Testing Generative AI for Circuit Board Design
blog.jitx.com
blog.jitx.com
Edit: Better at chain of thought, long running agentic tasks, following rigid directions.
It's very hard to evaluate whether a model is better than another, especially doing it in a scientifically sound way is time consuming and hard.
This is why I find these types of comments like "model X is so much better than model Y" to be about as useful as "chocolate ice cream is so much better than vanilla"
This is a great way to describe what I've been feeling / experiencing as well.
* Connection between points
* Flows better
* Eyes don't start-stop as much
Different readable than the more flowing, conjunct readable than yours (which is the more typical use of it I concede)
Figures that any article written on LLM limits is immediately out of date. I'll write an update piece to summarize new findings.
I'm no expert in the matter, but for "holistic" things (where there are a lot of cross-connections and inter-dependencies) it feels like a diffusion-based generative structure would be better-suited than next-token-prediction. I've felt this way about poetry-generation, and I feel like it might apply in these sorts of cases as well.
Additionally, this is a highly-specialized field. From the conclusion of the article:
> Overall we have some promising directions. Using LLMs for circuit board design looks a lot like using them for other complex tasks. They work well for pulling concrete data out of human-shaped data sources, they can do slightly more difficult tasks if they can solve that task by writing code, but eventually their capabilities break down in domains too far out of the training distribution.
> We only tested the frontier models in this work, but I predict similar results from the open-source Llama or Mistral models. Some fine tuning on netlist creation would likely make the generation capabilities more useful.
I agree with the authors here.
While it's nice to imagine that AGI would be able to generalize skills to work competently in domain-specific tasks, I think this shows very clearly that we're not there yet, and if one wants to use LLMs in such an area, one would need to fine-tune for it. Would like to see round 2 of this made using a fine-tuning approach.
But I think there's also the bitter lesson to be learned here: many times people say LLMs won't do well on a task, they are often surprised either immediately or a few months later.
Overall not sure what to expect, but fine tuning experiments would be interesting regardless.
I have my own library of nuances but how would you even fine tune anything to understand the black box abstraction of an IC to work out if a nuance applies or not between it and a load or what a transmission line or edge would look like between the IC and the load?
This is where understanding trumps generative AI instantly.
Do you have a challenge task I can try? What's the easiest thing I could get an LLM to do for circuit board design that would surprise you?
Edit: no VSWR constraint. Can add that later :)
Edit 2: oh or design a board for a simple 100Mohm input instrumentation amplifier which knows what a guard ring is and how badly the solder mask will screw it up :)
On top of that, making an AI that can regurgitate simple textbook circuits and connect them together in reasonable ways is only the first step towards a much more difficult goal. More subtle problems in electronics design are all about context-dependent interactions between systems.
I have experience building boards in Altium and found it rather enjoyable; my own knowledge was often a constraint as I started out, but once I got proficient it just seemed to flow out onto the canvas.
There are some design considerations that would be awesome to farm out to genai, but I think we are far from that. Like stable-diffusion is to images, the source data for text-to-PCB would need to be well-labeled in addition to being correllated with the physical PCB features themselves.
The part where I think we lose a lot of data in pursuit of something like this, is all of the research and integration work that went on behind everything that eventually got put into the schematic and then laid out on a board. I think it would be really difficult to "diffuse" a finished PCB from an RFQ-level description.
How exact is exactly the same time? Current solver matches to under 10fs, and I think at that level you'd have to fab it to see how close you get with fiber weave skew and all that.
Do you have a test case for a schematic design task?
The point is there’s a methodology to solve these problems already. Is this better? And can it use and apply it?
I find I spend an enormous amount of time on boring stuff like connecting VCC and ground with appropriate decoupling caps, tying output pins from one IC to the input pins on the other, creating library parts from data sheets, etc.
There's a handful of interesting problems in any good project where the abstraction breaks down and you have to prove your worth. But a ton of time gets spent on the equivalent of boilerplate code.
If I could tell an AI to generate a 100x100 prototype with such-and-such a microcontroller, this sensor and that sensor with those off-board connectors, with USB power, a regulator, a tag-connect header, a couple debug LEDs, and break out unused IO to a header...that would have huge value to my workflow, even if it gave up on anything analog or high-speed. Presumably you'd just take the first pass schematic/board file from the AI and begin work on anything with nuance.
If generative AI can do equivalent work for PCBs as it can do for text programming languages, people wouldn't use it for transmission line design. They'd use it for the equivalent of parsing some JSON or making a new class with some imports, fields, and method templates.
AI can't do it itself (yet), and having it call the higher level functions doesn't save that much time...
Heh. This is very true. I think perhaps the thing I'm most amazed by is that simple next-token prediction seems to work unreasonably well for a great many tasks.
I just don't know how well that will scale into more complex tasks. With simple next-token prediction there is little mechanism for the model to iterate or to revise or refine as it goes.
There have been some experiments with things like speculative generation (where multiple branches are evaluated in parallel) to give a bit of a lookahead effect and help avoid the LLM locking itself into dead-ends, but they don't seem super popular overall -- people just prefer to increase the power and accuracy of the base model and keep chugging forward.
I can't help feeling like a fundamental shift something more akin to a diffusion-based approach would be helpful for such things. I just want some sort of mechanism where the model can "think" longer about harder problems. If you present a simple chess board to an LLM or a complex board to an LLM and ask it to generate the next move, it always responds in the same amount of time. That alone should tell us that LLMs are not intelligent, and they are not "thinking", and they will be insufficient for this going forward.
I believe Yann LeCun is right -- simply scaling LLMs is not going to get us to AGI. We need a fundamental structural shift to something new, but until we stop seeing such insane advancements in the quality of generation with LLMs (looking at you, Claude!!), I don't think we will move beyond. We have to get bored with LLMs first.
Is that true, especially if you ask it to think step-by-step?
I would think the model has certain associations for simple/common board states and different ones for complex/uncommon states, and when you ask it to think step-by-step it will explain the associations with a particular state. That "chattiness" may lead it to using more computation for complex boards.
> Is that true, especially if you ask it to think step-by-step?
That's fair -- there's a lot of room to grow in this area.
If the LLM has been trained to operate with running internal-monologue, then I believe they will operate better. I think this definitely needs to be explored more -- from what little I understand of this research, the results are sporadically promising, but getting something like ReAct (or other, similar structures) to work consistently is something I don't think I've seen yet.
There is such a mechanism - multiple rounds of prompting. You can implement diverse patterns (chains, networks) of prompts.
https://redwoodresearch.substack.com/p/getting-50-sota-on-ar...
What about the topic, it is impossible to synthesize STEM things not in the manner an engineer does this. I mean thou shalt to know some typical solutions and have all the calculations for all what's happening in the schematic being developed.
Textbooks are not a joke and no matter who are you - a human or a device.
If you are interested I highly recommend this + your favorite llm. It does not do everything but is far superior to some highly expensive tools, in flexibility and repeatability. https://github.com/devbisme/skidl
One thing I've been personally really intrigued by is the possibility of using self-play and adversarial learning as a way to advance beyond our current stage of imitation-only LLMs.
Having a strong rules-based framework to be able to be able to measure quality and correctness of solutions is necessary for any RL training setup to proceed. I think that skidl could be a really nice framework to be part of an RL-trained LLM's curriculum!
I've written down a bunch of thoughts [1] on using games or code-generation in an adversarial training setup, but I could see circuit design being a good training ground as well!
Yes, as well as dealing with a variable-length window.
When generating images with diffusion, one specifies the image ahead-of-time. When generating text with diffusion, it's a bit more open-ended. How long do we want this paragraph to go? Well, that depends on what goes into it -- so how do we adjust for that? Do we use a hierarchical tree-structure approach? Chunk it and do a chain of overlapping segments that are all of fixed-length (could possibly be combined with a transformer model)?
Hard to say what would finally work in the end, but I think this is the sort of thing that YLC is talking about when he encourages students to look beyond LLMs. [1]
There is one posted on HN every week. How many more do we need to accept the fact this tech is not what it is sold at and we are bored waiting for it get good? I am not say "get better", because it keeps getting better, but somehow doesn't get good.
It's not "general intelligence", so it's over hyped, and They get so whiny about the inevitable criticism, and They are ignoring that it's so mindnumbingly boring to have people making the excuse that "designed a circuit board from scratch" wasn't something anyone thinks or claims an LLM should do.
Who told you LLMs can design circuit boards?
Who told you LLMs are [artificial] general intelligence?
I get sick of it constantly being everywhere, but I don't feel the need to intellectualize it in a way that blames the nefarious ???
*waves*
Everyone means a different thing by each letter of AGI, and sometimes also by the combination.
I know my opinion is an unpopular one, but given how much more general-purpose they are than most other AI, I count LLMs as "general" AI; and I'm old enough to remember when AI didn't automatically mean "expert level or better", when it was a surprise that Kasparov was beaten (let alone Lee Sedol).
LLMs are (currently) the ultimate form of "Jack of all trades, master of none".
I'm not surprised that it failed with these tests, even though it clearly knows more about electronics than me. (I once tried to buy a 220 kΩ resistor, didn't have the skill to notice the shop had given me a 220 Ω resistor, the resistor caught fire).
I'd still like to call these things "AGI"… except for the fact that people don't agree on what the word means and keep objecting to my usage of the initials as is, so it would't really communicate anything for me to do so.
We discovered this nearly-magical technology. But now the novelty is wearing off, and the question is no longer "how awesome is this?". It's "what can I do with it for today?".
And frustratingly, the apparent list of uses is shrinking, mostly because many serious applications come with a footnote of "yeah, it can do that, but unreliably and with failure modes that are hard for most users to spot and correct".
So yes, adding "...but without making up dangerous nonsense" is moving the goalposts, but is it wrong?
For example, we are using it to do meeting summaries and it is remarkably good at it. In fact, in comparison to humans we did A/B testing with - usually better.
Another thing is new employee ramp. It is able to answer questions and guide new employees much faster than we’ve ever seen before.
Another thing I’ve started toying with it with, but have gotten incredible results so far is email prioritization. Basically letting me know which emails I should read most urgently.
Again, these were all things where the state of the art was basically useless 3 years ago.
While it might be "moving the goal posts" the issue is that the goal posts were arbitrary to start with. In the context of the metaphor we put them on the field so there could be a game, despite the outcome literally not mattering anywhere else.
This isn't limited to AI: anyone dealing with customers knows that the worst thing you can do is take what the customer says their problem is at face value, replete with the proposed solution. What the customer knows is they have a problem, but it's very unlikely they want the solution they think they do.
So are you happy that a 1940s tic-tac-toe computer "is AI"? And that's going to be your bar for AI forever?
"Moving the goalposts is a metaphor, derived from goal-based sports such as football and hockey, that means to change the rule or criterion of a process or competition while it is still in progress, in such a way that the new goal offers one side an advantage or disadvantage." - and the important part about AI is that it be easy for developers to claim they have created AI, and if we move the goalposts then that's bad because ... it puts them at an unfair disadvantage? What is even wrong with "moving the goalposts" in this situation, claiming something is/isn't AI is not a goal-based sport. The metaphor is nonsensical whining.
I think it would take a human about 2.6 million (waking) years to actually read Common Crawl[0]; though obviously faster if they simply absorb token streams as direct sensory input.
The strength of computers is that transistors are (literally) faster than synapses to the degree to which marathon runners are faster than continental drift; the weakness is they need to, too — current generation AI is only able to be this good due to this advantage allowing it to read far more than any human.
How much this difference matters depends on the use-case: if AI were as good at learning as we are, Tesla's FSD would be level 5 autonomy years ago already, even with just optical input.
[0] April 2024: 386 TiB; assuming 9.83 bits per word and 250 w.p.m: https://www.wolframalpha.com/input?i=386+TiB+%2F+9.83+bits+p...
As the saying goes - "make it work, make it right, make it fast".
Soon everything you see and hear will be built up through a myriad of AI models and pipelines.
It is so bizarre that some people view this as a positive outcome.
Your argument rhymes with:
- "Let's keep using horses. They're good enough."
- "Photography lacks the artistic merit of portrait art."
- "Electronic music isn't music."
- "Vinyl is the only way to listen to music."
- "Digital photography ruins photography."
- "Digital illustration isn't real illustration and tablets are cheating."
- "Video games aren't art."
- "Javascript developers aren't real programmers."
Though I'm paraphrasing, these are all things that have been said.
I bet you my right kidney that people will use AI to produce incredible art that will one day (soon) garner widespread praise and accolade.
It's just a tool.
The population of people who want to create art is higher than the people who have the classical skills. By sheer volume, the former will dominate the latter. And eventually most artists will begin to use AI tools when they realize that's what they are -- tools.
1. (real illustration vs digital illustration)
2. (composing on sheet music vs composing in a DAW)
and
3. illustration vs Stable Diffusion
4. composing vs generative music models such as Suno
What's different is the wide disparity between input and output. Generally, art has traditionally had a closer connection between the "creator" and the "creation". Generative models have married two conventionally highly disparate mediums together, e.g. text to image / text to audio.
If you have zero artistic ability, you'd have about as much success using Photoshop as you would with traditional pencil and paper.
Whereas any doofus can type in the description of something along with words like "3D", "trending on artstation", "hyper-realistic,", and "4K" and then proceed to generate thousands of images in automatic1111 which they can flood DeviantArt with in a single day.
The same applies to music composition whether you are laboriously notating with sheet music or dropping notes using a horizontal tracker in a DAW like Logic. If you're not a musician, the fanciest DAW in the world won't make you one.
I spent 48 hours two weeks back (with only a few hours of sleep) making an AI film. I used motion capture, rotoscoping, and a whole host of other tools to accomplish this.
I know people who have spent months making AI music videos. People who painstakingly mask and pose skeletons. People who design and comp shots between multiple workflows.
These are tools.
Many times this also happens with artists themselves. After a point, you are getting way more commissions than you can produce yourself, so you employ a small army of understudies that learn your techniques and make your pieces for you. So what you describe has existed for hundreds of years.
A short list could include old ones like Rembrandt or Rubens and a new ones like Jeff Koons or Damien Hirst.
Artists and "creative" people have long held a monopoly on this ability and are now finally paying the price now that we've automated them away and made their "valuable" skill a commodity.
I've seen a lot of schadenfreude towards artists recently, as if they're somehow gatekeeping art and stopping the rest of us from practicing it.
I really struggle to understand it; the barrier of entry to art is basically just buying a paper and pencil and making time to practice. For most people the practice time could be spent on many things which would have better economic outcomes.
> monopoly
Doesn't this term imply an absence of competition? There seems to be a lot of competition. Anyone can be an artist, and anyone can attempt to make a living doing art. There is no certification, no educational requirements. I'm sure proximity to wealth is helpful but this is true of approximately every career or hobby.
Tangentially, there seem to be positive social benefits to everyone having different skills and depending on other people to get things done. It makes me feel good when people call me up asking for help with something I'm good at. I'm sure it feels the same for the neighborhood handyman when they fix someone's sink, the artist when they make profile pics for their friends, etc. I could be wrong but I don't think it'll be entirely good for people when they can just have an AI or a robot do everything for them.
The Facebook feed is AI; Google PageRank is AI; anti-spam filters are AI; A/B testing is AI; recommendation systems are AI.
It's been a long time since computers took over from humans with designing transistor layouts in CPUs — I was hearing about the software needing to account for quantum mechanics nearly a decade ago already.
It's frustrating because it's infantalizing, it derails the potential of an interesting technical discussion (ex. Here, diffusion), and it misses the mark substantially.
At the end of the day, it's useful in a thousand ways day to day, and the vast majority of people feel this way. The only people I see vehemently arguing the opposite seem to assume only things with 0 error rate are useful or are upset about money in some form.
But is that really it? I'm all ears. I'm on a 5 hour flight. I'm genuinely unclear on whats going on that leads people to take this absolutist position that they're waiting for ??? to admit ??? about LLMs.
Yes, the prose machine didnt nail circuit design, that doesn't mean whatever They you're imagining needs to give up and accept ???
Irony: humans think in very black-and-white terms, one could even say boolean; conversely LLMs display subtly and nuance.
When I was a kid, repeats of Trek had Spock and Kirk defeating robots with the liar's paradox, yet today it seems like humans are the ones who are broken by it while the machines are just going "I understood that reference!"
When we get to that level, we're all out of work.
In the meantime, LLMs are already basically as good as the scriptwriters made the TNG-VOY era starship computers act.
At the same time I do appreciate the actual performance and potential future promise of this tech. I have to remind myself that the wolf and sheep show is a side attraction, but for some people it’s clearly the main attraction.
The problem with everything today is not only that it’s hype-centric, but that that carries away those who were otherwise reasonable. AI isn’t any special in this regard, it’s just “crypto” of this decade.
I see this trend everywhere, in tech, socio, markets. Everything is way too fake, screamy and blown out of proportion.
I'm sure you see it, I'd just love for someone to pause their internal passion play long enough to explain what they're seeing. Because I refuse to infantalize, I refuse to believe it's just grumbling because its not 100% accurate 100% of the time, and doesn't do 100% of everything.
The problem with hype is that it can become a pathological form of social proof.
fwiw if someone's really into Google minutae: I'm not so sure it is relatively okay anymore, it's kinda freaky how many posts there are on Blind along the lines of "wow I left X for here, assumed i'd at least be okay, but I am deeply unhappy. its much worse than average-white-collar job I left"
in general id recommend Ian Hickson's blog post on leaving. I can't remember the exact quote that hit hard, something like decisions moved from being X to Y to Z to being for peoples own benefit.
I'd also add there was some odd corrupting effects from CS turning into something an aimless Ivy Leaguer would do if they didn't feel like finance.
Why? Is this naive engineering refusing to acknowledge the same old design flaws? Nefarious management fast tracking enshittification? Or do users actually want their write-a-naughty-limerick goofs to get mixed up with their serious effort to fast track circuit design? I wouldn’t want to appear cynical but one of these explanations just makes more sense than the others!
The core tech such as it is is fine, great even. But it’s not hard to see many different ways that it’s already spiraling out of control.
I cannot help but think there are some similarities between large model generative AI and human reasoning abilities.
For example if I ask a physician with a really high IQ some general questions about say anything like fixing shocks on my mini van … he may have some better ideas than me.
However he may be wrong since he specialized in medicine, although he may have provided some good overall info.
Let’s take a lower IQ mechanic who has worked as a mechanic for 15 years. Despite this human having less IQ, less overall knowledge on general topics … he gives a much better answer of fixing my shocks.
So with LLM AI fine tuning looks to be key as it is with human beings. Large data sets that are filtered / summarized with specific fields as the focus.
> The AI generated circuit was three times the cost and size of the design created by that expert engineer at TI. It is also missing many of the necessary connections.
Exactly what I expected.
Edit: to clarify this is even below the expectations of a junior EE who had a heavy weekend on the vodka.
AI happy as it worked the first 10ns of the cycle.
It's a distinction I fear many people will have trouble keeping in-mind, faced with the misleading eloquence of LLM output.
What natural language processing does is just make a much smarter (and dumber, in many ways) parser that can make an attempt to infer the intent, as well as be instructed how to recover from mistakes.
Personally I'm a skeptic since I've seen some hilariously bad hallucinations in generated code (and unlike a human engineer who will say "idk but I think this might work" instead of "yessir this is the solution!"). If you have to double check every output manually it's not that much better than learning yourself. However, at least with programming tasks, LLMs are fantastic at giving wrong answers with the right vocabulary - which makes it possible to check and find a solution through authoritative sources and references instead of blindly analyzing a problem or paying a human a lot of money to tell you the answer to your query.
For example, I don't use LLMs to give me answers. I use them to help explore a design space, particularly by giving me the vocabulary to ask better questions. And that's the real value of a conversational model today.
- https://www.damninteresting.com/on-the-origin-of-circuits/
- https://www.sciencedirect.com/science/article/abs/pii/S03784...
Agree with OP that the raw models aren't that useful for schematic/pcb design.
It's why we build flux from the ground up to provide the models with the right context. The models are great moderators but poor sources of great knowledge.
Here are some great use cases:
https://www.youtube.com/watch?v=XdH075ClrYk
https://www.youtube.com/watch?v=J0CHG_fPxzw&t=276s
https://www.youtube.com/watch?v=iGJOzVf0o7o&t=2s
and here a great example of levering AI to go from idea to full design https://x.com/BuildWithFlux/status/1804219703264706578
It kind of grosses me out that we are entering a world where programming could be just testing (to me) random permutations of programs for correctness.
Most people are wrong that AI won't be able to do this soon. The same way you can't expect an AI to generate a website in assembly, but you CAN expect it to generate a website with React/tailwind, you can't expect an AI to generate circuits without having strong functional blocks to work with.
Great work from the author studying existing solutions/models- I'll post some of my findings soon as well! The more you play with it, the more inevitable it feels!
Are you able to accomplish this with prompt-engineering, or are you doing fine-tuning of LLMs / custom-trained models?
A lot of the netlists are electrically nonsense when it's doing synthesis for me. Have you found otherwise?
My impression is that synthetic datasets and finetuning will basically completely solve the problem, but eventually it’ll be available in general purpose models- so it’s not clear if its worth it to build a dedicated model.
Overall the article’s analysis is great. I’m very optimistic that this will be solved in the next 2 years.
Can you? Because last time I tried (probably about February) it still wasn’t a thing
I don’t think pedantry helps here, it doesn’t add to the conversation at all.
I agree that using llms for generating things like schemas, components, build scripts etc is a good use of the technology, but we’re no closer to saying “make a saas landing page for X using vercel” and having it ready to deploy, then we were a year ago
The industry does not like sharing, and the openly available datasets are full of mistakes. As a junior EE you learn quite quickly to never trust third-party symbols and footprints - if you can find them at all. Even when they come directly from the manufacturer there's a decent chance they don't 100% agree with the datasheet PDF. And good luck if that datasheet is locked behind a NDA!
If we can't even get basic stuff like that done properly, I don't think we can reasonably expect manufacturers to provide ready-to-use "building blocks" any time soon. It would require the manufacturers to invest a lot of engineer-hours into manually writing those, for essentially zero gain to them. After all, the information is already available to customers via the datasheet...
Datasheets get incredibly confusing incredibly fast, and every single detail is critical. It's quite common for one datasheet to describe multiple parts at the same time, even in the same tables and diagrams, and have contradictions between the two parts. You end up with pins labeled "EN/SET" where the xxx1 variant has the pin act as Enable and the xxx3 variant have the same pin act as Setpoint. If you don't generate two separate symbols for those, the symbols are essentially useless because they can't be trusted. And that's just about the easiest thing you're going to come across.
This isn't a problem which can be solved downstream. Even trained experts are often confused because the input data is just really bad. You can't throw garbage into AI and except diamonds to come out, the only way to solve it is to convince all the manufacturers to switch to a to-be-developed universal documentation protocol.
I don't know how feasible it is. This would probably take low $millions or so of training, data collection and research to get not trash results.
I'd certainly love it for trying to diagnose circuits.
It's probably not really that possible even at higher end consumer grade 1200dpi.
And the devices, in this case, bluetooth aux transceivers, they all do the same things. They've even more or less converged on all being 3 buttons. When optimizing for cost reduction with the commodity chips that everyone is using to do the same things, the manufacturer variation isn't that vast.
In the same way you can get 3d models from 2d photos because you can identify the object based on a database of samples and then guess the 3d contours, the hypothesis to test is whether with enough scans and schematics, a sufficiently large statistical model will be good enough to make decent guesses.
If you've got say 40 devices with 80% of the same chips doing the same things for the same purpose, a 41st device might have lots of guessable things that you can't necessarily capture on a cheap flatbed
This will probably work but it's a couple million away from becoming a reality. There's shortcuts that might make this a couple $100,000s project (essentially data contracts with bespoke chip printers) but I'd have to make those connections. And even then, it's just a hobbyist product. The chances of recouping that investment is probably zero although the tech would certainly be cool and useful. Just not "I'll pay you money" level useful.
They are already far ahead of many others with respect to next generation EE CAD.
Judicious application of AI would be a big win for them.
Edit: adding "TL;DRN'T" to my vocabulary XD
Adding Skynetn't to company charter...
"If we make a really really good specialty text-prediction engine, it could be able to productively mimic an imaginary general AI, and if it can do that then it can productively mimic other specialty AIs, because it's all just intelligence, right?"
few really understand what the limits of the tech are. and if it will even unlock the usecases for which it is being touted.
TLDR: We test LLMs to figure out how helpful they are for designing a circuit board. We focus on utility of frontier models (GPT4o, Claude 3 Opus, Gemini 1.5) across a set of design tasks, to find where they are and are not useful. They look pretty good for building skills, writing code, and getting useful data out of datasheets.
TLDRN'T: We do not explore any proprietary copilots, or how to apply a things like a diffusion model to the place and route problem.
* Failed to properly understand and respond to the requirements for component selection, which were already pretty generic.
* Succeeded in parsing the pinout for an IC but produced an incomplete footprint with incorrect dimensions.
* Added extra components to a parsed reference schematic.
* Produced very basic errors in a description of filter topologies and chose the wrong one given the requirements.
* Generated utterly broken schematics for several simple circuits, with missing connections and aggressively-incorrect placement of decoupling capacitors.
Any one of these failures, individually, would break the entire design. The article's conclusion for this section buries the lede slightly:
> The AI generated circuit was three times the cost and size of the design created by that expert engineer at TI. It is also missing many of the necessary connections.
Cost and size are irrelevant if the design doesn't work. LLMs aren't a third as good as a human at this task, they just fail.
The LLMs do much better converting high-level requirements into (very) high-level source code. This make sense (it's fundamentally a language task), but also isn't very useful. Turning "I need an inverting amplifier with a gain of 20" into "amp = inverting_amplifier('amp1', gain=-20.0)" is pretty trivial.
The fact that LLMs apparently perform better if you literally offer them a cookie is, uh... something.
The number of times I've had to entirely redo a circuit because of one misplaced connection, yeah, none of those circuits worked for any price before I fixed every single error.
I think Gemini could definitely do that microphone study. Good test case! I remember spending 8 hours on DigiKey in the bad old times, looking for an audio jack that was 0.5mm shorter.
That's absurd to me, it took so long to figure out which random sequence of letters was the smallest in overall PCB footprint.
Maybe we found it, we think it's the AYU2T-1B-GA-GA-ETY(HF) but sure would be nice if Digikey had a search by footprint dimensions.
Yet strangely the physical ability of a device to fit into a location you need it is not in the list of things I can search. Takes ten seconds to find the numbers -- after I download and open the PDF file.
https://www.digikey.com/en/products/filter/coaxial-connector...
Just so strange, but so common. And digikey is heads and shoulders above average, McMaster might be the only better one I know of at it and they're very curated.
But the bottom line is that it's a task that a novice could have solved with a Google search or two, and the LLM fumbled it in ways that'd be difficult for a non-expert to spot and rectify. LLMs are generally pretty good at information retrieval, so it's quite disappointing.
The cookie thing... well, they learn statistical patterns. People on the internet often try harder if there is a quid-pro-quo, so the LLMs copy that, and it slips past RLHF because "performs as well with or without a cookie" is probably not one of the things they optimize for.