Maybe I'm missing it, but I don't see how long it took to create these, or how much they cost (in tokens.)
PCB layout is a weird problem. As the article says, you can get halfway through before realizing you need to re-do everything, and there's no way to predict that from the outset. It's an inherently self-bottlenecked process.
There is also a lot of judgement that goes into it, that I'm not sure could ever be fully covered with AI. You can never hit 100% perfection on every metric (crosstalk, parasitics, DFM, BOM optimization, trace length matching, etc.) so determining where and by how much these things can fall short of optimal is very much a part of the process, and seems like it would be an overfitting nightmare for an AI.
There are also simply decisions that have no empirical right or wrong answer, but are still critical to the system functioning. What frequency do you want to run your data busses at? How much power can it draw? How long does it need to last in standby? How hot can it get?
I agree but I think theres a wide swathe of hobbyist design space where you don't have to think about this stuff too hard.
I've made a few simple designs, e.g. 2 layer < 10 mhz boards, cc or codex can few shot them using kicad and just scripts for layout. The bit where I was impressed was where it pulled the datasheet and made a custom footprint for a relay, which was spot on. As someone who doesn't use the tools much it's a basic thing but it would have taken me a lot longer.
One silly issue I ran into in my workflow was not taking into account the WHOLE package (not just footprint), resulted in a cap package mechanically interfering with an IC, so its worth doing renders. I bodged around it.
A lot of stuff hobbyists want to do is just gluing together reference designs into combos.
Maybe I'm missing it, but I don't see how long it took to create these, or how much they cost (in tokens.)
PCB layout is a weird problem. As the article says, you can get halfway through before realizing you need to re-do everything, and there's no way to predict that from the outset. It's an inherently self-bottlenecked process.
There is also a lot of judgement that goes into it, that I'm not sure could ever be fully covered with AI. You can never hit 100% perfection on every metric (crosstalk, parasitics, DFM, BOM optimization, trace length matching, etc.) so determining where and by how much these things can fall short of optimal is very much a part of the process, and seems like it would be an overfitting nightmare for an AI.
There are also simply decisions that have no empirical right or wrong answer, but are still critical to the system functioning. What frequency do you want to run your data busses at? How much power can it draw? How long does it need to last in standby? How hot can it get?
I agree but I think theres a wide swathe of hobbyist design space where you don't have to think about this stuff too hard.
I've made a few simple designs, e.g. 2 layer < 10 mhz boards, cc or codex can few shot them using kicad and just scripts for layout. The bit where I was impressed was where it pulled the datasheet and made a custom footprint for a relay, which was spot on. As someone who doesn't use the tools much it's a basic thing but it would have taken me a lot longer.
One silly issue I ran into in my workflow was not taking into account the WHOLE package (not just footprint), resulted in a cap package mechanically interfering with an IC, so its worth doing renders. I bodged around it.
A lot of stuff hobbyists want to do is just gluing together reference designs into combos.
There is a comparison plot that had model Electronics benchmark scores vs $ cost. All tested models were between $2000 and $5000.
I’ve been messing with GPT 6.1 sol PCB routing, but it always comes out really ugly and inefficient. It mostly just runs an autorouter.
Sol IS really good at part selection given loose parameters and JLCPCB’s inventory.
We need a good modeling language for general purpose electronics, some sort of UML for hardware.