Are you using AI?

AI — especially modern LLMs — can be extremely effective for software and firmware development. In many situations, tasks that normally take significant time can be solved much faster. It’s also clear that AI often approaches problems in a fundamentally different way than the human mind, which means any solution integrated into a project must always be fully reviewed, validated, and traceable.

I'm building two busker organs, one working with 3D printed pipes, one working with accordion reeds. The latter one needs an Arduino Uno Q, which runs a Linux system, and is designed to be programmed in Python. I knew nothing about Python, but with the help of an AI assistant, I've learnt a lot. Using Visual Studio Code, I also get a lot of suggestions how to write the code line by line. I read those lines and accept them (or sometimes change them) only after I understand them completely.

During the process (which has taken over 2 years by now) I've also written a lot of music for the busker organs. First I had no interest in creating music with the help of AI. What I did instead was I wrote the music by heart (as I always do), but I had AI design a title page image for each tune. That was fun! Although I'd very much like some human artist to do the art work instead. But since I can't pay anyone, and there's no profit prospects in this, I haven't even tried that path yet. Here's some of the AI art:

But I am working on AI generated music, too, just to learn how neural networks work. I'm using my 15 original busker organ tunes as training material for a neural network, the task of which will be to produce music that sounds like my tunes.

I also tested Suno, just for fun. It was fun, and a bit creepy, how it turned one of my tunes into something different from mine. I fed the beginning of one of my tunes into Suno and the result was so funny, that I might implement a Suno interface into one of my busker organs. Then again, using AI to create images or music tends to be a thing where I am the one, who creates and refines the prompts, then selects the best outcome. And that is kind of a creative process, which is interesting. But I rather spend my creativety on writing the music myself.

I used Claude alongside KiCad primarily as an engineering assistant for schematic design rather than PCB layout.

It was especially useful for automating repetitive schematic tasks. For example, I could ask it to duplicate a level shifter stage around Q1 multiple times, saving me from a lot of manual editing. It also handled large-scale design updates, such as marking every 0402 resistor as DNP across the project.

Beyond automation, I used it as a design reviewer. I'd have it inspect the schematics for potential issues like startup conditions, sanity-check enable signal sequencing, and suggest where pull-up or pull-down resistors might be needed on control signals. It was a good second set of eyes for catching things I'd want to verify.

I didn't use Claude for PCB routing, though. Routing is the part of board design I genuinely enjoy—it's a creative process where I prefer to stay hands-on.

How did Claude help you design PCBs? What did it do for you besides research automation etc.?

I started using Cursor in December 2024, then switched to Roo, and eventually landed on Claude Code. If I have a full day to focus, I'll burn through 300M tokens without thinking twice.

These days, Claude has helped me complete several C/C++ MCU projects, design PCBs, select parts, and think through optimization strategies in areas where I had far less experience. It's been like having an expert collaborator, and I can feel how much my own skills have leveled up.

More recently, Codex got a Vivado environment running on an unsupported Mac in just a few hours. I can now compile overlays for PYNQ, and all Verilogs are entirely written by Claude.

The sky really is the limit. My biggest challenge now isn't tooling—it's being disciplined enough to challenge my own ideas. Nothing feels out of reach anymore. Almost everything seems achievable if I'm willing to put in the time. The real tradeoff is finding that time among everything else life demands.

AI is my unpaid (I don't have subscription) research assistant and analysis tool.

A part time job maintaining a network has given me reason to use it. I don't recall exact command line syntax but AI does. Yeah it gets it wrong sometimes but for the most part especially working across multiple operating systems it gets me in the ball park to continue.

Got log files that require analysis, throw it to chatgpt and it evaluate and provide summary. Here is the output from five network iperf3 files, what do you see? response, the output appears to be limited to 100MB. Oh yeah, now I remembers that bottleneck.

My years remaining in supporting IT are dwindling. Learning MS365 is not necessary if all you need to do is fix a DNS records. Research assistant, were do I find that change area in the MS admin tools? Yep, right on the mark. 

What is the difference between tracert & traceroute? A detailed description without  opening a man page. A CLI output that took me years to master but still don't remember that syntax. YES!

I've also used it to write Arduino code. I am never very good at it manually. I know what I want and if I can explain it correctly AI gives me a baseline to tweak. I will be damned if I can explain the loops but it works. The challenge is to provide the detailed explanation to make it happen.

It has drafted policy documents. Create a policy for credit card use in non-profit. Done, ten bullet points. I can work with that as a start.

Stumbled upon this PDF parser that helps AI process documents. 
https://github.com/opendataloader-project/opendataloader-pdf

Have only tried it on a few documents so far, but things that were non-sesnse to the AI previously are actually interpreted much more accurately after parsing, when using the json file output. 

I don't know if it's any comfort at all, but I'm not the only one on that boat. Here is a great piece: human-in-the-loop.bearblog.dev/.../

This is the part of AI, I don't find helpful. Investors and founders who think they've found the next unicorn.  As for fl*x trying to flex... Good luck to them. Let them burn their cash...

No, not really. – Not yet.

Why? – I hate data theft and identity theft.

It's so much easier with any kind of AI.

Just record a few words you say, and you can use those stolen words to conduct business or commit other crimes.

And what if your face was also recorded? Well, then you're done for – the damage done can never be repaired.

Then there are arguments here – "Help with data analysis, help with creating tables, help with writing programs, etc..."

Help with writing programs?  – Well, then you might not be able to check if the AI ​​is programming a "black box," and

suddenly you have a nasty virus in your program that will never be detected. Cool! 

Okay, okay… one more question: Can AI really distinguish between true and false information?

“Yes, AI has a lot to learn.” – Sorry, but that argument is a bit dubious.

Yesterday, the news on TV reported that AI is already being used in pharmaceutical companies.
But the authorities are currently prohibiting its use because the AI ​​is mixing and suggesting poisons.

So, if no one checks what the AI ​​is actually doing, then medications will be produced that don't cure the sick, but instead cause even worse problems…

I really don't want to discuss this argument to death here.

So, YOU - the human must have every time the control over the AI.

The argument "faster" is not really the 1. argument.

Security and live is the domain.

Regards

Gerald

---