I built a working thing last week in an afternoon. Auth, a database, a half-decent UI. Five years ago that's a sprint. Maybe two.
I didn't write almost any of it. I read it.
That's the part nobody really says out loud yet. The job is quietly flipping. You used to be the person who types the thing into existence. Now you're the person who looks at the thing and says yes, or no, or "why did you put a queue there."
Author to reviewer. It happened while I wasn't paying attention.
And here's the thing about reviewers — you don't need ten of them.
You needed ten engineers because each one could only hold so much in their head and type so fast. Take away the typing. Take away the "so fast." What's left is judgment, and one person with good judgment can cover a lot of ground. They can hold the whole shape of the system and answer the only question that's gotten more expensive, not less: is this right?
So the team of ten senior engineers becomes one experienced person who can say yes fast and say no faster. The org chart doesn't grow. It folds.
I'm not going to pretend that isn't a little scary to type. I have friends on those teams. I'm on one.
But there's a line everyone keeps repeating that I want to poke at. The line is: "AI can build anything, you just need a good spec." True. Genuinely true — give the machine a sharp doc and it ships.
The punchline nobody adds: a bad product mind also gets a good doc generator now. Feed it muddled thinking and it builds your muddled thing, beautifully, in an afternoon. AI doesn't know your idea is bad. It just makes it faster.
It's a garbage accelerator and a value accelerator. Same engine. The only variable is who's holding it.
Which means the thing that stays rare was never the typing.
It's taste. Knowing this design is generic sludge before anyone tells you. Knowing the user doesn't actually want the feature they asked for. Knowing why Redis and not the obvious thing — not the syntax, the why, the part where business and architecture and the shape of the team all press on one decision at once. The machine writes flawless code and still doesn't know your user. That gap is the whole job now.
I've got a word for the gut feel that flags bad code before I can explain what's wrong. The nose. You can't download it. You grow it by shipping things and watching some of them break at 2am.
The floor came up for everyone. A beginner ships real software today, and good for them, I mean that.
The ceiling didn't move. The ceiling is still made of people.
So the work isn't telling the machine what to build. It's being the one person in the room who knows what's worth building, and being willing to put your name on it.
That's the loop I keep landing on. Code went to nearly free and somehow I'm busier — just busy with different stuff. Less writing. More deciding.
I started these notes a year ago because something felt like it was shifting and I couldn't name it. I still can't, fully. It's the most exciting my job has been in a decade and also I keep catching myself wondering what exactly I'm reviewing toward.
Cheaper code. More of it. Faster.
Toward what, though.