A short field guide for anyone who has just watched a machine write working code

The typing was never the job

AI writes code well. That is a real change. It is not the same thing as “there is nothing left for people to do.” Here is the difference, drawn out slowly.


1

Cut open the job. Find the typing.

Picture the whole job of “making software that works” as one pie. Now find the slice that is a person typing code into an editor.

It is small. It was always small. Everything around it is the work of deciding what should exist, agreeing on what “right” means, and finding out where it is wrong.

typing the code working out what is even worth building deciding what “correct” means here checking, testing, fixing what broke getting people to agree on one answer typing the code
Plate 1. Rough shares, not measured ones — every team splits differently. The point is the shape: the machine got very good at the smallest slice first.

So the honest version of the worry is not “AI can code.” It is: can a machine also do the four bigger slices? That is a much harder question, and the answer today is “parts of them, unevenly, and never the part where someone has to be accountable for the result.”


2

The nail gun problem

Here is a pattern that has repeated for two hundred years, and it surprises people every single time.

A tool arrives that does the main task ten times faster. Everyone assumes the workers doing that task will disappear. Often the opposite happens — because the tool makes the whole activity cheap enough that people want far more of it.

the dip everyone predicts machines in use people employed tool arrives later
Plate 2. The shape, not the numbers. Cash machines and bank tellers is the famous case: machines spread fast, each branch needed fewer tellers, branches got cheap to open, so more branches opened — and teller jobs held up for decades before finally falling.

Software has an unusually strong version of this. There is a huge pile of software that is genuinely useful but has never been built, because it would cost £150,000 of engineering time and only save someone £4,000 a year. That software is not “not needed.” It is priced out.

Drop the price and the pile starts lighting up.

Drag the price down

Twenty-four real-ish jobs that a small business, a school or a charity would love to have built. Each lights up when building it costs less than it saves.

Nobody knows how far down that ladder demand actually reaches. It may reach a very long way. It may hit a ceiling nobody expected. But “the work all disappears” assumes the pile of unbuilt useful things is small, and it is not small.


3

Writing is the easy half. Checking is the hard half.

Try this. One of these lines is wrong. The code runs perfectly — no error message, no crash. It just quietly gives you a slightly wrong number, forever.

Find the line that lies

This is meant to add up a list of numbers and give the average. Tap the line you think is broken.

function average(numbers) {  for (let i = 0; i < numbers.length; i++) {  }}

 

If you found it: notice what you actually did. You did not read the code. You read it against an idea in your head of what an average is. The code was fluent. The code was confident. The code was wrong.

You cannot supervise something you do not understand.

This is the whole argument in one line. A machine that writes code faster than you does not remove the need for someone who can tell whether the code is right. It increases it, because now there is far more code arriving, far faster, and all of it looks equally plausible.

writing it checking it, choosing between answers, carrying the blame
Plate 3. Automate the light side and the scale does not balance out — it tips harder. The weight moves to the side a person still has to stand on.

4

Now the part that actually worries me

Everything above is the reassuring case, and I think it is largely right. Here is the case against it, stated as strongly as I can.

The comforting storyWhy it might not hold
Past machines took over physical tasks, and people moved up into thinking work. This time the thinking work is the thing being automated. There is no obvious floor above it to step onto.
Cheaper software means more software, so more jobs. More software could be built by the same number of people, each doing ten times as much. Demand has to grow faster than productivity, and it might not.
Someone still has to check the machine’s work. One good checker can check a lot of output. Ten writers can become one checker.

And underneath all three sits the sharpest problem, which is not really about unemployment at all.

The missing rung

Beginner work — the simple, well-defined, someone-check-it-afterwards tasks — is the most automatable work there is. It is also, and this is the trap, exactly how a beginner becomes an expert. You get judgement by doing a thousand small things badly and being corrected.

expert judgement starts here beginner how?
Plate 4. Take away the bottom rungs and the ladder still stands — for the people already near the top. Nobody has a good answer for the people at the bottom yet. That is a real, current, unsolved problem.

So the fair summary is not “everything is fine” and not “it is over.” It is: there is a great deal of space left for people, and it is spread very unevenly. The generic advice “just learn to code and you will be fine” is finished. That is a different claim from “coding is finished.”


5

Then should anyone still learn it?

Yes — but be honest about which reason is yours, because the reasons have not aged equally.

Reason for learningHow it is holding up
“It is a reliable ticket to a well-paid job.”Weaker than it was. Still true at the top, much shakier at the entry level.
“I want to be able to check and steer what the machine builds.”Stronger than ever. This is the job now.
“I want to build my own things without asking anyone.”Enormously stronger. One person can now build what needed a team.
“It teaches me to think clearly about steps, state and failure.”Unchanged, and it transfers to almost any systems-shaped work.

One warning, and it matters more than the rest

There is a new and very easy way to learn nothing at all: hit the first moment of difficulty, ask the machine, get a working answer, move on. Repeat for a year. You will have shipped things. You will not be able to tell when the machine is wrong.

real skill ask straight away struggle first, then ask week one year three
Plate 5. The grey line feels better for about six months. The gap between the two lines is roughly the thing employers will still be paying for.

The struggle was never a hazing ritual. It is the mechanism. Being stuck, forming a wrong theory, testing it, and finding out precisely how you were wrong — that is the process that builds the model in your head that you later use to catch the machine lying to you. The tool does not remove that step. It just makes skipping it far more tempting, and the cost of skipping it arrives much later.

So: use it constantly. Ask it everything. But ask it after you have formed a guess, and make it explain rather than just deliver. Read the answer as a colleague’s draft, not a verdict.


The short version

Machines got extremely good at the smallest slice of the job. The bigger slices — deciding what is worth making, deciding what “right” means, noticing when something is quietly wrong, and standing behind the answer — are still ours, and there is now more of that work, not less.

The genuine problem is not that there is no room for people. It is that the room is at the top of a ladder whose bottom rungs are being sawn off, and we have not yet worked out what to replace them with.