Web Development Lost Its Value, Hardware Didn't
About how AI is making basic code generation less valuable, why web development is no longer such an easy way into IT as it used to be, and why hardware now looks like a much more interesting niche. Plus a bit about how I use LLMs to design PCBs, what they can actually automate, and what you still have to test, solder, debug, and finish with your own hands.
A few years ago, to get into IT it was enough to go through some courses or spend a couple of months at home, do some layouts, mess around with React, make a few pet projects, and someone could already hire you for $1000+ a month. Now everything is a bit different. You either enter the market already at a mid-level, or go junior for some $300 and try to somehow survive among hundreds of the same candidates. The world changes, and you have to change with it.
AI friend or enemy?
For me, AI is more like a wild dog that you need to tame. If you actually care about your work, you can learn how to use it properly and boost yourself a lot. Although it can also be the opposite: some dude who was doing the bare minimum for years can quickly play around with agents, realize that half of his work can be automated, and continue making money while playing even more dota. I heard more than once about people who were sitting on mid positions for years, doing the absolute minimum and not giving a fuck. But anyway, if you completely ignore AI now, you start falling behind pretty fast.
AI really boosts productivity. When someone says that it doesn't help them at all, very often the problem is how exactly they use it. Someone asked chatgpt once to write some code, got some bullshit and decided that all of this doesn't work. Someone just treats it like a better google. But almost everywhere where the main work happens on a computer and you work with text, code or information, it can already take a huge part of the routine from you. Programming, accounting, documentation, analytics, working with text, the list can go on for a long time.
At the same time, AI doesn't have to do all the work instead of you. If you are a writer and just start generating all your text with it, you can very quickly erase your own style and get the same faceless shit that is already everywhere on the internet. But you can use it to edit text, find weak parts, work with structure, do research, generate images or just use it as another pair of brains. Programming is a bit different: there the point is exactly that the agent should write a big part of the code, while you think about what exactly it should write, how it should be built and whether it is generating another pile of slop for you.
The price of code — 20$
You can generate code like crazy now. But if you do it without any system and without experimenting, very quickly you get a mess that is hard to maintain: you add a new feature, an old one breaks, after a couple of weeks the agent itself no longer understands what it wrote, or some bug appears that neither you nor the coding agent can properly find. So code generation itself already almost means nothing. More important is how you organize the generation.
For me, for example, feature based architecture works best right now. When I started using it properly, a big part of the problems with architecture that the coding agent could create for itself just disappeared. There were fewer situations where one new feature started breaking half of the project. Same thing with prompts to AI: a vague prompt works fine where you still don't really know what you want.
For example, design. You can say make it prettier and run dozens of variants until something normal randomly comes out. Or you can spend 20 minutes, find some references, understand what exactly you like, and give a more targeted prompt. In the second case you simply save yourself a lot of time and nerves.
And here we come to the main point. The value now is not in whether you know how to generate code. The value is in product thinking: proper architecture, analytics, understanding the user, marketing, sales, distribution. Maybe being able to properly promote a product is now even more important than being able to generate it properly. Because now practically any random dude with codex, claude code or some other agent can crank out a saas in a few days, but not everyone can get their first five bucks from that saas.
And actually this is a very interesting time to try making money from your own products.
Hardware as an alternative
For the last few months I've been experimenting a lot with designing boards using LLMs. I mostly use codex and claude. Antigravity with gemini still handles this much worse, and different mcp for kicad that I tried also still look like complete garbage. There is another problem: designing boards eats a shit ton of tokens. A normal project can easily take 40+ million.
The process of generating a pcb board together with the schematic looks like this:
- A python script is generated that creates the schematic in kicad
- Based on the schematic, the pcb itself is generated
- Traces are routed partly automatically, for example through freerouting
But these are only the main steps. Between them there can be several regenerations of scripts because the board came out wrong, components were placed stupidly, or somewhere the agent misunderstood the schematic. If there is an mcu on the board, you can also throw together some basic firmware in parallel, compile it and try to at least partially run it in an emulator.
After that the real world starts. You buy components, build the schematic on a breadboard or protoboard and check if it actually works at all. Whether the components communicate properly, whether the firmware behaves how you expect, whether there are problems with power, whether there are some weird things that looked completely normal on the schematic.
If everything works, you order the first pcb revision, assemble it and test it. You check if there are any mistakes on the board, if the footprints are normal, if components physically fit properly, if there are no problems with the case, connectors, power and everything else. And only after that you can think about some normal production batch.
From the idea to the first actually working board it can take around a month and a half. It can be faster, it can be much longer, everything depends on complexity and on how many problems show up during prototyping. Because hardware debugging is a fucking pain.
And here, in my opinion, there is an interesting window right now. saas became extremely accessible. Almost anyone can make another web service now, because of that every second person generates something, launches it and spams it on Reddit, X, Product Hunt and everywhere else. In hardware the entry barrier is still much higher, because here it's not enough to press enter, enter, enter and wait until the agent generates something.
You need to order components, solder something by hand, wait for pcb, figure out why this shit doesn't start, think about the case, production, shipping, returns. Because of that there is simply less competition. At the same time there can be no less money in hardware than in saas.
But there is one thing. To properly make money on hardware, it's not enough to just be a good hardware engineer. At the same time you have to be a bit of a businessman, logistics guy, coder, product guy, buyer and marketer. That's why a normal team of like-minded people here makes even more sense than in regular software. You can do it alone, but it is much harder.
Conclusion
The time when you could learn how to crank out websites and automatically get a good salary is over. And this is not only about Ukraine. Knowing how to write code is slowly becoming less rare, because AI makes code production cheaper and cheaper.
The value moves higher: into the ability to come up with a product, design it properly, bring it to a working state and, most importantly, find people who are ready to pay for it.
So now the question is not really whether AI will take your job. The question is more about what you can make with AI that someone is actually ready to pay for.