I was puzzled a bit, then realized they only handle schematics. Saying "PCB schematics" is weird.
A schematic is just a representation of a netlist, something where text is more than fine since the graphical form is only for human consumption. An LLM is actually a pretty good fit to cross-reference datasheets and netlists.
Would it be actual PCB layout I would be skeptical as LLMs are quite poor at anything spatial. For schematics however, it could work quite well as a double check.
Some anecdata: This weekend as a lark I asked Claude Code to design a (fairly simple) analog circuit and simulate it in LTSpice to verify. It did three edit-simulate-fix cycles and to my surprise ended up with something that seemed pretty sane.
That said, schematics (as opposed to netlists) don't seem to be a practical I/O format yet. It did generate a KiCad schematic file when asked, but it was pretty bad (penguin on a bicycle level).
Anyway, somehow there does seem to be some electronic tools training happening, becuase I tried this maybe a year ago and it was pretty hopeless.
Ive had that argument with many of the schematic PCB ai startups. Online open source schematics and PCB designs are awful training sets by large. There are some gems out there publicly but that's gems in a sea of sand. Far different than training a LLM on all the published books of the world.
I was puzzled a bit, then realized they only handle schematics. Saying "PCB schematics" is weird.
A schematic is just a representation of a netlist, something where text is more than fine since the graphical form is only for human consumption. An LLM is actually a pretty good fit to cross-reference datasheets and netlists.
Would it be actual PCB layout I would be skeptical as LLMs are quite poor at anything spatial. For schematics however, it could work quite well as a double check.
I don’t buy that it would be very good at reliably finding problems in schematics either. There’s no big dataset on the internet to train on
Some anecdata: This weekend as a lark I asked Claude Code to design a (fairly simple) analog circuit and simulate it in LTSpice to verify. It did three edit-simulate-fix cycles and to my surprise ended up with something that seemed pretty sane.
That said, schematics (as opposed to netlists) don't seem to be a practical I/O format yet. It did generate a KiCad schematic file when asked, but it was pretty bad (penguin on a bicycle level).
Anyway, somehow there does seem to be some electronic tools training happening, becuase I tried this maybe a year ago and it was pretty hopeless.
4 replies →
Ive had that argument with many of the schematic PCB ai startups. Online open source schematics and PCB designs are awful training sets by large. There are some gems out there publicly but that's gems in a sea of sand. Far different than training a LLM on all the published books of the world.
Not sure how familiar you are with KiCad but its project files are just text. With a little bit of rearranging you can get things like:
"Component U1 is 74HCT02, has 14 pins"
"Component R2 is 4.7K resistor, has 2 pins"
"Signal +5V connects U1 pin 14, U2 pin 9, R1 pin 2, R7 pin 1...."
"Signal /RESET connects U1 pin 2, U4 pin 3, U7 pin 8...."