Sending a robot

Over the last few months, pretty much every PR I’ve shipped has been generated.

Pre-AI, reviewing coworkers’ code was valuable because I could apply my experience of building in the same codebase to reviewing it. I knew the terrain and the shape of the system. Now, most code is sort of… good enough, probably fine. I don’t feel its correctness so strongly.

The agents still need direction. I hold them to a high bar and the results are usually pretty good.

But something feels different.

The old flow state born of friction and understanding meshing together is weaker now. My own PRs sometimes feel foreign. I review generated code, feed comments back into the agent, then move on.

The result is usually correct, but do I understand it as deeply as I used to?

It’s easy to outsource thinking, but growth requires presence and work. When I stop working out, I get weaker.

Coding has always felt like refined thinking to me. Challenging enough to train the mind through the act itself. Moving through a codebase forms an intuitive understanding of the landscape.

Maybe I’m still building this understanding, just from a different vantage point.

Perhaps value is shifting upwards into judgement and taste. I think they’re distilled from experience. The hands-on building of the trees helps inform the forest.

Is there a danger if people skip developing the low-level intuition that higher-order judgement historically emerged from?

There’s a parallel with when personal computing first took off. Productivity was meant to skyrocket. Every major abstraction layer shift makes big promises.

You don’t really know a place until you’ve been there yourself.

Sending a robot isn’t quite the same thing.