The bubble can pop and nothing goes back

Let me start with the uncomfortable part, because pretending otherwise makes the rest of this useless.

Most of the code I shipped this year, I did not type. I described what I wanted, read what came back, rejected some of it, adjusted the rest, and merged. On the days it goes well I feel like a very well-paid reviewer. On the days it goes badly I feel like a man holding a steering wheel that isn't connected to anything.

That second feeling is impostor syndrome, and the standard consolation for it — you're more competent than you think — doesn't land anymore, because this time the premise changed underneath the feeling. The classic version says you're wrong about yourself. The 2026 version says something true: you really didn't write it.

So the honest question isn't "am I a fraud?" It's "is this job ending?"

The bubble is real. The technology still isn't going away.#

I don't think the current AI economy makes sense. Inference loses money at the prices people pay for it, training runs cost more each generation, and the whole sector runs on a promise that profit arrives later — after scale, after efficiency, after something. The datacentres draw power and water at a scale that has become a local political issue in several countries. A lot of very confident capital is betting on an outcome nobody has demonstrated yet.

Two futures, roughly. Either the bet pays — the capability keeps climbing until the economics work — or the money runs out and there's a correction, and some analysts think a loud one, because this much capex concentrated in this few companies doesn't unwind quietly.

Here's the part that matters for your career: both futures still have LLMs in them.

The weights already exist. Open models keep landing, and the Chinese labs in particular keep demonstrating that you can serve something close to frontier quality for a fraction of what the incumbents spend. Nobody is going to un-train a model because a valuation corrected. And the strategic layer doesn't care about margins at all — this is a technology race between states as much as between companies, and the data flowing through these products is the asset. Data is the ore. Model access is the pickaxe you hand out cheaply so people keep digging on your land.

A crash changes who sells you tokens and what they cost. It doesn't change that your employer will keep buying them, because even at three times today's price they're absurdly cheap next to a salary.

So the feeling is real and the conclusion is wrong#

Impostor syndrome assumes there's an authentic version of you who types everything, and that you're impersonating them. Fine — but ask when that version existed. I didn't write the HTTP stack. I didn't write the query planner that makes my indexes work. I didn't write React, Postgres, or the compiler. My whole career has been assembling other people's work and being paid for the judgment in the assembly.

What's genuinely different is the altitude. The abstraction used to stop below my code. Now it reaches into it, and the thing I'm judging is no longer only "which library" but "is this implementation correct, and should it exist in this shape at all."

That's a harder job, not a smaller one. But it is a different job, and the discomfort is the accurate perception that the skill that got you here — producing correct syntax quickly — is now the cheapest input in the pipeline.

Two doors#

I think there are honestly two.

Leave. Not a joke and not a slur. Electrician, plumber, welder, anything where the work is physical and local and the bottleneck is hands. Those trades are undersupplied nearly everywhere, they pay better than most people assume, and they are structurally hard to automate. If what you liked about programming was the money and the indoor chair, this is a legitimate answer and I'd rather someone say it out loud than spend four years hoping.

Get better. Not "learn a new framework" better. Better at the parts the model is worst at, which turn out to be the parts that were always the actual job and that many of us — me included — got away with doing at half strength because the typing filled the day.

I know which door I'm taking. The harder question is what "better" concretely means when the code stops being the deliverable, and that needs its own post: Getting better at the job the model left you.

One thing I'd ask you to drop either way: the hope that a market correction hands the old job back. It won't. The bill might get bigger and the vendor might change, and the work will still arrive already half-written.