Can AI Now Design PCBs That Just Work?

With the recent release of its GPT-6 Astra model, OpenAI explicitly pushed the claim that it is capable of designing complete circuit boards in KiCad, starting from a provided schematic and outputting a fully routed PCB that theoretically could be sent off to be manufactured. This of course raises the question whether this is just a nifty party trick that works under strictly controlled conditions like most auto-routing tools, or whether there’s more to it. In a recent [EEBench] blog post, OpenAI’s claims are put to the test. Back in 2024, we looked at how LLMs handle circuit board design, starting with the schematic. The conclusion was that you might as well just do it all by hand. Tracking progress here, [EEBench] is an electrical engineering agent benchmark that tests how effective these so-called AI agents are at performing useful hardware engineering work. As their methodology already makes clear, creating a populated and routed PCB from a schematic is just one step of many. Consequently, GPT-6 Astra scores 69.3% (+/- 10%) on their benchmark, roughly in the same ballpark as Claude Opus 5, albeit cheaper and faster. It should be noted that [EEBench] is run by the developers behind Atopile, which is a code-based system for creating PCBs in KiCad with a strong focus on use by such AI agents. Despite this, the blog post makes it clear that there is still a long way to go before hardware engineers can receive their pink slips and take up another profession. After all, while vibe-coding a quick prototype PCB can be a nice shortcut, for a one-off project PCB, most boards are expected to handle all the edge cases, be fully validated by multiple independent engineers, and have passed stringent testing before committing to a production run. (Top image: from OpenAI GPT-6 Astra KiCad video) I’ve vibe coded 3 pcbs that I’ve sent off to manufacture now using Fable. Each one has come back working first time. It was done via just pointing claude at empty kicad files with no skills/mcp etc. It usually takes me at least one re-spin or some hand rework for boards I’ve done by hand. Complexity wasn’t huge for the boards, but they also weren’t totally trivial. The routing can look a little wonky (which started me down the rabbit hole of making a rust fork of freerouting), but its really wild how well it works. It’s worth noting that programs to route connections appropriately, taking into account coupling and Delta I and the rest, have existed for decades in the serious chip design world. AI can only do an inferior job to algorithms which have semantic awareness. This is definitely a case of when the only fastener you have is a nail, every tool looks like a hammer. They couldn’t do it at all last month. They will be better than all that stuff in a little while. People always critique this stuff as a snapshot in time instead of as a continuing trend. I completely agree, these ai things will improve, and quite rapidly too. But on the other side of the equation there there are the huge ai companies, investing a trillion dollars, wreaking havoc with the economy while making huge promises that simply are not true yet. And that turns them into lying bastards. And that the “leaders” of these companies are the scum of the earth is not hard to see. Just have a look at the many billions that the sugar mountain guy pays in settlements, and for a big part that is to avoid having to testify under oath. The dark side of me sometimes thinks all these “hallucinations” of the ai things are on purpose. They are learning to be better liars, and learning how to mislead more people more effectively. People who have gotten ridiculously rich by lying and cheating are now teaching monstrous machines to do the lying and cheating for them. One of the latest scandals is that he set up a system to guide teenagers into insecurity and depression, just to make some profit on advertisements for “beauty products”. It’s now gotten to a point that these machines (and their companies) are evolving at a much greater speed then society can handle, and he main goal it’s used for is to extract more of the worlds wealth into an ever smaller group of sociopaths who already have more money then they can handle. Some people are going to be like seriously psychologically broken when they finally have to face the possibility that it’s not all hype… so many have invested a huge amount of their psyche into the anti-AI op (which was funded by like four or five billionaires (who all work in AI lol)) Hi, professional hardware developer here, For software developing our society accept, for some unknown reason, that software never works well. There is always an update, there is pressure to the user to use updates, user put pressure to the developer for update. Strange if you think of it? I guess it was different when software was in the mask rom of your devices. :-) For hardware developing it is different. If I made an issue and the device start to fail 1-2 year later my boss has to lie to the customer (it is your problem!), had to exchange the hardware, the customer go to court, newspaper write about it. There is much higher pressure to the KI do deliver 100% solution. For software the people accepting 97%. Agree that the bar is higher for hw than sw. But this is because: 1) HW cannot be changed once shipped to customers. The cost to recall and rework/replace could sink a company. 2) HW is several orders of magnitude simpler than a modern software package. It’s simply not feasible to even find all the sw bugs, let alone fix them. A well designed pcb can be made extremely rebust. That said, I don’t think we ACCEPT buggy sw exactly. We’re quick to complain and report bugs. And we are very insistent that the ‘vendors’ fix those bugs immediately – Especially when we haven’t paid anything for them – ask any FOSS team. It’s more that the sw vendors set their release metrics below 100% for the reasons given above. If the bar were anywhere near 100% companies would never ship anything. Also, re the 97% sw quality metric. For small apps, maybe. But any feature rich app (like kicad for example) it will be much lower. The release requirements might be no known ‘show stopper’ bugs, and overall bug count declining and less than some percentage. Note also that complex electronics devices (such as cpu’s) have similar constraints to sw. Though they are still simpler, they are complex enough that perfection is unrealistic. The ‘no show stopper’ bugs target is true there too. The onus on validation to identify the show stopper bugs is just much higher because it can’t be changed once it goes to the fab. I fail to see how a CPU is less complex than (most) software. Especially if considering the necessary production processes tightly coupled to the design. An article about AI. I am sure the usual HaD commenters will be very normal and rational about this. Variations on all of the existing designs out there should work fine, but edge cases could be problematic. It is like the “octopus in a spacesuit” problem for image generators they all failed badly for over a year because there was nothing close enough in the training data. Google nano banana was the first to pass that test. “… vibe-coding a quick prototype PCB can be a nice shortcut …” indeed. 😁 I believe from a perspective of caution we must ask if LLM design of power supplies could be accepted as is? Like if we just build what it outputs, it will be totally safe right? Would you use an LLM to determine if a foraged mushroom is toxic or not? Your product still needs to be tested for compliance with safety and EMC regulations. This hasn’t stopped manufacturers though from e.g. bringing those dangerous usb chargers to the market. As creating electronics junk becomes accessible to more people, we can expect that more of it will be coming.