Crafting Was the Rehearsal. Reviewing Is the Job You Have to Keep Training For.

I’ve been chewing on a question that isn’t really about AI: what part of programming was I actually in love with? Not the software. The act. And once I’m honest about the answer, the “am I worth less now” panic going around stops making sense to me. Not because the panic is silly, but because it’s aimed at the wrong target, and because the comforting answer to it is also wrong.

The question everyone’s asking is the wrong one

ThePrimeagen’s talk “I suck”, a talk he gave under the working title “what is my value,” opens with the question he says he gets from his audience constantly, sometimes from people in real distress: “Is everything I’ve done for the last… twenty years, is it worthless? Am I worth less now?”

He tries two answers and rejects both, and he’s right to. Value isn’t “taste,” the word he mocks as VC-speak for making things look nice. A model can do that too. And it isn’t raw output either: he clocked around 15,000 lines a week at his peak, and someone else does more than double that in a single day. You cannot out-type a tool built to type. Neither of those was ever really what made the job feel good, for him or for me.

The chromium tweet

The turn in his talk is a screenshot: someone asks an AI assistant for help with what they call a trivial problem, and the model’s internal monologue casually suggests forking Chromium. He doesn’t let it slide:

“How do you build a web project in which the result of whatever you’ve made choices of as engineers led you to forking chromium? This is the worst decision I’ve ever seen.”

I laughed at that line the first time, then it stopped being funny. The model wasn’t wrong about how to fork Chromium: compiling it takes hours, but it’s a well-documented process. It was wrong about whether to, and nobody in the loop caught it before it became a live suggestion. That’s the job. It was always the job. Crafting was just how you trained for it.

Where the comforting version of this stops short

Here’s where I think most people land after watching this talk, and where I think they stop one step too early: “okay, so I’m not worth less, because humans still need to review the agent.” True, but treated as a stable fallback position, a job that survives no matter how good generation gets, requiring no upkeep of its own, that’s wrong. And it’s wrong in a way that matters more than the reassurance is worth.

The scar tissue that let ThePrimeagen catch the Chromium suggestion in two seconds wasn’t handed to him. He built it, specifically, by writing the wrong version of something, watching it go badly, and doing the work of understanding why. He gestures at this himself with what I think is the most important line in the talk, on “toxic productivity”: failing repeatedly to understand something is more valuable than shipping something clickable. He doesn’t follow it to its conclusion, but I will: if that failure-and-understanding loop is how judgment gets built in the first place, and tools now remove most of the occasions to run it, the supply of people who can catch the next Chromium fork doesn’t refill itself for free.

Reading a plausible diff and checking it against intuition you already have is a different cognitive act than building that intuition in the first place. It’s not obviously true that doing a lot of the first keeps you good at the second, any more than a pilot who only monitors an autopilot stays sharp at hand-flying by watching it happen. Currency requirements exist in aviation because that skill decays specifically when you stop exercising it, even while you’re still technically “in the loop.” I don’t see why code review would be exempt from the same pattern.

So, am I worth less?

Only if the only thing you were ever worth was your typing speed. If it was your judgment, the years of getting it wrong until you stopped, that didn’t get cheaper. It got more load-bearing, because it’s now the only thing standing between five generated options and the one that quietly forks Chromium. But it also isn’t a stock you built once and get to keep spending down forever. It’s a flow, and it needs the same kind of input that built it in the first place.

That reframes “toxic productivity” from a nice sentiment into an actual prescription. Don’t just review the next agent-generated change and move on. Every so often, before you look at what it produced, write your own first pass badly, slowly, by hand, and then compare. Not because your version needs to win, but because that’s the rep that keeps the muscle from atrophying while everyone assumes it’s permanent.

ThePrimeagen closes his talk on a DHH quote: “it’s fun to be competent.” I think that’s true, but competence just moved, and it won’t stay wherever it moved to on its own. It used to live in your hands. Now it lives in the two seconds before you approve someone else’s diff, including the agent’s, and that’s a muscle you keep by using it, not by having used it once, twenty years ago.

Go watch the source talk, rough audio and all. Then find one place this week where you can afford to do it the slow way on purpose.