What AI actually multiplies
I wrote earlier about why I’m not buying Jira: a solo user can stand up a custom kanban board with an afternoon of AI help and skip the heavyweight tool. So I did. And the part that stuck with me afterward wasn’t the speed, it was how little of the work was mine, and which part of it was.
The reflex is to call it a ten-times multiplier and move on. But that credits the tool with something that was always mine.
Software experience splits into two layers. One is judgment: knowing what the thing needs to be, smelling a bad design before it ships, stopping the agent from doing something nonsensically complex, understanding that three interfaces over one data model have to agree. That takes a decade to build and decays slowly. The other layer is the churn: whatever framework of the month, whichever database is fashionable this year, the config files that have to line up exactly like a COBOL screen, the forty pages of docs hiding the one flag you need. That turns over every eighteen months or so, and it was never the craft. It was just a tax that “good engineers” have to pay constantly.
I was a senior engineer from the new millennium to about 2006, then I moved into leadership and stopped paying that daily tax. The judgment kept compounding in architecture reviews and postmortems while the stack knowledge went stale. I came to be seen through a kind of software ageism, one of those affable “I used to code” guys, out of touch with the bleeding-edge paradigms and shiny objects. However, the experienced judgment I retained, paired with the stack I’d let lapse, it turns out, is exactly the profile these tools reward: they flatten the churn and leave the judgment alone. It can’t tell me what to build or whether the design is sound, only spare me this season’s incantations. So I’m building things by hand I haven’t touched since 2006, because the engineering was never the bottleneck. The disposable machinery was.
It would be dishonest to stop there, because the number isn’t stable. On a bounded greenfield build the multiplier is enormous. On genuinely novel work, where the model is guessing right alongside me, it drops to about 2 or 3x. On a subtle bug it can go negative, the whole afternoon lost to correcting confident wrong answers instead of fixing the thing I already understood. And the speed is only safe because I can read the output, catch the broken persistence layer, steer it back. The tool multiplies judgment, it doesn’t supply it. Applied to none, it multiplies zero.
This isn’t really about code. I used to publish a post or article maybe every six months, if that. The older posts here show it. Not for lack of opinions, I’ve never been short on those, but because the gap between a finished thought and a live page used to be wide enough to put me off. Now I don’t even see it. I say what I mean, and the push to GitHub and the deploy out to the CDN take care of themselves. Close that gap and the rate climbs, until the only thing left limiting you is how many ideas you have worth keeping.
So I’ve stopped being impressed by “10x engineer, now with AI.” Build the judgment. It’s the only thing here worth multiplying.