GDP Can't See the Work That's Replacing Your Job

It's five in the morning. You wake up. You brush your teeth, iron something,
throw clothes on, grab breakfast. You run out the door because the train is
coming and you're already late.

Train, ferry, another train. You sit down and work for eight hours. You buy
lunch. You sit in meetings. You do a few things that were actually yours. Then
you reverse the whole trip to get home, cook, get the kids sorted, find twenty
minutes of downtime, and do it again Tuesday. And Wednesday. And Thursday. And Friday — and on the days you work from home the shape of the day doesn't change. You still get up and go to work. The benefit is that there is no commute.

I think that kind of work is dying.

Not quickly, though. That's the part I get wrong. Erik Brynjolfsson has spent
thirty years measuring what technology does to jobs, and his best comparison is
electricity: roughly thirty years passed between electric motors arriving in
American factories and anyone being able to see the productivity gains. He thinks AI moves faster — three to five years — but not overnight.

Two versions of what comes next

S.B. Divya's Machinehood is set in 2095, and the thing that stuck with me
isn't the gig economy. It's the pills. Flow for focus. Zips and buffs for strength and speed. Juvers to heal faster. People dose themselves daily — not to feel good, but to stay economically competitive with machines.

That's the dark version. You keep your job by chemically becoming more like the
thing that's replacing you. People converted into infrastructure.

Naval Ravikant has the bright version. He thinks work ends up looking more like
a Hollywood production: a team assembles for a project, does the work, disbands, and you decide when you want to work on the next project. His mechanism is Ronald Coase — firms exist because coordinating outside them is expensive, and as that cost falls, firms shrink. The part I keep returning to is what he calls specific knowledge: knowledge you cannot be trained for. If society can train you, it can train your replacement.

Two potential futures. I don't know if we get either — more likely something
along the spectrum. But both are answers to the same question: what keeps a person worth paying.

The part nobody is counting

Here's what I keep thinking about.

Brynjolfsson's point is that GDP measures where money gets spent, not where
value gets created, and those two have come apart. When something has a zero
price it carries zero weight in GDP — by construction, not by oversight. So he
built a second measure, GDP-B, where the B is benefits. Instead of asking what
you paid, it asks what someone would have to pay you to stop using it.

He's run that across six hundred goods and services. The average person values
Wikipedia at around ten dollars a month. Value from language models rose about
seventy percent in nine months. Trillions of dollars, he says, that appear
nowhere in the official numbers.

Now go back to that commute.

Every part of it is GDP. The fare, the ferry, the lunch you buy, the office
somebody leases, the seven hours you are counted as present. All of it
registers.

The work that is quietly replacing it mostly doesn't. It happens at kitchen
tables at five in the morning, on tools that are free or nearly free, and the
economy has no line for it.

What got free, and what didn't

So here's what I think that means.

Execution got free. Not cheap — free, and largely uncounted. Work that used to
cost a weekend costs twenty minutes, and no measure we have picks it up.

What did not get free is knowing what was worth making.

Brynjolfsson's answer to what stays scarce is agency. I'd narrow it. Agency is
the wrong size — everyone has some, and "have more agency" is the kind of advice
that sounds true and changes nothing. What twenty-plus years in IT actually
bought me isn't the ability to build faster. The machines did that, and they'd
have done it for anyone. What it bought me is knowing which of the things I
could build is worth building, and which ones look useful and aren't. That
doesn't come out of the tools, and it doesn't get cheaper as they improve.

That's the claim. Here is the evidence.

How this was made

I haven't typed a word of this.

I'm dictating into Wispr Flow, which strips out every um and and and you
know before any of it reaches you. When I stop talking, I run the transcript
through a model to check whether what I said about Divya and Naval is actually
true.

It wasn't, quite. I had Naval's conclusion roughly right and his mechanism
wrong — I'd never have produced Coase from memory. And I'd forgotten the pills
in Machinehood entirely, which turns out to be the most important detail in
the book for what I'm arguing.

Which is the part I keep turning over. The machine caught me being wrong, and
I'm glad it did — this piece is more accurate than the one I would have written
alone. But I can't check the checker. I accepted the correction because it
sounded right and because it came with a source. If it had invented a plausible
mechanism and hung Coase's name on it, I think I'd have taken that too.

That's real leverage and real dependency in the same move. I don't have a
resolution for it. I'd rather write it down than pretend the workflow is clean,
because everyone publishing these workflows leaves this part out.

There's also an agent running in the background while I talk, doing deeper
research on this. When it comes back I'll go through what it found and pick what
stays.

That's my whole job in this piece. Deciding what's worth keeping.

Three machines did nearly all of the making, at a cost the economy cannot see.
The one step I couldn't hand off was choosing what stayed in. Twenty years ago
that would have been the part I was worst at.

Which leaves the question I can't put down. When the machine can do the work —
what is worth making?