The plumber was right

Last March my kitchen sink started making a sound like a small animal dying inside the wall. Not my problem, technically. I’m a renter. But I was home, the landlord was in Normandy, and the message thread was going nowhere.

So I called a guy. Marc. He showed up in a van that smelled like coffee and PVC glue, took one look under the sink, and said something in French that translates roughly to “yeah, that’s bad.” Then he fixed it in forty minutes with tools I didn’t know existed.

I went back upstairs to my desk, where I’d been fighting a segfault in a custom memory allocator for an inference chip. Six hours in. Still broken.

And I had this stupid, petty thought: he’s going to be fine.

Not “rich.” Not “fulfilled.” Just fine. People will always have pipes. Pipes will always fail. Marc doesn’t need a demo day. He doesn’t need to explain his value to a board that thinks “AI strategy” is a personality trait.

I build the things that are supposed to replace people like me. I know that sounds dramatic. I’m not having a breakdown about it. I’m just being honest about what I do for a living.

The asymmetry nobody talks about at meetups

There’s a weird split forming and I see it every time I take the RER into Paris.

On the train you’ve got people in Patagonia vests talking about “leveraging LLMs for workflow optimization,” which is a sentence that means “we’ll fire half the analysts and the other half will babysit a chatbot.” Fine. That’s their job.

Then you get off at Châtelet and walk past the guy re-laying cobblestones by hand. Or the baker who gets in at 4am. Or the woman at the marché who can tell you which tomato will actually taste like something this week.

None of them are on LinkedIn. None of them need to be.

The asymmetry is this: knowledge work is getting cheaper at the top of the skill curve and more precarious in the middle, while physical work is still bounded by physics. You can’t prompt-engineer a burst pipe. You can’t fine-tune a sourdough starter (well, you could, but you’d still need hands and an oven and the willingness to wake up before the sun).

I say this as someone who loves what I do. I like compilers. I like watching a biped robot learn to walk badly and then less badly. I like the moment a training curve finally bends upward after you’ve been staring at flat lines for a week.

But I would not bet my children’s rent on my job existing in the same form fifteen years from now. Marc probably would.

What I actually think is coming

I’m not a futurist. I don’t have a Substack about “the singularity.” I have a laptop and too much caffeine and opinions I’d rather write here than scream at a dinner party.

What I think is coming is boring and brutal at the same time.

A lot of white-collar work is going to get hollowed out. Not all at once. Not with robots marching through offices. More like: one team of ten becomes two people and a cluster of agents, and the eight who left don’t all land somewhere better. Some do. Most don’t, if history is any guide.

The jobs that stay will cluster at the extremes.

At one end: enormous organizations running enormous projects. Think the kind of company where “shipping” means coordinating four thousand people across twelve time zones to put a new model behind an API. The work is less about any single person’s craft and more about staying attached to a machine that’s too big to fail and too complex for any one human to understand.

At the other end: very small. A bakery. A repair shop. A three-person studio that makes furniture. A farm stand. Things where the customer can point at the person who did the thing.

The uncomfortable part is the middle. The fifty-person agency. The regional bank’s IT department. The “we do consulting for mid-market companies” shop. That’s the part I think gets sanded down.

So what do you do with that

I don’t have a clean answer. Anyone who does is selling something.

What I’ve started telling friends who ask (and they do, more lately) is less “learn to code” and more “learn something that can’t be emailed.”

Plumbing, obviously, is the boring example. But also: welding, nursing, growing food, fixing bikes, installing heat pumps, doing electrical work where you still have to physically verify the thing won’t burn the building down.

Not because those jobs are easy. They’re often hard and poorly paid at the entry level. But because the failure modes are visible. A bad weld you can see. A bad spreadsheet from an AI agent might look perfect until the quarter closes.

I’m not quitting tomorrow. I still think there’s a window to do interesting work in ML systems if you’re close to the metal and willing to move when the wind shifts. But I’m also signing up for a woodworking class in the 20th. My girlfriend thinks it’s a hobby. I think it’s a hedge.

Marc charged 120 euros. My allocator bug took another two days. He was right.