here's a meme we've all posted some version of.
"My code works and I have no idea why."
Classic. Universal. Every developer since 1970 has felt it.
But it changed this year, and nobody really announced it. Now it goes:
"The AI wrote code that works. Neither of us knows why."
Same joke. Completely different problem.
The old version was funny because it was temporary
Back then, "I don't know why it works" meant you'd get there eventually. You wrote it. You'd tripped over something, and if you sat with it for twenty minutes you'd figure out what.
The confusion was yours. So was the code. So was the fix.
The new version is different. You didn't write it, you can't debug it, and there is no twenty-minute session that gets you there, because you were never in the loop to begin with.
You approved it. That's a different verb.
The numbers are genuinely wild
By 2026, 92% of US-based developers are expected to use AI coding tools daily, and 41% of all code globally is projected to be AI-generated. Roughly 40% of junior developers deploy code without fully understanding it.
Sit with that last one. Not "wrote without understanding." Deployed.
Then there's my favourite study, the one nobody wants to talk about: developers believed AI made them about 20% faster, and measurement showed they were actually around 19% slower.
We're not just shipping code we don't understand. We're wrong about whether it's even helping.
Meanwhile, at the top
Andrej Karpathy, who coined "vibe coding" in the first place, called the term passé back in February. His argument was that the frontier had already moved past autocomplete to agents that plan, test and correct themselves across whole codebases.
The guy who named it retired it before most of us finished arguing about it.
And underneath, junior software developer employment has now declined for 33 consecutive months.
So the entry-level job is disappearing at the exact moment the main skill being demanded is judgement, which is the one thing you can only get by doing the entry-level job.
Great. Love that for us.
The part I actually think about
I'm a final-year student. I'm supposed to be the target market for all of this.
And the honest version is that AI is incredible when I already know what I'm looking at, and quietly dangerous when I don't. Same tool. The difference is entirely me.
When I understand the domain, it's the best pair programmer I've ever had. When I don't, it's a very confident stranger handing me something and walking away.
The trap isn't that AI writes bad code. It writes fine code, usually. The trap is that working code feels like finished code, and those have never been the same thing.
My rule now, for whatever it's worth:
If I can't explain why the code works, I haven't finished. I've just stopped.
Not "read every line." Nobody does that and you know it. Just: can I say out loud what it does and why that's the right approach? If not, I'm not done, no matter how green the tests are.
Your turn
I want the honest answers, not the LinkedIn ones.
What's the worst thing you've shipped that you didn't fully understand?
I'll start. I once merged a fix for a bug I never actually reproduced. It worked. It's still in production. I still don't know if it fixed the bug or just moved it somewhere I haven't looked yet.
Drop yours below. No judgement, we're all in the same boat and the boat was generated by an LLM.
Sources: daily.dev on vibe coding in 2026 · Forbes on entry-level coding jobs
Top comments (0)