I think using the vision decoder baked into modern LLMs is the way to go. Have the LLM iterate; make sure it can assert placement qualities and understands the hard requirements. I think it can be done.
Dont know about LLM, but AI in general isnt such a stupid idea as one might think and Chinese are particularly well positioned to take advantage.
Take for example something like XinZhiZao (XZZ), ZXW, Wuxinji, diyfixtool. They have huge databases with pictures, diagrams and boardviews of pretty much every phone, laptop and graphics card. With all this data you could build AI system ripping of^^^^^ "suggesting" routing for your design based on similarity to stole^^^training data. That way you start with layout that worked in devices shipped by the millions.
This could be build in stages, starting witch much weaker system trained on just pcb pictures + layer count. This should be enough to suggest ~optimal initial chip placement for classical auto-router.
author here: I think synthetic data, generated by ~brute force iteration with LLMs, with every DRC analysis imaginable and more, will yield a more consistent/usable/larger dataset than any existing dataset. It's a mistake to put too much weight in anyone's existing data. This is why we work hard to make algorithms that LLMs can use, because they have emerging spatial capabilities that excel when coupled with detailed analysis.
They do a have a vision decoder like many other LLMs, so in theory it should be able to write the positions textually, then call a render command, then look at the rendered bitmap. I's all very opaque though; I'd love a visualisation of the latent space data that it's converting the image to. I found that very long vertical images throw Opus off completetely for example. It's very interesting to experiment with this. Let it play with placement and let it call a render command. Then let is describe in detail what it sees. I'll be looking into this a lot this year. Maybe there will be niche models that will be smaller but have better vision capabilities then Opus. A world where one model rules would be incredibly depressing (kinda like what we saw with some software companies since the 90s).
Take for example something like XinZhiZao (XZZ), ZXW, Wuxinji, diyfixtool. They have huge databases with pictures, diagrams and boardviews of pretty much every phone, laptop and graphics card. With all this data you could build AI system ripping of^^^^^ "suggesting" routing for your design based on similarity to stole^^^training data. That way you start with layout that worked in devices shipped by the millions.
This could be build in stages, starting witch much weaker system trained on just pcb pictures + layer count. This should be enough to suggest ~optimal initial chip placement for classical auto-router.