I see many people arguing that just like the industrial revolution, AI will lead to the creation of many new jobs. Personally, I have a hard time believing that - especially concerning software.
I believe that every new job that we can come up with in the virtual world is subject to the same automation patterns we are already using, to automate existing work.
In the real world, a carpenter builds furniture. They can do it faster with modern tools, but they are still needed (until the robots come 🤖). Of course you can buy off the shelf items from Ikea, but you trade in uniqueness and individuality.
In the virtual world, software developers (aka internet carpenters) have been focusing on removing themselves from the equation of building websites and applications since the dawn of the internet, while keeping uniqueness and individuality. From Dreamweaver to Squarespace, Wix, Retool, Shopify, Lovable and a host of content management systems in between, all of them focus on allowing non-developers to build their own little internet furniture.
However, LLMs have changed the game substantially. Where before, people were constrained by the capabilities of their chosen platform, meaning they couldn’t just build a website that would randomly open a picture of a dog every few seconds without a developer present (sorry cat people, dog person here 🙃). Now they can. With just words. And with Omarchy, even Linux distributions and operating systems are starting to become malleable jellyware.
For most medium/big applications, you still need internet carpenters to architect and build your digital house and apply some taste to the invisible and visible interiors. However, the more we reflect on our taste and the architectural choices we make and write them into text files, the better LLMs will become at even that. Platforms will have these skills baked in and therefore allow non-developers to build even more.
Now you can say: “Well, but Matthias, look. Like, no problem that means we’ll retreat to hosting and maintaining the infrastructure. Like the good old days where people still knew what they were doing.” To which I would reply: “Do you know how good these models are at server admin?” As soon as we can host a DeepSeek v4.1 Flash capable model on remotely affordable, local hardware, even that starts slowly going out the window.
The point I am trying to make is that the mechanism of “solving” each step in the abstraction ladder is the same. Reflect on what you are doing, write it down, give it to/discuss it with your LLM of choice. Repeat. Until the main question that’s left is: “What do we actually even want to build?”
And I am not even sure that question is fully safe from the same recursive mechanism. Interesting times.