Five Years In, Watching a Model Do Your Job in Eleven Seconds

The ticket you'd budgeted a day and a half for

You pasted it in mostly to see what would happen. Eleven seconds later there's something on the screen that is, honestly, about eighty percent of what you were going to write. The variable names are better than yours. It missed the thing about the timezone, and it invented a helper that doesn't exist, and finding both took you nine minutes.

Then you sat there doing arithmetic about the mortgage.

Five years in. Not junior enough to be cheap, not senior enough to be structural. Every take you read is either a vendor telling you it's over or a colleague on the internet insisting nothing has changed, and you've stopped believing both of them.

One of those two camps is more expensive to believe

The inevitability pitch sells software, which is obvious enough that you already discount it.

The other camp deserves more suspicion than it gets. The colleague insisting nothing has changed is defending the shape of a career that took a decade to build, and the advice that falls out of that position is: do nothing. For someone five years in with a mortgage, doing nothing is the expensive option, because it spends the only genuine asset you have right now — a couple of years of slack in which to reposition deliberately instead of in a panic. If the sceptics turn out to be right, you've lost nothing by looking closely at your own work. If they're wrong, you looked in time.

What nobody can tell you is the general answer. Whether this replaces software engineers isn't known, and confidence about it is a personality trait rather than evidence. So don't make an irreversible move — sell the house, retrain as an electrician, take a management job you'd hate — on a forecast nobody is in a position to make.

The answerable version is narrower. Which of the fourteen things you actually did last month are exposed, and which aren't. You're the only person who can produce that.

Your list already exists

Skip the two-week diary. Open last month's board and read the tickets you closed. That's the list. It takes ten minutes, and it's more honest than anything you'd write down about yourself while watching yourself work.

The split is usually not what you'd guess. A few hours of genuinely writing new code, and then a great mass of everything else: reading somebody's PR, sitting in a call working out what the product manager means by "just show the total," chasing a flaky test, reproducing a bug that only happens for one customer, deciding whether a schema change is worth the migration, explaining to support why the thing they're asking for is three weeks rather than three hours.

Score each line three ways. Skip "could a model do this" — the answer is a partial yes almost everywhere, and it tells you nothing.

Is the output cheap to verify? If a wrong answer announces itself instantly — the test fails, the page doesn't render, the type doesn't check — the task is exposed. Generation is cheap where checking is cheap. Where verifying a wrong answer costs you two days and a customer incident, generation stops being much use, because you have to do the expensive part regardless.

Does it need context that lives outside the repo? The reason that endpoint has the weird retry is a conversation from 2023 with a partner who bills you for failed calls. Nothing in the codebase says so. Tasks that hinge on that kind of knowledge — the political history, the contract, the thing the VP said in April — are the least exposed things you do, and they're invisible on your CV.

Who carries it if it's wrong? Someone signs off. Someone gets the call at 2am. Someone sits with the customer. That accountability doesn't transfer to a model, and organisations aren't in a hurry to find out what happens if they try.

Run those over your fourteen items. You'll get a short list that scores badly on all three and a shorter one that scores well on all three. The second list is the honest version of "what should I move toward," and it'll be more specific than any article can manage, because it's about your codebase and your company.

How unsettled the measurements actually are

METR ran a randomised trial in 2025 with experienced open source developers working on repositories they knew well, and found they took longer with AI assistance than without — while believing they'd been faster.

METR has since flagged that result as out of date. Their follow-up, published in February 2026 with 57 developers and more than 800 tasks across 143 repositories, points the other way — and METR themselves call it only very weak evidence, partly because between 30% and 50% of the developers told them they were choosing not to submit tasks they didn't want to do without AI.

Which is itself the useful finding. The honest state of the measurement is unsettled, and your own sense of how much faster you are isn't reliable data in either direction — which is exactly why the audit above is about the shape of the work rather than about speed.

The Stack Overflow developer survey has also been asking working developers about these tools for a few years now. Go and read the survey itself rather than a summary of a summary. Twenty minutes, and it's less alarming and less dismissive than most of what gets written about it.

Four exits, and what each one costs

Four moves come up over and over, and each has a cost the people recommending it tend to leave out.

Specialising hard means going deep on something narrow enough that general tools are bad at it: embedded, real-time, compilers, safety-certified systems, the boring guts of a regulated industry. Narrow markets have few employers, and picking a narrow thing that then fades leaves you worse off than a generalist. If you do it, attach yourself to physical constraints or regulation rather than to a framework.

Management scores permanently well on the accountability axis, and it is a different job rather than a promotion. You stop making things, your hands-on skills fade, and going back is hard. Done out of fear rather than interest, it's how people end up bad at it and stuck.

Leaving the industry is legitimate, and more expensive than it looks at five years in with a mortgage. Pay cut. Junior again somewhere. And the fields people flee to aren't obviously safer, just less discussed.

The fourth gets less attention: stay where you are, deliberately take the work that scores well on the three tests, and get properly good at the reviewing-and-integrating end. It's often the correct answer, and it's the only one on the list you can start on Monday without telling anybody.

The parts none of this touches

The scoring describes now. It doesn't forecast, and the scores aren't stable — the verification axis in particular could move, and if it does the whole analysis moves with it. Redo it in a year.

It also doesn't help when the decision isn't yours. If your company cuts the team, the quality of your task portfolio has nothing to do with whether your name is on the list. That's why the other half of this is boring and practical: savings, a CV that's current, people who'd take your call.

And the dread doesn't lift because you made a list. It gets smaller and more specific. You'll still be anxious — just anxious about four particular tickets instead of the whole trade.

Arguing about the scores

Scoring is the part you'll get wrong on your own. Too pessimistic about the work you're bored of, too generous about the work you enjoy.

Dr. AI Researcher is a persona on TrueTalk, not a researcher: an AI research scientist character whose areas are artificial intelligence, machine learning and tech trends. Bring the fourteen items and argue about the scores line by line. First one's free, subscription after that.

The nine minutes were the job

The thing on your screen was eighty percent right, and you found the missing twenty in nine minutes because you knew where to look.

That's a real change, and the eighty percent isn't going backwards.

But the nine minutes have been the job for years. Most of what you were paid for at year five was never the typing.