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.
AI is exceptionally good at RP2040/RP2350/ESP32-based boards, we've ordered ~15 unique designs, mostly RP2040-based that were designed and autorouted with AI using tscircuit and usually OpenAI models.
Currently the major mistakes that AI makes: 1) Improperly rotated connectors (backwards USB port) 2) Endless routing loops for complex boards 3) Really huge/unreadable single page schematics
I'm extremely bullish on AI for PCB design, but there is a huge, huge tooling gap. We basically need to soup up all our collective design rule checks so that AI can work in corrective loops. It needs to check for things like "opamp schematic layout conventional" and "trace width closely matches reference design for chip". One thing that makes me super excited is supply-chain resiliency when chip/subcircuit-swapping becomes super easy
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.
I've been working on several PCB projects, and agents are very helpful for automating many parts of the workflow. They're excellent for schematics. Complex routing, though, is still a human's job for now; I've watched Freerouting fall on its face many times. If you're doing six-layer high-speed DDR routing, AI won't solve that for you today.
I'm curious if this even works at high speed though I'm seeing what appears to be length-matched traces. Is the tool using an EM simulator in the process?
Was non-trivial (e.g. days and days of cumulative work) but worked with Codex to build up a parametric generator where I could adjust the star size, number of rings, leds per ring. It did all the LED-to-LED routing, panel layout with mouse bites, placed decoupling caps. At the time Codex wasn't as good at spatial reasoning, I expect that Astra would do this much faster.
4 x layer PCB with GND and 5V inner layers. WS2816B LEDs (not available in strips yet). 3D printed a frame so that two stars could be opposite each other (video shows only one side). ESP32 S3 on a 3.3v to 5v shifter board sandwiched in between. ESP32 wifi antenna pokes through and sits on a keep out section. Received control signals from a central controller via ESP Now.
In our company, we are doing high speed PCB designs with FPGAs, memories, ADCs, DACs etc. We are not yet using AI much in our PCB design workflow because our designs seem to be too complex for the AI solutions we have tried. Are there any areas where people are successfully incorporating AI to their PCB design workflow, even with complex, high speed designs? Schematic/layout review? Routing? Footprint creation?
Why wouldn't an LLM be able to do high speed designs? Are you inventing new types of design rules for every board? Or is it simply because the automatic layout tools are not up to the task yet?
The feedback loop is too long, it's not like a software development harness that can direct the model to write some tests for complex behavior, and iterate until they pass.
There a lots of things that can be subtly wrong, and it gets much worse as your signal frequency increases, or you have to deal with high power/radio/etc. The model (agent) has no way to validate its work.
It is just that the current tools out there seem to work for small or hobbyist boards currently. We couldn't apply them on our complex board designs. But maybe you had a different experience? I am curious on what kind of ways big companies are incorporating AI in their PCB design workflows currently.
> Upon Claude's compilation of the project, I asked it to prepare the required files for JLCPCB, and tell me what to select on their GUI
It would be an interesting step to get Claude (or other LLM) to do order the PCBs itself on the JLCPCB website.
We're nearly at the stage where you can describe what a device should do, let it design the circuit (using available components on JLCPCB), design a PCB and enclosure then order PCBs, assembly and 3d printed enclosure and get it sent to you for final assembly.
But: the PCB itself is not looking beautiful in terms of routing and a lot of schematic details yet. Horrible to to be more honest. But that's a minor issue.
I'm not overly shocked. 90% of creating PCBs is copying reference schematics from vendor datasheets and connecting them. That's LLM territory. This mostly failed for now as they just did pdf2text and killed all documents.
The last 10% are brutal (HF, EMC, high frequency multi-layer designs, etc.). But I managed LLMs to automatically create ngspice models using vision onto datasheets.. lol
It just sends the PDF to Gemini Flash and asks it to extract semi structured data (semi structured in the sense that it's not fully structured extraction, but its a JSON file to feed into the next step of the LLM design process).
Works pretty well in general, though I should update to the latest Gemini version.
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.
Machine routing is very old thing. Early Eagle did it fine. I worked 12 years ago on 4 layer pcb initially routed with Eagle. Layers with horizontal and vertical wires were unusual at the time, but the board worked fine. It’s the connector placement according enclosure design files, part management/availability/obsolescence and design for test what needs reliable automation. The tool may not hallucinate here.
I researched a bit on how to manufacture PCBs at home, and i came across Voltera V one. The price is indeed way too high, but I am hoping somebody will make a foss version of it. There are already pretty interesting foss PnP machines.
For a 2-layer pcb I would either go chemical etching or a small routing cnc. Chemical etching kits , uv lamps and blank pcbs are not "too" expensive, but it's more of a professional setup than a hobbyist. Just think of all the PCBs yoi can send to manufacture (and with 4 layers and more) for the price of even getting half the stuff.
I saw the board cost, but how about the total cost in tokens to design it? Can empathise with the good and meh feeling you got from it. For me I try to stick with manual struggles with new concepts which grows my brain and allows me learn something new. But its impressive what LLMs can do.
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!
It's pretty interesting to watch an LLM drive ngspice. ... clearly an area that could use more training, but good enough to make it less annoying for me to use.
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.
> That said, I was not feeling much of an accomplishment, rightfully so. I used to enjoy the learning and the struggle that came with it.
> Yes it does suck that the joy we had while building has been sucked out of us
> For my hobby projects tho, I will continue to YOLO it and build stuff fast without necessarily knowing about the details.
I don't understand this conclusion. LLMs sucked the joy out of it, didn't feel rewarding, you're forced to use them at work, so you decide... just keep doing it at home, too? Why not actually learn to build a proper PCB knowing you can do much better than this horrible (despite being dead-simple) board that only got there with the help of your expert colleague's guidance anyways? Don't succumb to utter laziness and deny yourself the satisfaction of learning new skills, that's letting the thing rot your brain out worse than any TV ever did.
You are right. What I meant in my conclusion was more about : "I will use these tools to generate parts of a project I don't necessarily care about." I already started learning about proper PCB design, and I am actively fighting against choosing an LLM instead of grasping a concept fully. It's something I am struggling with, but all this is new for most of us. Hoping I will find a good balance at the end. Thanks for your remarks :)
Fair enough. I think there is great value to using LLMs to assist in learning new hobbies, helping lessen the friction, and doing research. I just think it's sad if you let them replace your thinking completely instead of augmenting it. Cheers for the human-written blog post, I did enjoy reading it.
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
AI is exceptionally good at RP2040/RP2350/ESP32-based boards, we've ordered ~15 unique designs, mostly RP2040-based that were designed and autorouted with AI using tscircuit and usually OpenAI models.
Currently the major mistakes that AI makes: 1) Improperly rotated connectors (backwards USB port) 2) Endless routing loops for complex boards 3) Really huge/unreadable single page schematics
I'm extremely bullish on AI for PCB design, but there is a huge, huge tooling gap. We basically need to soup up all our collective design rule checks so that AI can work in corrective loops. It needs to check for things like "opamp schematic layout conventional" and "trace width closely matches reference design for chip". One thing that makes me super excited is supply-chain resiliency when chip/subcircuit-swapping becomes super easy
Major mistakes that AI makes:
AI is exceptionally good at RP2040/RP2350/ESP32-based boards.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!
Bought to us by Fable 5 and the more-experienced colleague who foresaw potential issues.
I've been working on several PCB projects, and agents are very helpful for automating many parts of the workflow. They're excellent for schematics. Complex routing, though, is still a human's job for now; I've watched Freerouting fall on its face many times. If you're doing six-layer high-speed DDR routing, AI won't solve that for you today.
However checkout EDA tools https://github.com/l3wi/claude-eda/ https://github.com/drandyhaas/KiCadRoutingTools
That layout does not look very good:
https://github.com/drandyhaas/KiCadRoutingTools/blob/main/do...
I'm curious if this even works at high speed though I'm seeing what appears to be length-matched traces. Is the tool using an EM simulator in the process?
This LED star PCB was 95% generated by Codex: https://www.youtube.com/shorts/8a52LjovhcE
Was non-trivial (e.g. days and days of cumulative work) but worked with Codex to build up a parametric generator where I could adjust the star size, number of rings, leds per ring. It did all the LED-to-LED routing, panel layout with mouse bites, placed decoupling caps. At the time Codex wasn't as good at spatial reasoning, I expect that Astra would do this much faster.
4 x layer PCB with GND and 5V inner layers. WS2816B LEDs (not available in strips yet). 3D printed a frame so that two stars could be opposite each other (video shows only one side). ESP32 S3 on a 3.3v to 5v shifter board sandwiched in between. ESP32 wifi antenna pokes through and sits on a keep out section. Received control signals from a central controller via ESP Now.
In our company, we are doing high speed PCB designs with FPGAs, memories, ADCs, DACs etc. We are not yet using AI much in our PCB design workflow because our designs seem to be too complex for the AI solutions we have tried. Are there any areas where people are successfully incorporating AI to their PCB design workflow, even with complex, high speed designs? Schematic/layout review? Routing? Footprint creation?
Why wouldn't an LLM be able to do high speed designs? Are you inventing new types of design rules for every board? Or is it simply because the automatic layout tools are not up to the task yet?
The feedback loop is too long, it's not like a software development harness that can direct the model to write some tests for complex behavior, and iterate until they pass.
There a lots of things that can be subtly wrong, and it gets much worse as your signal frequency increases, or you have to deal with high power/radio/etc. The model (agent) has no way to validate its work.
It is just that the current tools out there seem to work for small or hobbyist boards currently. We couldn't apply them on our complex board designs. But maybe you had a different experience? I am curious on what kind of ways big companies are incorporating AI in their PCB design workflows currently.
Yeah I've built about a dozen boards using GPT-sol 5.6 and my favorite is my Pi CM5 carrier board, to replace a laptop board in an old Sony Vaio p laptop. https://x.com/kanwisher/status/2089629411364745351?s=20
Did it work ?
It's funny, I made a CM5 board for my N40L not too long ago https://www.youtube.com/watch?v=AzW4rRsD8Uk
Was sol5.6 good enough to do the actual HDMI / PCI-E and USB 3 Routing ?
> Upon Claude's compilation of the project, I asked it to prepare the required files for JLCPCB, and tell me what to select on their GUI
It would be an interesting step to get Claude (or other LLM) to do order the PCBs itself on the JLCPCB website.
We're nearly at the stage where you can describe what a device should do, let it design the circuit (using available components on JLCPCB), design a PCB and enclosure then order PCBs, assembly and 3d printed enclosure and get it sent to you for final assembly.
There's a plugin for Kicad https://github.com/bouni/kicad-jlcpcb-tools
If it's functional: that's a breakthrough.
But: the PCB itself is not looking beautiful in terms of routing and a lot of schematic details yet. Horrible to to be more honest. But that's a minor issue.
I'm not overly shocked. 90% of creating PCBs is copying reference schematics from vendor datasheets and connecting them. That's LLM territory. This mostly failed for now as they just did pdf2text and killed all documents.
The last 10% are brutal (HF, EMC, high frequency multi-layer designs, etc.). But I managed LLMs to automatically create ngspice models using vision onto datasheets.. lol
Shameless plug for the slop coded CLI I use in my workflow: https://github.com/akiselev/datasheet-cli
It just sends the PDF to Gemini Flash and asks it to extract semi structured data (semi structured in the sense that it's not fully structured extraction, but its a JSON file to feed into the next step of the LLM design process).
Works pretty well in general, though I should update to the latest Gemini version.
That's pretty cool! How reliable has the footprint extraction been?
I tested it With Gemini 2.5 Flash. Works great.
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.
Yeah, we're still very far from a machine routing PCBs. Nice project, though, looks pretty fun to do it with an AI.
Machine routing is very old thing. Early Eagle did it fine. I worked 12 years ago on 4 layer pcb initially routed with Eagle. Layers with horizontal and vertical wires were unusual at the time, but the board worked fine. It’s the connector placement according enclosure design files, part management/availability/obsolescence and design for test what needs reliable automation. The tool may not hallucinate here.
I researched a bit on how to manufacture PCBs at home, and i came across Voltera V one. The price is indeed way too high, but I am hoping somebody will make a foss version of it. There are already pretty interesting foss PnP machines.
For a 2-layer pcb I would either go chemical etching or a small routing cnc. Chemical etching kits , uv lamps and blank pcbs are not "too" expensive, but it's more of a professional setup than a hobbyist. Just think of all the PCBs yoi can send to manufacture (and with 4 layers and more) for the price of even getting half the stuff.
Can it shop autonomously on LCSC? Is it allowed to?
I saw the board cost, but how about the total cost in tokens to design it? Can empathise with the good and meh feeling you got from it. For me I try to stick with manual struggles with new concepts which grows my brain and allows me learn something new. But its impressive what LLMs can do.
I am on Max 5x plan, so there were no additional costs. But the would-be API cost was around 450 usd.
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!
It's pretty interesting to watch an LLM drive ngspice. ... clearly an area that could use more training, but good enough to make it less annoying for me to use.
The machine designed the circuit board in two hours, which was impressive, and the human understood it in never, which is becoming traditional.
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.
Yes indeed. I paid 7.5 euros extra per e-ink display. So it was 33.5 euros per board.
> That said, I was not feeling much of an accomplishment, rightfully so. I used to enjoy the learning and the struggle that came with it.
> Yes it does suck that the joy we had while building has been sucked out of us
> For my hobby projects tho, I will continue to YOLO it and build stuff fast without necessarily knowing about the details.
I don't understand this conclusion. LLMs sucked the joy out of it, didn't feel rewarding, you're forced to use them at work, so you decide... just keep doing it at home, too? Why not actually learn to build a proper PCB knowing you can do much better than this horrible (despite being dead-simple) board that only got there with the help of your expert colleague's guidance anyways? Don't succumb to utter laziness and deny yourself the satisfaction of learning new skills, that's letting the thing rot your brain out worse than any TV ever did.
You are right. What I meant in my conclusion was more about : "I will use these tools to generate parts of a project I don't necessarily care about." I already started learning about proper PCB design, and I am actively fighting against choosing an LLM instead of grasping a concept fully. It's something I am struggling with, but all this is new for most of us. Hoping I will find a good balance at the end. Thanks for your remarks :)
Fair enough. I think there is great value to using LLMs to assist in learning new hobbies, helping lessen the friction, and doing research. I just think it's sad if you let them replace your thinking completely instead of augmenting it. Cheers for the human-written blog post, I did enjoy reading it.
Thank you :) That means a lot to me.