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.
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'm still waiting on my order of 15 PCBs done with Sol - a mini-BT-keyboard that I hope will be comparable to the Psion 5 / Gemini PDA ones, using Kailh PG1316S switches. I still need to design and manufacture a housing (I do have a lot of hobbyist experience with this though), but the idea is to have a magsafe-connectable keyboard I can clip onto my Pixel and make use of the new Linux subsystem (or Termux) and do some productive stuff on the go. It's just been manufactured by JLC but not shipped yet. Funny enough, production itself was ~$70, then $80 for shipping, taxes and fees (tarriffs?).
Instead of directly generating Kicad, I chose to let it write "generators" in Python that output Kicad (S-Exps!) pcb and project - with mixed results - but I can generate a version with/without per-key Neopixels and customize certain parts of the board, including debug-stuff. I also _think_ this makes it easier to resume working on this in new sessions but not sure. Otoh, I never actually looked at the Python (vibed...) so it's still a lot of hard-coding.
I also landed on and used KicadRoutingTools after trying FreeRouting, combined with some minimal "auto-placing" code for larger ICs and such that are placed with certain constraints. My routing runs are way slower though - on the order of minutes.
I always wanted to try atopile - that might be a good match for AI/PCB work. The only thing that really held me back was the lack of routing.
> I always wanted to try atopile - that might be a good match for AI/PCB work.
Unfortunately all public open source work halted 6 months ago and they pivoted to a browser based IDE product to monetize: https://atopile.io/blog/atopile-v16
I followed them for a while before the AI craze started. It looked promising back then, but I wasnt willing to invest time and attention back then on something so new.
Would you be willing to share the full design files? Considering most of the PCB is well-trodden territory, it would be quite interesting to see the details what it actually came up with!
I built a project nearly identical to this one so knowing the level of complexity required, I'm not surprised it worked. AI is indeed raising the floor. Even simple hobby designs were once out of reach, but now we can "YOLO it and build stuff fast without necessarily knowing about the details" as the author puts it.
That said, without experience it's easy to pierce the veil of "simple circuitry" by accident, just like in XKCD: Tasks[1].
Looking at preview of the author's next project at the end of the article, I'm not so sure the same strategy will work. The board is probably two orders of magnitude more complex. Without knowing the circuit details, I see many red flags (randomly placed ics and decoupling, likely wrong length matching, random thick traces indicating lack of foresight on power distribution, backwards connectors, the probably unnecessary use of BGA footprints, etc).
This seems to fall into the same category I've seen with basically all agent driven software as well. That is to say, "Amazing that someone with no experience could do this, but not interesting at all in its own right". Call me when AI is laying out the next TSMC node.
I sent in a larger, slightly more complex board to JLCPCB over the weekend and it was $175 for 5 boards with components and assembly. Shipping, tariffs (ouch), and sales taxes added another ~$100 or so.
> For five fully assembled boards I paid 130 Euros, ordered the E-ink displays from a local shop
I built a project board nearly identical to this one a few years ago, and this seems about right. I don't think the displays themselves are included in the 130 euros. They can vary a lot in price depending on spec and supplier.
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
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'm still waiting on my order of 15 PCBs done with Sol - a mini-BT-keyboard that I hope will be comparable to the Psion 5 / Gemini PDA ones, using Kailh PG1316S switches. I still need to design and manufacture a housing (I do have a lot of hobbyist experience with this though), but the idea is to have a magsafe-connectable keyboard I can clip onto my Pixel and make use of the new Linux subsystem (or Termux) and do some productive stuff on the go. It's just been manufactured by JLC but not shipped yet. Funny enough, production itself was ~$70, then $80 for shipping, taxes and fees (tarriffs?).
Instead of directly generating Kicad, I chose to let it write "generators" in Python that output Kicad (S-Exps!) pcb and project - with mixed results - but I can generate a version with/without per-key Neopixels and customize certain parts of the board, including debug-stuff. I also _think_ this makes it easier to resume working on this in new sessions but not sure. Otoh, I never actually looked at the Python (vibed...) so it's still a lot of hard-coding.
I also landed on and used KicadRoutingTools after trying FreeRouting, combined with some minimal "auto-placing" code for larger ICs and such that are placed with certain constraints. My routing runs are way slower though - on the order of minutes.
I always wanted to try atopile - that might be a good match for AI/PCB work. The only thing that really held me back was the lack of routing.
> I always wanted to try atopile - that might be a good match for AI/PCB work.
Unfortunately all public open source work halted 6 months ago and they pivoted to a browser based IDE product to monetize: https://atopile.io/blog/atopile-v16
I followed them for a while before the AI craze started. It looked promising back then, but I wasnt willing to invest time and attention back then on something so new.
Good thing I didn’t
Thanks :(
That sounds like a cool project both for your pixel and the keyboard! Keep us posted how it goes!
Would you be willing to share the full design files? Considering most of the PCB is well-trodden territory, it would be quite interesting to see the details what it actually came up with!
I built a project nearly identical to this one so knowing the level of complexity required, I'm not surprised it worked. AI is indeed raising the floor. Even simple hobby designs were once out of reach, but now we can "YOLO it and build stuff fast without necessarily knowing about the details" as the author puts it.
That said, without experience it's easy to pierce the veil of "simple circuitry" by accident, just like in XKCD: Tasks[1].
Looking at preview of the author's next project at the end of the article, I'm not so sure the same strategy will work. The board is probably two orders of magnitude more complex. Without knowing the circuit details, I see many red flags (randomly placed ics and decoupling, likely wrong length matching, random thick traces indicating lack of foresight on power distribution, backwards connectors, the probably unnecessary use of BGA footprints, etc).
[1]: https://xkcd.com/1425/
This seems to fall into the same category I've seen with basically all agent driven software as well. That is to say, "Amazing that someone with no experience could do this, but not interesting at all in its own right". Call me when AI is laying out the next TSMC node.
Very cool! Can you reveal how much this custom board cost you?
I sent in a larger, slightly more complex board to JLCPCB over the weekend and it was $175 for 5 boards with components and assembly. Shipping, tariffs (ouch), and sales taxes added another ~$100 or so.
> For five fully assembled boards I paid 130 Euros, ordered the E-ink displays from a local shop
I built a project board nearly identical to this one a few years ago, and this seems about right. I don't think the displays themselves are included in the 130 euros. They can vary a lot in price depending on spec and supplier.