Comment by embeng4096
2 days ago
Idk, I'd like to see AI design a schematic and lay out a board. Even the occasional "Show HN" that combines AI+PCB stuff is about analysis and checks. It doesn't seem like anybody is even near any kind of ECAD using AI. People still complain about autorouters. Even frontier models need architecting and guard rails in software to avoid spaghetti, it seems like an even harder problem to get them to choose the correct parts, create accurate footprints and symbols, and wire those all up together.
"This is wrong, fix it" + recompile, can happen twenty times in ten minutes, but "we discovered the layout is wrong, fix it" is precluded by the cost and time of a new board spin.
AI + text (code) seems like a good match but (E)CAD seems a lot harder to interface AI with. If I'm wrong, I'd like to share that with the EEs on my team, though.
> Idk, I'd like to see AI design a schematic and lay out a board.
It's being worked on, and as the other commenter said, it's a data problem not a fundamental limitation. That's what's scary to me.
That said the SOTA tools in this space still suck, so it's not here yet :)
Its just a data problem, there's nothing special about the task. It just happens that there's a lot of data about programming already.
PCB routing is still a murderously-hard NP problem, AI or no AI. Not least because a fully-automated schematic-to-PCB flow has to be driven by component placement, which in turn is driven (or at least influenced heavily) by requirements further up the chain.
I'm sure it'll happen -- if you can defeat the world Go champion you can certainly route traces -- but so far it seems to be an exception to the bitter lesson, where proprietary hand-tuned algorithms are still on top. Given the lack of large quantities of PCB CAD files that can be appropriated for training, it might be one of those things that has to be self-learned by the AI, again reminiscent of how Alpha Zero worked.
Everything is an exception to the bitter lesson until you get enough data.