As someone who does both hardware design (mostly high precision analog electronics) and software engineering, the difference between both fields in AI capability is striking. Models like Fable and Astra can provide lots of interesting insights into the design of specialty electronics when prompted correctly. Yet they can hardly put that knowledge to work if instructed to generate a schematic or lay out a PCB.
Of course, for readers here with a solid grasp on how LLM's operate it's not exactly surprising. Still, I feel this is exactly the kind of thing that drives home that the way we reason about (general) intelligence does not map cleanly to LLM operation, no matter how advanced the current models are.
If anything, it is surprising that the LLM is so bad at stealing a ready-made RP2350 design! There are plenty of easy-to-find tutorials like [0] and [1] that handhold you through the entire design process, and a simple search will give you dozens of design files for open-source boards.
At this level of board complexity you almost have to try to make it completely non-functional, so although it is nice to see it in practice I am not exactly surprised that a LLM with a decent bunch of assistance managed to do it. But there's quite a gap between the equivalent of poorly copying a "Hello World" from StackOverflow and making actual production-ready boards.
[0]: https://pip-assets.raspberrypi.com/categories/1214-rp2350/do...
[1]: https://jlcpcb.com/blog/how-to-design-layout-with-rp2350
You have to be more specific about what you mean by design. The models are excellent and the netlist and the components along with tradeoffs.
Design means PCB routing. A schematic is the PCB equivalent to a Mermaid diagram.
It’s the training data. Finished work isn’t training data because parts are sooooo different, but maybe if the finished designs were properly labeled into the training pipeline it might work? Or the iterative work itself might need to be the training corpus.
i noticed the same treying to just generate 3D models using AI, i wonder if a well trained very specific world model would be better for the task
Maybe it's not properly benchmarked.
A standard benchmark using kicad to create or fix electronic designs would go a long way in motivating and measuring improvement.
There's eebench, but maybe it's not exactly what you were thinking