Writing
AI, Layoffs, and the Capability Gradient
People ask whether AI will take jobs as if jobs were one thing. They aren't. What AI actually changes is something I think of as the capability gradient: work that is well-scoped, repeatable, and easy to check gets cheaper, while work that requires framing a messy problem — and being trusted with the consequences of getting it wrong — gets more valuable. The gradient always existed. AI is making it steeper.
I have a closer view of this than most, and closer than I'd sometimes like. My job at Roblox is building AI systems that do work humans used to do: reviewing content, making moderation decisions at a volume no human team could handle. I don't have to speculate about whether models can absorb real work. I watch it happen. I also watch exactly where it stops.
Where it stops is verification
Models compress work fastest where the task is easy to specify and easy to check. That's the whole gradient in one sentence. If you can write down what success looks like and confirm it cheaply, the work is on its way to being automated. If success is hard to define, or expensive to verify, or only legible to someone with context the model doesn't have, the work resists — for now.
This is why the gradient doesn't map cleanly onto status. Some high-status white-collar work is, if you're honest about it, repeatable synthesis with a nice title. Some lower-status work is dense with messy local context, negotiation, trust, and physical-world checks no model can do yet. The market is going to reprice along the gradient, not along the title, and that's going to surprise people on both ends.
The uncomfortable part
The implication I like least is that AI widens the gap between workers rather than narrowing it. Someone with strong judgment who uses these tools well becomes dramatically more productive. Someone whose work was mostly executing well-defined tasks watches the market reprice that execution downward, through no fault of their own.
This is also why layoffs in this world aren't one-for-one replacement — no robot slides into your chair. What happens instead is that organizations discover a smaller group of high-leverage people plus AI systems can cover what used to take a much larger team. From the outside it looks like ordinary cost-cutting. The mechanism underneath is different, and pretending otherwise doesn't help anyone plan.
Moving up the gradient
So what do you do about it. The honest answer is that I don't know how fast this plays out — nobody does, and you should distrust anyone who gives you a date. But the direction seems clear enough to act on: move toward the work where ambiguity, ownership, and consequence stay high.
Concretely, that means shifting from doing the known procedure to deciding what should be done. Define the problem. Design the verification loop — when the cost of a wrong answer is high, somebody has to decide what "checked" means, and that somebody is doing work the model can't. Own the production boundary, the place where the system meets reality and someone has to answer for it. Do those things and AI multiplies your output instead of replacing it.
The career strategy is not to compete with the model at knowing things. You'll lose, and the margin grows every quarter. The strategy is to be the person who decides what the model should do, how to check whether it did it, and when to turn it off.