The industrial revolution of software
Every time technology makes something effortless, a quieter shift happens too. I think software is next.
For quite some time now, I’ve been wanting to write about this. Let’s call it a prediction. And as any prediction, there’s a high chance I’m wrong. But at least right now, this is what I believe will likely happen with software development in the age of AI.
The industrial revolution
Around 1750, the Brits started to develop ways to make human labour a bit more efficient. Spinning machines, the steam engine, iron production and coal mining all contributed to this.
Locally produced and handmade goods such as clothing, tools or furniture became widely available and mass produced.
Remember that until then, simple and common things such as nails were built one by one by a blacksmith—someone who currently designs and builds custom handmade knives. And clothing was so labour intensive that people owned just a few.
Around 1850 the revolution hit the European continent, allowing railways to be built so that goods could be shipped further away from where they were made.
Guess what happened?
Around the late 19th century, a movement started in—you guessed it—Britain, called Arts and Crafts.
I believe I just made my college history teacher very proud with this one.
Arts and Crafts was based in the idea that machines were degrading the products. Especially around value and quality. It was believed that good design required honest human craftsmanship.
For instance, furniture veneered to look like walnut was very much frowned upon. Materials should look like what they are. Pine—cheap and abundant—disguised as walnut—expensive and scarce—was a no go.
The result was handcrafted goods became premium. And craftsmen like William Morris began selling mostly to wealthy industrialists.
Imperfections and signals of the human touch were considered a feature, not a defect.
In the age of AI
Can you see where we are going already? Currently, AI is present on most developer’s workflows. From the time we write code, to the time the code is deployed, at some point, AI will play its part.
But mostly, AI-generated code—commonly referred to as ‘slop’—is becoming more and more common. The truth is that the code is generated almost instantly, and most of the time, depending on which model was used, it works as expected. So the benefits, especially for someone who cares about quantity more than quality, are pretty obvious.
The code that a developer would need around a week to generate can now be generated by AI in just a few minutes. It’s impressive on its own—let’s not ignore that!
But it’s also imperfect.
You see, as impressive as it may be, the more input you throw at it, the more guardrails you put in place, the better directions you give it, generally speaking, the better the output. So you still need to be there. A human. Mostly to tell it what you want and double-check it gave you what you asked for.
So it became obvious that our job as developers could be on the line. Not because an AI model could replace an entire team, but because perhaps instead of a team of 10 developers, you might just need 2. Along with a subscription—read AI salary.
But the reality is that what AI does really well is writing code. And our job as developers was never to write the code. It’s to think it through. It’s to solve the problems in the most efficient and readable way. Writing code can be seen as a side effect of that. We are not paid to write, we are paid to think.
Handmade code
Machine code is something unreadable for us, humans. It’s ones and zeros put together in gibberish-looking instructions that a machine actually understands.
The code we write is mostly for humans, not machines. You need to write it in a way that your teammates understand. Write it in a way that the future you understands. And document it so that it’s easier to get back into a few months or years later.
Thinking about the code means using the best architecture you can come up with for the problems you have at hand. It means thinking about where each file should go so that it makes sense for a human. It means thinking as a human and knowing what a human might do to decide which edge cases you need to account for.
As said above, AI is good at following instructions, not so much at guessing what you might need or want. That’s on you. You either think about it upfront and tell it all, or you risk edge cases and bugs you aren’t even aware of.
Handmade code means every structure is considered, every interaction is intentional.
It’s the difference between a template and a design that exists only for you.
Security vulnerabilities
Recently, AI models have been finding vulnerabilities in open source code all the time. It’s through these vulnerabilities that things like Shai-Hulud take place. Although AI models can scan entire codebases for a good purpose, it feels like nowadays it’s mostly done by malicious people.
Either way, AI-generated code is not necessarily bug-free.
Besides, unless you have a human in place reviewing the code before pushing, you just gave an AI model a green light to push code to your private repo. How does that feel?
Handmade code is not perfect—far from it—but it means someone is accountable for it. Someone saw the codebase entirely. Someone can explain what every single bit does and why it’s there.
When a stakeholder asks something about your AI-built website and you don’t know the answer because you didn’t write it yourself, you’ll need to defer and ask AI the same question.
A person building a website by hand will know the answer to most questions you could ask. We have a mental model we can tap into anytime. An AI model? Too much context and it may hallucinate. Too little and it may not even know what you’re talking about—but it will still act like it knows everything.
Trust is the ultimate luxury good. A hand-made site comes with a hand-made promise: you own every decision, every dependency, every edge case. That’s something no AI model can offer.
It’s not cheap either
You’ve certainly heard about OpenAI by now, the company behind ChatGPT. But did you know ChatGPT—a tool used by millions worldwide—is not profitable? That’s right. And apparently it’s not expected to be until at least 2030. These big AI companies are bleeding money through what’s called inference costs—that’s paying for every user query along with the electrical bills that come with it. They are spending billions a year and not even making a profit yet.
Hardware prices for components such as RAM sticks and graphics cards—which AI uses a lot—have been rising to rates such as 400–500% in the past couple of years. A kit of 2x32GB of DDR5 memory at 5600MHz is currently costing around 1000 EUR. Read that again and let it sink in. Memory sticks alone have the potential to make your computer build budget rise by more than 100%.
This is not sustainable, and at some point something will give.
My prediction is that hardware prices will not drop significantly by late 2026—contrary to what’s currently believed. Manufacturers have noticed that even the common consumer is willing to pay 2x to 4x more than just a few years ago. I don’t think they are ever going back to ‘the usual’ prices. They may drop them if and when AI companies decide to buy less, but I don’t think RAM prices will ever drop significantly.
Then there’s AI access. Or, who can use it. Right now it’s fairly easy to benefit from AI chatbots or even coding agents. Claude Code has a free tier. ChatGPT has a free tier. Google and Microsoft also have their free tier counterparts. Even DuckDuckGo has a free—and private—chatbot you can use at duck.ai. I personally used it a lot before moving to ChatGPT—which I barely use now anyway.
But there is no such thing as a free lunch. Someone paid for it. And as we’ve learned, electrical bills aren’t cheap. As the adoption of electric cars spreads worldwide, the strain on the electrical grids is only going to make that more evident.
My prediction is that free AI will be a thing of the past in the coming years. You either run your own local LLMs and pay the bill yourself, or rely on paid tiers from third-party providers, paying the bill along with the service.
Let’s go back
As AI companies start to embrace the metered pricing model—you pay as much as you use—it becomes unpredictable how much you are going to pay at the end of the month. And people are going to be extremely more conscious about how much they use AI. In the eyes of a company with several developers, the unpredictability is even higher.
That’s also where projects like Caveman take place. It’s actually borderline genius.
But that still implies you are willing to pay the bill. If you aren’t—and let’s face it, most developers at small companies might not be—you will need to go back to a time where you wrote code by hand. Like a craftsman.
It’s a valuable skill. And thinking through our problems is a valuable skill too. Not just for code.
When machines make everything abundant, what a human makes by hand becomes the only thing worth paying attention to. I believe code is no different.
The old way—sitting down, thinking through a problem, and writing code by hand—turns out to be not just more intentional, but more predictable. You pay once, you own it, and you understand every piece of it.
Let’s not look at it as a workflow from the past. Because it might be the only sustainable one left.
