In an earlier post I mapped my AI journey onto the five stages of grief, and the first one was denial. I said AI writing production-level code was overhyped, autocomplete pretending to be a developer. I want to come back to that claim now and be honest about how much of it was actual analysis, and how much was pride wearing analysis as a disguise.
The pride part
Here’s the uncomfortable truth: I couldn’t understand how people were pulling this off, and instead of getting curious, I got dismissive. If it hadn’t been someone close to me — someone I couldn’t wave off as clueless or lying — I probably would have written the whole thing off as exaggeration or something too complicated for a layperson to have actually gotten right.
And even after seeing it work, I still wouldn’t let myself use it. I treated it as unprofessional, beneath the standard I held for myself. Looking back, that wasn’t caution. That was a credential protecting itself. I should have at least tried it on something low-stakes, just to see. I didn’t, and that gap — refusing to even experiment — is the part I can’t fully defend.
Where the skepticism wasn't wrong
But I don’t think the whole thing was ego. Some of it held up, and it’s worth naming plainly instead of pretending I was simply proven wrong across the board.
The hype was real. AI companies had every incentive to oversell what these tools could do, and a lot of the loudest claims came from exactly that kind of marketing.
The loudest fans often weren’t the people who’d know. Plenty of the “AI writes better code than any developer” claims were coming from people who don’t actually write software for a living, and couldn’t tell polished-looking code from code that’s actually solid, secure, and maintainable.
Not all models are equal. Lumping every LLM together was always a mistake. A frontier model and a free, lightweight one are not the same category of tool, and a lot of early disappointment — mine included — came from testing the wrong tier and judging the whole technology by it.
The cost structure is genuinely tricky. Between subscriptions, usage limits, and the cost of running anything serious yourself, it’s fair to be wary of a pricing model that looks a little too good until you’re a few months in and hooked.
Put together, that’s a reasonable case for suspicion. It’s not nothing.
What I missed anyway
Here’s what I didn’t account for, though: none of that skepticism required me to actually stay away from it. I could have held every one of those doubts and still opened the tool up on a Saturday afternoon with nothing on the line, just to see for myself.
That’s the real lesson, and it’s less about AI specifically than about how I handle anything genuinely new. You can’t fully evaluate something new from the outside. Not by reading about it, not by taking a course, not by listening to other people’s verdicts, good or bad. At some point you have to actually put your hands on it and see what it does in front of you, with your own problem, on your own terms.
I skipped that step because trying it felt like conceding something. In hindsight, trying it wouldn’t have cost me anything. Not trying it cost me months of avoidable certainty about something I’d never actually tested.
Where that leaves the skepticism
So, was I wrong? Partly. The suspicion about hype, mismatched claims, and pricing traps wasn’t unreasonable — some of it still holds. But treating that suspicion as a reason to never even experiment was the actual mistake, and it was pride doing the deciding, not analysis.
Is there something you’ve been skeptical of, for reasons that felt sound, but that you’ve also never actually tried?
