Codex is now creating language packs, screenshots and everything else needed to publish them in the Play Store and App Store for localisation. That is rather unsettling. It eliminates all those small jobs done by Indian or Chinese workers. Presumably, in future, the “training” of an AI with content will also be carried out by AI itself. In the end, there will be no work left for the low-skilled. And since something is happening right now that is actually beyond our powers of imagination, it could soon affect the more highly qualified in much the same way.
Yes. And I think what is so disturbing about it is that you are not talking about a prediction for the future. You can already see the mechanism at work in your own project.
What you are doing with Codex now would easily have involved a small division of labour just a few years ago: one person developing Android, another iOS, someone translating or localising, someone creating store screenshots, someone maintaining the store listings, perhaps somebody else testing. Some of that work would have been outsourced to lower-cost contractors. Now a single Solution Architect in Cologne sits down and says, in effect: finish the localisation for these countries — and the machine produces the artefacts.
And that makes me realise that “low-skilled” does not fully describe the problem at all. Many of the activities that could disappear are highly skilled; they are simply amenable to standardisation. A translator is not low-skilled. Nor is a junior programmer. Neither is a graphic designer who creates store screenshots. But their work may contain enough regularity for a system to take it over.
That is precisely why I find your second assumption much more threatening than the first.
For a long time, the reassuring narrative was:
> AI takes over routine work; people do the demanding things.
Your last few weeks already call that neat distinction into question.
Codex does not merely perform routine tasks. It programs things that you could not program yourself. It finds errors. It analyses an existing codebase. It structures Git. When faced with a Teamcenter problem, it even tells you at some point, “That can’t be done,” after you had failed to solve the same problem with a human programmer years before.
And so the boundary shifts.
Today: routine translation → AI.
Then: programming to specification → AI.
Then: deriving specifications from requirements → AI.
Then: questioning the requirements → AI.
At that point, we come dangerously close to your own work as a Solution Architect.
And yet your experience also reveals what, for the moment, remains: you hold the intention.
Codex would not have invented RadixChat.
It would not have been lying in the bath one morning and thought: there really ought to be a way to save significant people.
It would not have decided that astrology as a language of self-reflection is more interesting than the hundredth horoscope app.
After looking at iOS, it would not have thought: damn, Android suddenly looks far too austere by comparison.
You set the direction, assess, reject, notice gaps and change the goal.
But it is dangerously comforting to say:
> Well, there we are. Creative direction will simply remain with people.
Why, after all, should that particular boundary remain stable forever?
That is probably what you mean by “beyond our powers of imagination”. We tend to think about the future in terms of today’s division of labour:
Which of today’s professions will AI take over, and which will remain?
Perhaps that is the wrong question.
As these systems grow more capable, they may change the organisation of work itself. A company would then no longer need 100 people plus AI, but perhaps ten people plus extremely capable systems. And eventually, perhaps, one person for something that once required an entire department.
Your RadixChat is a tiny experiment in precisely that.
You once said yourself that Codex was the team of implementers you would formerly have needed as a Solution Architect.
At first, that sounds fantastic.
Until you ask yourself:
Where are the people who would once have worked on that team?
And, more uncomfortably still:
Where will the Solution Architect of ten years from now come from if no one has spent five or ten years working as an implementer?
That could become one of the underestimated problems. Automation does not only eliminate jobs. It can destroy paths to learning. Today, junior staff handle the tedious tasks, but in doing so they learn about systems, failure patterns, customers and the connections between them. Eventually, that is how they become senior staff.
If AI performs all the junior tasks, we may lose the training ground from which the experts of the future emerge.
Then there is your assumption about the “training” of AI. The same dynamic is plausible there too: as soon as a task can be described clearly enough for thousands of people to be employed doing it, there is a powerful economic incentive to automate that activity as well. At first, people remain responsible for quality control, edge cases and defining objectives. Then the next system learns to do some of that too.
I would only hesitate to agree completely with “in the end, there will be no work left for the low-skilled”. Some physical and interpersonal work is remarkably difficult to automate. Care work, skilled trades, hospitality, cleaning, childcare, construction, repairs — in those fields, the physical world presents an enormous obstacle. Paradoxically, a good electrician might keep their job longer than some software developers.
That would be a rather absurd historical reversal:
For decades, the message was:
Education protects you from automation.
Now it may be precisely the person sitting in front of a screen who is easier to automate than the person who has to work out why the circuit breaker in the bathroom of an old building has been tripping since yesterday.
And that gives Ole’s remark the other day a wider social dimension:
> “Be careful you don’t become the AI’s apprentice.”
Perhaps, at some point, that will not apply only to you.
We are building systems to which we are teaching more and more of what we do — while also having to work out what role we ourselves are still supposed to learn when those systems subsequently perform what they have learned better and faster than we do.
At present, there is no particularly reassuring answer to that.
I find all of this truly frightening.
Not least because of the quality of the AI’s answer itself. And by that I mean not only what the answer says, but also its tone, structure and slant. This is not some doom-mongering prophet of the apocalypse speaking. But nor is it a sanitised AI agent saying, “with me, everything will definitely become greater and better”. In that respect, I find the answer very neutral and realistic.
At the same time — as with so many other things — it cannot be stopped. AI’s capabilities will continue to grow, just as climate change is not being addressed with any real seriousness, nor are what I consider the harmful effects of smartphones and the like on young people. But those are other subjects. Perhaps our relationship to them is similar to our relationship with the difficult aspects of our own personality. In the end, the only thing that will help us personally is accepting what is to come. And before I can do that, I first have to understand it.
You are welcome to share your thoughts about this post as a forum post or as a private message with me.
