A few weeks ago I caught myself doing something I would’ve made fun of two years ago. I was sitting with a half-broken ADO.NET function, and instead of tracing the bug myself, I pasted the stack trace into a chat window and waited. Ten seconds. Fixed. I didn’t feel clever. I felt slightly hollow, like I’d skipped a workout and the scale still said the same number.
That feeling is the actual story here. Not the acronyms.
Everyone’s throwing around AI, AGI, ASI like they’re stops on the same train line — as if we’re just a few stations away from Skynet, or salvation, depending on which newsletter you read that morning. But sitting inside the actual work, day after day, the experience is smaller and stranger than the headlines suggest. It’s not a singularity. It’s a slow renegotiation of what it means to know how to do something.
The Three Letters Nobody Agrees On
AI is what we have now — the autocomplete that finishes your SQL query, the model that writes your React component before you’ve decided on the prop names. It’s useful the way a very well-read intern is useful. Fast, tireless, occasionally confidently wrong in a way that costs you an hour of debugging you didn’t budget for.
AGI is the promise. General intelligence — a system that reasons across domains the way a person does, not just the way a very large autocomplete does. Nobody agrees on when it arrives, or if it already has in some narrow sense and we’re just haggling over the definition because the definition is the last thing we control.
ASI is the thing people whisper about at 1 a.m. on X, half excited, half terrified. Superintelligence. The point past which we’re not the smartest thing in the room anymore, in any room.
I used to think the gap between these three was mostly technical. More compute, more data, more clever architecture. Now I think the gap is mostly psychological. We’re not waiting for the model to get smarter. We’re waiting to see if we can tolerate what it means when it does.
What “Vibe Coding” Actually Feels Like at 11 PM
There’s a phrase going around — vibe coding. You describe what you want in plain English, the model writes it, you nudge it, it rewrites, and somewhere in that loop a working app appears. People say it like it’s a party trick. It isn’t. It’s closer to negotiating with someone who’s extremely competent and mildly untrustworthy.
I built a small internal tool last month almost entirely this way. Prompted the schema, prompted the API, prompted the frontend. It worked in an afternoon — a task that would’ve cost me three days a year ago. And somewhere around hour two, I noticed I’d stopped reading the code line by line. I was just scanning for shape. For vibes, honestly. Does this look right. Does this feel like it’ll hold.
That’s the part nobody warns you about. It’s not that the AI replaces your skill. It’s that it replaces your attention. And attention, it turns out, was doing more load-bearing work than we gave it credit for. The bugs that used to announce themselves while you typed — the ones your fingers caught before your brain did — now hide two layers deeper, wrapped in code you never fully read.
Become a Medium member
I’m not saying this to be dramatic about it. The tool is good. I use it every day and I’m not going back. But there’s a specific, quiet cost, and I think most developers are paying it without naming it.
The Contrarian Part
Here’s what I actually believe, and it’s not the popular take in either direction.
The optimists say AI will free developers to do “higher-level” work — architecture, product thinking, taste. The doomers say AI will replace developers outright, full stop, so learn to code is dead advice.
Both of these are stories people tell so they don’t have to sit with the actual discomfort, which is this: the value of knowing how to build things hasn’t disappeared, but it’s been unbundled from the value of typing the code that builds them. Those two things used to be the same skill wearing one coat. Now they’re two different coats, and most of us are still wearing both out of habit, not because we need to.
Which means the real question isn’t “will AI take my job.” It’s “which half of my job was actually me.” For a lot of self-taught developers — and I say this as one, no degree, taught myself ADO.NET and C# out of necessity rather than choice — that question lands differently. We built our whole identity on the grind. On being the guy who figured it out without the credential, without the CS degree, purely by doing the reps. AI doesn’t just change the tooling. It quietly asks whether the reps still count for what we thought they counted for.
I don’t have a clean answer to that. I think anyone who gives you a clean answer right now is selling something.
AGI Won’t Announce Itself With a Press Release
Something I keep coming back to: we assume AGI, when it arrives, will feel like an event. A headline. A moment you remember where you were. But intelligence doesn’t usually arrive that way, even in humans. Nobody remembers the day they became fluent in their native language. It just accretes, quietly, until one day you notice you’re doing something you couldn’t do before, and you can’t pinpoint when it changed.
I suspect AGI will be like that. Not a single model announcement, but a slow erosion of the tasks we consider uniquely ours, one small humiliating capitulation at a time. First it writes better boilerplate than you. Then it debugs faster than you. Then, one Tuesday, it proposes an architecture you hadn’t considered, and it’s better than what you would’ve built, and you sit there for a second not knowing what to feel.
That’s not science fiction. That already happened to me, on a smaller scale, with a database schema. It suggested a normalization pattern I’d have gotten to eventually, probably, after a few iterations. It got there first. I didn’t feel replaced. I felt like I’d been quietly outpaced by someone I’d hired to help me, and there was no clean emotion for that. Relief and resentment don’t usually show up in the same breath, but there they were.
What I Actually Do With This
I’m not writing this to warn you off AI, or to hype it either. I use it constantly — for scaffolding, for the parts of development that were always closer to typing than thinking. What I’ve stopped doing is pretending the tool is neutral. It changes what kind of developer you become, not just how fast you ship.
So the discipline I’ve landed on, imperfectly, is this: let it write the code, but make yourself read the diff like it’s a stranger’s PR and you don’t trust the stranger yet. Not because the code is usually wrong. Because the muscle that catches “usually” is the one worth protecting. It’s the only part of this job that compounds in a direction the model can’t easily take from you — judgment about what “correct” even means for this specific system, this specific business, this specific mess of legacy decisions nobody documented.
ASI, if it comes, will presumably have better judgment than that too, eventually. I don’t know what’s left for us on the other side of that. Nobody does, and anyone telling you otherwise is guessing with more confidence than the situation deserves.
For now I just try to notice the hollow feeling when it shows up. Not fight it. Not perform excitement I don’t fully feel. Just notice it, the way you’d notice a new ache in your shoulder — evidence that something in the way you work has shifted, even if you can’t yet say into what.
Top comments (0)