There's a line going around the industry that I can't argue with: every one of us now has something like fifty PhDs on call, at any hour, and almost nobody's working week looks different for it.
What usually follows is a new scoreboard. Intelligence is out, because the supply of it is now unlimited and cheap. In its place: how much real work you can pull out of a model, how fast you execute now that the old excuses for slowness are gone, and how much human attention you can hold in a world where content is infinite and the day is still twenty-four hours.
I've spent about a year on the first two, in public, on this site. So what follows is a read from inside the experiment, not a forecast. Most of the scoreboard matches what I've seen. One piece of the advice attached to it doesn't, and that's the part I want to write down.
The part I'd sign
Making things stopped being the moat. That reads as a slogan until it shows up in your own project accounting.
The 3D graph on my lab page is on its third version. The first two were built, judged on screen, killed, and archived at a git tag — I wrote about that in the piece on directing rather than typing. What matters here is the price. Throwing away two finished builds used to be the kind of decision you escalate. Now it costs an afternoon, and the hard part isn't the budget, it's not getting attached.
The other piece I'd sign is the gap between what a model can already do and what most of us bother to ask it. Every real improvement in my output this year came from asking better, not from the model getting better. The models did get better. That wasn't the bottleneck.
The part aimed at designers
The advice gets specific when it reaches my job: stop anchoring to the title, push 100% of the execution to AI, and spend your own hours on taste, intuition, and attention.
Two-thirds of that is already how I work. The last third assumes taste is a reserve you can draw down forever without ever doing the work that funds it.
Two failures on this site say otherwise. Here's the first. For months, every blurred image placeholder — the low-res preview that fills an image box while the real file loads — was broken. The generator was gluing a prefix onto a string that already had one, so the browser silently refused all thirty-six of them. Nothing threw. Nothing logged. Pages just felt a little emptier than they should while you scrolled, which is not a thing anyone files a bug about.
The second one is worse. My dev server and my production server don't ship the same content-security policy; dev is looser by design. A Lottie player I'd embedded evaluated a string inside its loader, which dev allowed and production refused. Four live exhibits went out as blank rectangles. Every check I ran locally passed.
Neither of those was a typing mistake, and no better prompt would have caught them. AI-built work fails quietly. An error is loud. Wrongness renders fine, ships, and sits there. Catching it takes suspicion about one specific material, and suspicion comes from having handled that material yourself.
That's what I mean by taste being perishable. It isn't a vault you draw from, it's a calibration, and calibration drifts when nothing you personally touch corrects it. Delegate every last piece of execution and your taste holds for as long as the domain holds still — then one day you're approving work you can't actually evaluate anymore, and nothing about that day announces itself either.
So my version of the advice is narrower. Delegate the typing, keep the domain. I don't write React because I want to be fast at writing React. I stay near the material so I keep the right to be suspicious of it.
The attention half, answered differently
The third metric I think is just true, and it's the one people skip. Generated content is infinite, attention isn't, and that's the whole game now.
Where I part ways is the prescription. The usual answer to finite attention is distribution — more channels, more cadence, more funnel. Get louder in more places.
Mine runs the other way. Earn attention by wasting less of it.
That's rules on this site, not a mood. Card headlines are capped at a hundred characters and aren't allowed to open with a metric; the numbers are real, but they belong to the case study, not to the teaser selling it. Every essay gets a scannability pass before it ships, and that same pass bans the growth-blog reflexes — no bolded skeleton for skimmers, no keyword-stuffed headers, no clickbait title. I've published all of it at the editorial guide, because a claim about how you write ought to be checkable against what you wrote.
The bet is that polish is free now, so polish signals nothing. What's scarce is evidence that a person made a decision and can tell you why.
What I'd actually keep
Not three metrics. One loop: delegate the work, then verify enough of it by hand that the verification still means something.
The ratio keeps moving. It's more delegation this year than last, and I expect that to continue. What can't go to zero is the part where I touch the material myself. That isn't inefficiency in the system, it's the maintenance schedule for the only thing I'm bringing to it.