An operator engineer at DeepSeek wrote an essay saying he probably won't lose his job, but may have to change professions.
He had just finished writing the main attention operator for a new model, and was proud of it; he also knows clearly that in six months to a year, an AI-written operator will likely be as good as his, or better. His line: if I have to be overthrown, I'd rather the one overthrowing me is me.
It's the most honest piece of career reasoning I've read in a while. But I think "change professions" is the wrong phrase.
He isn't changing industries, he's changing layers
The next stop he describes for himself: no longer writing every line by hand, but dispatching models, decomposing tasks, evaluating results, and owning the final system.
That isn't another industry. That's the layer directly above the thing he has been doing all along.
That layer was always there. It just didn't need him — capacity at the layer below was limited, so writing an operator to its limit was the scarce part, and deciding which operator to write was comparatively easy. Now capacity below has been multiplied by tens, and the bottleneck moved up wholesale: what to build, whether it's right, whether it's worth it.
Every layer AI eats makes the layer above it more expensive. That isn't consolation, it's supply and demand.
Three conditions for moving up
Once you see it this way, the question changes. Not "will I be replaced" but "have I prepared for the layer above." I think three things have to be present at once.
One: you have to have actually done the work below. Not in order to keep doing it, but to keep the judgment. Someone who has never written an operator can't see what's wrong with one an AI wrote; someone who has never built a system can't see where a generated architecture will collapse. This is the one most often misread as "technical skill doesn't matter any more" — the opposite is true. Once the layer below is automated, the ability to judge its output becomes scarcer, not less scarce.
Two: you have to know what the layer above looks at. The questions up there are always the same set: what is this worth, who does it save what for, why now, what happens if we don't. Anyone who hasn't practised that set will simply be waiting for instructions in a new place. It isn't a talent; it's whether you asked one extra question while doing the work below.
Three: you have to be willing to leave while you're still very good at it. This is the hardest one, and the rarest.
What people find hardest to give up isn't the thing they failed at; it's the thing they are best at. An operator good enough for the main path is years of accumulation, and part of an identity. What makes that essay worth reading is that its author started thinking about the next layer while this was still his strongest asset.
Most people wait until they're forced. And by the time you're forced, the layer above is already full of the people who left early.
Two conclusions that don't follow
"So hurry up and move into management" doesn't follow. The layer above isn't the management layer, it's the judgment layer. It doesn't need someone to coordinate people; it needs someone who can own an outcome and actually tell whether that outcome is any good. These get conflated all the time, and then someone who can't judge is placed in a judging seat with a fleet of agents beneath them, producing errors at scale.
"So don't bother learning the fundamentals" follows even less. Weak engineering paired with AI produces slop several times faster than before. People who can't move up usually aren't people who can't use AI; they're people who can't evaluate what AI handed them.
So
The real question isn't "will AI replace me." It's two more concrete ones: what is the layer above the one I'm on, and while I work this layer, am I also practising that one?
My own path was added a layer at a time. Data analysis first, until I found that data is only one part of a business; then incubating a business, until I found that what a model can replace and what it can amplify are two entirely different questions; and then AI. None of those moves happened because I got bored with the layer below. They happened because the layer below ran out.
What the next layer up is, I'm still working out. But one thing is certain: it won't wait for me to be pushed into it.