On "AI doing the thinking"
A small defense of keeping the thinking part. Also a small complaint about demos.
There’s a framing I keep hearing: the goal is to have AI “do the thinking” so that we can focus on higher-level work. I understand the appeal. I also think it’s mostly wrong, at least for the kind of work I do.
What the demos show
AI tool demos have a consistent shape. Someone has a problem. They type a prompt. Something impressive appears. The audience nods.
What the demo doesn’t show: the person already understood the problem well enough to frame it as a good prompt. They already knew what a good answer looked like, which is why they could recognise that the AI’s output was a good answer. They already had the context to know what to do with the result.
The “thinking” didn’t go away. It got compressed into the prompt and the evaluation.
The skill you can’t outsource
In infrastructure work, the thinking I’m most glad I can’t outsource is the one that happens when something is wrong and nobody knows why.
You’re looking at a system that’s behaving unexpectedly. The logs are ambiguous. The documentation doesn’t cover this case. Nobody’s written a runbook for it.
AI is useful here — it can suggest hypotheses, help you read unfamiliar code, look up API behaviours. But it can’t replace the mental model of the system you’ve built up over months of working with it. That model is the thinking.
A small complaint about demos
Demos show AI doing impressive things. They rarely show AI doing confidently wrong things, which happens just as often. The result is a calibration problem: people go in expecting impressive and are surprised by wrong, when the correct expectation is “impressive and wrong, in roughly equal measure, with no reliable way to tell which is which without the underlying expertise.”
The tools are genuinely useful. The framing around them is often not.
What I’d say instead
Instead of “AI doing the thinking,” I’d say: AI compresses the mechanical parts of some kinds of thinking, which frees up time and attention for the parts that are actually hard.
That’s less exciting as a pitch. It’s also, I think, closer to what’s actually happening.