KiCAD-GPT is here
I’ve been experimenting with a Beta version of KiCAD that adds an AI based interface to the a popular electronics design application.
It has been trained on example circuits found in datasheets, old magazines and discussion forums. Rather than needing to manually design circuits, its easy to use text based interface allows you to simply request the circuit you desire. So I put it to work creating the classic “hello world” of the electronics world, the blinking LED.
The results were quite remarkable, a fully operational circuit was created in response to a simple prompt. I have quite an old PC with an i3 processor and it took about 4 hours to compute the new circuit. But with a newer processor you should see better performance. And it did take a few attempts to get the AI to understand exactly what I was asking for.
Unfortunately auto-routing is not currently working and it does have a tendency to prefer older components. So I’m not sure it is going to replace the electronics design engineer just yet.
But it is just a first version and I’m sure it will improve with future releases.
Posted in: Electronics
Tags: APR1
Metal sculpture in brass, mounted on an 1800s French Coin
was this April fools?
Hi Edward, yes I’m afraid that the articles marked APR1 are there for your amusement rather than providing facts.
Bad idea with this type of fools…sad.
Hey Greg, I appreciate your feedback, I add an APR1 article most years so you should be able to avoid the future ones.
I think AI is going to be one of the biggest hype bubbles and it could burst soon, hence this article is making a mockery of all those companies who have just added AI without thinking “why”. And also a slight dig at those who always suggest a 555 is the answer to all problems.
KiCAD is open source so you can always add in your own AI add-in if you really think that might be something that would be useful.
Hey that looks really promising! Did you see flux.ai ? Maybe some of their stuff can be implemented into KiCAD, too.
greetings,
André
Sorry Andre this was a spoof article from April 1st, playing on the “should have used a 555” comments we get on forums. Flux.AI has a lot of potential. It has the disadvantage of being online which could lead to users being locked out of their designs by aggressive business tactics or risking IP leaking into the hands of competitors.
I’m a developer that works on AI integrations like this. Whenever this was originally written, LLMs (AI) was interesting, but still bad at everything. Current gen AI is VERY capable of assisting even beginners with PCB design. In fact, I’m currently learning PCB design as a hobby, and almost everything I know I’ve learned from ChatGPT, or something it sent me to. It’s also remarkably good and reading images of PCBs, better than I am anyhow.
… and this technology will be 10x better in two more years. Honestly, everyone should stop all other development and be all over this.
AI is going to enable millions of people to do things they couldn’t do before. Including electronics design. It’s interactive knowledge to the masses.
The date of the post gives a strong hint to the satirical nature. Although many of the points still stand. Training data for circuits is more sparse than for code so that will constrain the components used. As an LLM does understand electronics but is repeating patterns, by itself won’t necessarily spot that in my example a key pin is not connected. This can be fixed by passing the output through the electronic equivalent of a Lint tool e.g. a design rule check. I’m sure things have improved in the last couple of years but I suspect this area will develop more slowly than areas like natural language, images and code.