal measure to be living through the part where it's still figuring itself out.