Skip to content
Josh Benzel
← Writing

Living in Peak Stupidity

The first unit test ChatGPT ever wrote for me looked beautiful. It also didn’t work. I sent screenshots to my family anyway.

Published
Reading time
9 min
Tags
ai

In late 2022, I was studying computer science and working part-time as a full-stack developer at a small company. We had a respectable backlog of software and a much less respectable collection of tests. I’d spent more hours than I care to admit writing slightly different versions of the same testing sequences, so when I stumbled across ChatGPT, I was pretty interested in the idea of a machine doing some of that work for me. I’d studied machine learning in college, including k-nearest neighbors, linear regression, Naive Bayes, and the basics of neural networks. But this felt different. Most of my programming assignments had involved figuring out the rules myself and then spelling them out to a computer in painful detail. With machine learning, you could give a computer examples and let it figure out some of the rules on its own. That seemed like black magic to me. I remember wondering why I didn’t see it used more often, although, looking back, I probably just didn’t recognize it when I did.

I approached ChatGPT the way years of Googling had taught me to approach an empty search box: I typed a few keywords and hoped for the best. Except this thing answered in complete sentences. I tried asking it for a simple little Python script, something specific enough that I doubted anyone had asked the exact same question before. The result was rough, but it was there. Then I handed it one of my unit tests, and by my generous estimate it did about eighty percent of the work. By the test runner’s estimate, it did zero percent. Somehow, that didn't bother me in the slightest, because I was staring at something that felt genuinely impossible. Here was a computer taking a request written in plain English and producing code I hadn't explicitly taught it to write. Not a search result or a snippet copied from a forum, but an answer to my question. I couldn't stop thinking about what that meant, or what it might mean a few years down the road.

I was convinced I'd just witnessed the beginning of something world-changing. And I don’t mean that in the way people say every new phone is going to change everything. I mean I genuinely thought I'd stumbled into a glimpse of the future, something so extraordinary that I couldn't understand why the entire world wasn't talking about it. I told my friends. I told my family. I told classmates who had actually studied machine learning with me, as though I were personally delivering news of a scientific breakthrough. I showed people screenshots of broken tests as if I’d discovered fire. They probably saw a chatbot that was pretty good at guessing Python. I saw the beginning of a technological revolution, and I was absolutely certain they just weren't seeing what I was seeing.

Looking back, I can place myself pretty neatly on that familiar Dunning-Kruger-style curve, right at what people often call the peak of stupidity. The idea is that a little knowledge can give you an enormous amount of confidence, long before you've learned enough to appreciate what you don't know. The popular curve is more of an illustration than a precise scientific graph, but it's a painfully good picture of how I felt. I wasn't necessarily wrong about the potential. I was just making sweeping predictions about the future of computing after roughly a week of playing with a chatbot that couldn't even pass my unit tests. My confidence was running a little ahead of my experience.

Me, week one with ChatGPT (late 2022)

The familiar confidence-versus-experience sketch. A metaphor for my experience, not a literal measurement of the effect.

Eventually, reality caught up with the excitement. ChatGPT could produce plenty of code that looked right, but getting it to actually work was another story. I did find a job it could handle pretty reliably: I’d give it the structure of a test, let it write the repetitive setup, and then figure out what needed testing and write the assertions myself. My great technological revolution had become a reasonably good boilerplate generator. Still, as someone who hates typing the same code twice, I was happy enough with that. I stopped expecting it to do everything and started using it for the few things it did well.

Me, after enough broken tests

Same curve, a few broken tests later.

In early 2025, I joined an AI and machine learning group at General Dynamics Information Technology, surrounded by engineers who knew far more about the field than I did. You might think that would have transformed the way I wrote software, but my day-to-day habits barely changed. I’d occasionally ask ChatGPT to remind me of a command or explain a concept, then go right back to writing code the way I always had. Around that time, a friend at work told me he hardly wrote code by hand anymore. He was using Claude Code, describing what he wanted and working with whatever it produced. I remember asking him to clarify because surely hardly didn’t mean hardly. It did. I thought that sounded a little ridiculous and went back to my desk to write code.

It took a different coworker to make it click. He pulled a couple of us aside to show some performance work he’d been doing, complete with really good Matplotlib figures and a bit of HTML to present the results. This wasn’t someone who would happily accept a sloppy result just because a model had produced it. I knew how particular he was about his work, and seeing what he’d made carried more weight than hearing someone talk about how much code they no longer had to type. That was the demonstration that finally convinced me to pay for a subscription.

One weekend in July 2025, I bought the Pro plan and spent the next couple of days doing what I suppose I have to admit was vibe coding. I still have trouble explaining just how unbelievable that weekend felt. I would describe what I wanted to build, watch Claude Code start putting it together, ask it to change something, and then watch it go off and make those changes too. I wasn't just getting a few lines of boilerplate anymore. I was watching a machine take a rough idea, work through the pieces, and turn it into something real while I sat there trying to wrap my head around what I was seeing. Every time I thought I'd found the limit of what I could reasonably ask it to do, I would try something else, and there it went again. I kept waiting for the trick to wear off, for the moment when I'd have to go back to doing everything the old way. By the end of the weekend, I felt like I’d skipped several years into the future.

It brought back exactly the feeling I'd had in 2022, except this time the thing I was so excited about wasn't a broken unit test that hinted at what might someday be possible. It was software being built right in front of me. I remembered those screenshots I'd sent my family and how convinced I'd been that the world was about to change. If you'd shown the 2022 version of me what I was doing that weekend, he'd have thought it was science fiction. And somehow, even after all that early excitement, the real experience still managed to surprise me. Apparently, there was another mountain of stupidity waiting for me farther down the curve, and I spent that entire weekend happily climbing it.

Me, July 2025

The same curve, with a second peak and an open end.

Since that weekend, I’ve had some interesting conversations with other engineers about what all of this means. Some went through almost the same progression I did, from skepticism to excitement. But for a few of them, the excitement eventually turned into something closer to grief, and I can understand why. There’s a special kind of satisfaction in spending three days stuck on a problem before finally realizing what you’d been missing. For a few minutes, you feel like the smartest person alive, even if you’re also the idiot who caused the problem on Monday. If an AI tool gives you the answer on Tuesday morning, you get two days of your life back. But you also miss those few minutes when everything finally clicks.

I also think about the engineers who spent years building skills one evening at a time. Imagine watching someone turn out, in an afternoon, work that used to take you countless nights to learn how to do. That doesn’t mean experience suddenly has no value. Knowing whether the result is correct, maintainable, or even worth building still matters enormously. But the finished product doesn’t always make it obvious who has that experience and who just knows how to ask for something convincing. I can be excited about what these tools make possible for younger engineers while also understanding why someone who spent years mastering a craft might feel like something has been taken from them.

What I find especially strange is how people talk about AI outside of those conversations. A few years ago, if someone told me it was useless, I could usually assume they hadn’t had much reason to try it. Now I’ll hear people dismiss it in one breath and use ChatGPT in the next to plan a trip, work through a difficult decision, or generate a ridiculous picture of themselves. We’ve even made hallucinate a normal word to use when talking about software. Try explaining that to a computer science student in 2022. It’s perfectly reasonable to find something useful and still dislike what it’s doing to your profession, but I’m more puzzled by the gap between what people say these systems can do and how much they trust them when they actually need help.

There’s a pretty remarkable example of this in the news as I write. On November 20, 2026, a mathematician at Anthropic announced that he and Claude had found what they believe is a proof of the Riemann hypothesis, one of the most famous unsolved problems in mathematics. The claim still needs independent scrutiny, so I'm not calling it settled. I keep coming back to what that would mean. A problem that has resisted mathematicians for more than a century and a half may have just been cracked with the help of AI. Is that the moment that makes people stop and reconsider what we're dealing with? I have my doubts. If the proof holds up, I suspect we'll barely have time to absorb it before somebody reminds us that these systems still make up facts and occasionally fail at basic arithmetic. Both things can be true, of course. I just wonder how much black magic has to become ordinary before we admit how strange it is.

There’s also something a little funny about my own part in this: I work in AI, but when someone asks what I do for a living, I usually say I'm a systems engineer or software engineer. Both are accurate, and I'm not hiding some secret disagreement with the work I do. I've just learned that mentioning AI can turn a perfectly normal introduction into an unsolicited debate about everything that's wrong with the world. Hating AI, or at least being loudly unimpressed by it, seems pretty popular right now. Plenty of the criticisms are fair, but I don't always feel like defending an entire field to someone who has already decided it's more of a nuisance than a masterpiece. So I leave those two letters out, and we can all get on with our day. It's a strange change for the same person who once practically chased people down to show them a chatbot writing Python.

So where does that leave me on the chart? I keep wondering whether we're watching the beginnings of another industrial revolution and collectively treating it like an annoying new feature we can't figure out how to turn off. Maybe the technology has become so easy to use that we've lost sight of just how remarkable it is. Maybe we’re so caught up in what it might take away that we’re overlooking what it might make possible. Or maybe this is exactly what someone standing at the top of the Dunning-Kruger curve would say. After all, I was convinced I'd seen the future back in 2022, then spent years underestimating what was happening right in front of me. Who's to say I haven't mistaken another spectacular view for a clear understanding of the landscape?

I can’t tell whether the rest of the world is missing the mountain, or whether I’ve simply climbed the peak of stupidity again. Either way, I’m pretty happy up here.

ESC
↑↓ navigate · ↵ selectJosh Benzel