#40 Writing code by hand will soon be as exotic as punching a card

A developer in 1970 would have found it absurd to write a program without going through punch cards. A developer in 1990 would have found it absurd to code without managing memory by hand. Every generation mistakes the abstraction it inherited for the nature of things, and the next one for dubious magic. We are no different.

Computing has never done anything but move up one level

Let's take the series from the start. At first you program in binary, literally by setting switches. Then comes assembler, which replaces numbers with words. Then compiled languages, Fortran and C, which let you write a formula instead of a sequence of machine instructions. Then managed-memory languages, Java and Python, which remove an entire class of errors. Then frameworks, which remove the plumbing. Then no-code, which removes the syntax.

At every floor, the same movement: you describe more of the intent and less of the mechanics. And at every floor, the same objection: « we are losing control », « it will be less efficient », « this is not real development ». The objection has always been partly right and broadly without effect.

Seen this way, a language model is not a UFO. It is the next floor up, the one where intent is described in plain English. The break is not one of kind, it is one of scale: it is the first floor where humans no longer have anything to learn of the machine's language, because the machine has learned theirs.

The keyboard as a bottleneck

Look at throughput. A trained adult types between 40 and 60 words per minute. They speak between 130 and 160. The ratio is one to three, and it only measures the visible part.

The invisible part costs more. When you type, you compress. You cut, you rephrase, you abandon the nuance because it would lengthen the paragraph. You filter your thinking before it reaches the screen, and that filter is applied by fatigue, not by judgement. Advocates of assisted dictation in 2026 make exactly this argument: the bottleneck in AI is no longer the model, it is the human's input bandwidth. The argument deserves to be taken seriously, even if it often comes from people who sell microphones.

Then add the body. The keyboard and the screen impose a posture, a distance, a fixed gaze. Any interaction that runs through them mechanically excludes the situations where the human has their hands full, their eyes elsewhere, or simply someone sitting in front of them. A doctor in consultation, a front-desk agent, a technician on site, a mountain guide. These are precisely the jobs computing has penetrated the least, and that is no accident.

What is really shifting: the scarce skill changes place

For fifty years, the digital skill that carried value was mastery of the tool. Knowing how to handle a spreadsheet, a database, a line-of-business application, a programming language. That skill is losing value, not because it is becoming useless, but because the barrier to entry is collapsing.

The skill on the rise is more thankless: knowing how to frame a problem. Saying what you want, in what context, under what constraints, and being able to recognise a bad answer when you see one. It is a framing skill, not a handling skill. It looks far more like the work of a project manager or a consultant than that of a software engineer.

One honest reservation is needed here, and it matters. Natural language is ambiguous, code is not. What you gain in accessibility, you lose in determinism. An instruction in English can be read three ways, a line of Python only one. The organisations that will suffer over the next two years are the ones that will have confused « this is easy to ask for » with « this is reliable to execute ». The extra layer of abstraction does not remove the need for verification. It displaces it, and it makes it far less visible.

So what now?

Two practical consequences for a local authority, a tourist board or a small business.

The first: stop recruiting on mastery of tools that will have changed within eighteen months, and recruit on the ability to describe a problem properly. The same holds internally, in the way you train your teams.

The second: every gain in accessibility must come with a control mechanism, or it gets paid for later and at a higher price. Who reviews? Against what criterion? Who is accountable for the output? These three questions cost nothing to ask today and become very expensive to settle after an incident.

That leaves the question of timing. Have these movements already begun, or are we talking about a decade? The answer is in the next article, with what is actually documented for the next six months.