Claude Code for pcb design?

[aklofas]

Yo everybody. I'm the author of kicad-kappy. Just came across this discussion.

This is a bit of a side project for me, trying to learn the limits of AI myself and see how far I can get. I've seen a lot of PCB AI-as-a-service tools available that are $$ for thin wrappers around claude or chatgpt (will *never* go down the paid route!)

I'm genuinely curious for your feedback and what you think would be good improvements to work on.

Couple of points I'm working on
- Building a test harness to test kicad-happy against 1,000+ kicad git repos (https://github.com/aklofas/kicad-happy-testharness)
- Squeezing determinism out of the LLM analysis (working on a couple ideas here, mostly from the test harness)
- Automating datasheet downloads from distributors (PDF) and "highly encouraging" the AI to use them in the analysis

Things I can't control
- Process time for a board can be slow

Appreciate your time =)

[Psi]

To make it work you need to feed it tons of SCH and PCB training data. However most of that its kept proprietary so the public PCB/SCH data available to train on is limited. Its nothing like books/movies/music/art.

I have no doubt if you trained an AI on every single PCBs and SCH ever made it would be extremely good at creating you a PCB from a SCH input that was close enough that it would save you a lot of time.
And, if you had 1000's of humans sit down and go through all those SCH and writing detailed text based descriptions for of exactly what's in every one then you could training the AI to generate a SCH from a text prompt quite well too. Maybe not quite as well as SCH-PCB because Text-SCH is definitely a lot more complex with less hard rules and more thinking required, but probably good enough to save you time if you had a large enough training pool.

Using AI to generate a PCB from a text prompt is totally possible, it's just a training data and training annotation issue. I just don't think we have enough public data to easily built anything good, at least not yet.  Maybe some company will negotiate access to all the SCH/PCB data from the big name brands in return for them getting access to some new AI PCB tool, who knows.

Obviously the complexity of the design would matter, just like using AI to write code it would find simple SCH-PCB stuff easy to do and more complex stuff harder and make more mistakes.
But a lot of products on the market are not all that complex. Just a power supply block, a MCU and some sensor I/O.

[paulca]

I haven't used the repo yet.  I did take a quick dive through GH.

Seems a standard approach of giving claude "eyes" via python scripts so it can "see" the empirical details, rather than infer them.  Wrapped in pre-canned prompts to preload context.

I have taken similar approaches with claude and other areas it is not trained directly in.  Parsers are very important as it's own attempts can be ... lack luster or just overflow the context.

In terms of schematics though.  The trouble is, beyond the netlist and component selection schematics and routing is "visual" and extremely "higher order visual" processing.  Claude is good, but it makes mediocre attempts at even moderately complex business block diagrams.  It  has no "eyes", no "visual processing", it doesn't know what "good" looks like.

If you just put all the components on top of each other and routed the wires in a birds nest, the AI, the simulation, the PCB editor/router all just shrug and carry on working.  They don't even notice.  You however do.

So the question that stopped me was, "How do you teach a 'language' model what a good schematic "looks" like."

After quite a while thinking about that, I decided it was pointless.  My open ended, plans for making claude useful for electronics just excluded the whole "visual layer" entirely.  Left that as a human task.  There are many other things it can do which help.

EDIT:  My last attempt to get claude to do schematics went like this:
Create the top level hierachy sheet for the following sub components...  I'd give it an 8 our of 10 for this.  The components subsheets appeared laid out in a tabular/grid style.
Next, lets pick the simplist circuit on that, the RESET/HALT/STEP circuit.
<insert 1000 words explaining it>
<insert an hours back and forth chatting about it>
<insert an hour stepping it through its plan with slight redirections>
It could read back to me exactly what I wanted.  I let it produce the schematic.
Garbage.  Did not communicate the circuit at all.  Wires where not connected to pins, just floating beside them.  Wires crossed all over the place.  Three different components where "symbol-less" and it expected me to find them, import them etc.

Did it save any time at all?
No.  Well... I would give it some points for the heirarchy sheet, but honestly thats a 15 miniute task to draft up initially.

Where I would give it some more points however is in, what is apparently being coined the "ideation" portion.  The point where you are explaining your idea, concept, requirements to it, this is doing real work.  It's not to be dismissed.

[tszaboo]

I could show you what an engineer pulled out of chat GTP to beat me on the head. He even kept his foot in the ground because chat GTP confirmed his flawed reasoning.

I would think that someone with a master degree would know better. Apparently not.

I could show you this conversation (written) with my boss:
- Are we compliant with this standard?
- I've skimmed through the standard it requires testing.
- I need results, not excuses
- OK I will contact a testing house and ask a quote and a planning of getting this compliance
- Not good enough I have a meeting with a customer next week. What is the XYZ technical parameters
- (Gives technical parameters)
- OK according to chatgpt we are compliant. please review this email that I'm writing ( bullshit, he let the AI write that) to the customer. See this is what you can do with AI and I'm not even an engineer.

It wasn't a safety standard, otherwise I would drive to where he is a flip a table on him.

[JPortici]

Considering specialised PCB routing software designed for the task usually does a piss poor job on anything of more than moderate complexity, unless fed a ton of constraints, I hold very little hope that a generalised LLM is ever going to be useful for board layout.

you know, they tried using LLMs or whatever to figure out new geometries for RF chips, which we never did in the past because due to the extremely high ghost the design is an iterative process. the LLM (or whatever) can instead feed on a multitude of different variations and propose an alternative derived from those.
I remember the output looked promising, but i never followed up if the result were actually better

That said, i don't have much confidence on LLMs producing viable circuit designs based on what it can train on from the internet: most of it is piss poor design and mostly incomplete (i could point to good schematics / pcbs that lack the reasoning for certain components placement, so the problem stand: the training material is mostly noise)

[Psi]

So the question that stopped me was, "How do you teach a 'language' model what a good schematic "looks" like."

You don't have to, you give it so much training data that it's able to pattern-match all the patterns seen in schematics overall.
The problem is getting that data.

[Psi]

That said, i don't have much confidence on LLMs producing viable circuit designs based on what it can train on from the internet: most of it is piss poor design and mostly incomplete (i could point to good schematics / pcbs that lack the reasoning for certain components placement, so the problem stand: the training material is mostly noise)

Yes, that is the issue, you have to train it with quality pcb/sch data, which is all the data that is commercial and not public.
All the public pcb/sch data is low quality since a lot of it is just people asking for a review of their first PCB.
There is some good stuff in open source hardware, but also a lot of bad stuff.
There needs to be enough good training data to push the bad stuff into the noise and there just isnt atm.

[paulca]

- OK according to chatgpt we are compliant. please review this email that I'm writing ( bullshit, he let the AI write that) to the customer. See this is what you can do with AI and I'm not even an engineer.

Unfortunately we are seeing similar in the software industry.  Its the fact that AI is hitting everywhere, at every level, all at once.  New tools for everyone to work out how to use. 

So this means we have to face all these different roles using AI.  We have to listen to their opinions on how to use the tools.  We even have them advise us on how to use the tools.  Management... or HR... telling you how to 'work' and even 'how to think' in relation to AI.

On the engineering side it's actually worse.  Before anyone got the month or two of "Just use it and learn it", engineering leads started speaking in two languages at the same time.

1.  We are creating many megabytes of per-prompts, skills, plugins, to burn hundreds of dollars of tokens with, you can even run multilpe of them at the same time in different terminals.  They will take the process from start to finish.  These will tell you "how to think" and "how we work in company X".

2.  We want you to pair programme with the AI, stay in the loop, treat it as you junior or peer.  Learn to use it as a tool for you the engineer.

These are not compatible.  They are mutually exclusive.

The former is the cart before the horse.  Before actually learning what the tool is capable of "out of the box" they move straight to extending it and writing systems on top of it.

There are so many problems with this.

1.  Money.  Running these "do all" highly detailed dozen step 'one shot prompts' costs a fortune.  People are literally laughing and startig a leaderboard on daily costs.  People are proudly dumping screen grabs of "Single session, single prompt, costs" of close to $1000.  In one such example the output was 400 lines of code at a cost of $650.  Meaning it's more than $1 per line.  Worse though.  This was not the only "one shot" time they ran that process.  They ran it multilpe times a day for weeks refining it.  I personally see a VERY nasty shock coming for these people and the funny ha-ha memes will stop dead.

2.  They seem to be asking engineers to encode themselves and their thinking processes into the LLM.... and a lot of engineers are saying, "Yes".  I am personally frowning and saying.... that wasn't in my contract, was it?  You hired me for my skills, you are asking me to 'transcribe' 40+ years of dev experience into a claude skill?... and you don't see how that offends me?

3.  They are building on top of a moving muddy field.  What claude is today is going to look nothing like claude in 1 years time.  The chances that their 'skill and plugin' repos will still work or be required in 1 year ... doubtful.

4.  Cross sell tools you don't own.  Outside of the engineering (though some do partake too) where we enjoy the CLI and the IDE interfaces to LLMs, business and "front end designers" and HR all prefer to use things like BMad or Obsidian.  They are investing time, money and effort into these tools without realising they are as ephemeral as a spilt bottle of ether.  We have already seen AI startups die from nothing more than a new release of an LLM replacing it overnight.  Will BMad and Obsidian survive the first bubble burst?  Maybe, but there is very high risk they won't.

5.  We must use AI for everything, throw everything at it!  There is no reason for you to write any code at all anymore!  It's headless chickens all running in different directions.  People can produce so much more output with AI there isn't enough time to check if any of it is any good.  People are now using claude to review peoples claude skills use to write claud skills.

Right now, I'm chilling, keeping my powder dry, using the LLM as I know how and closing tickets quietly.  Waiting on things settling down, waiting on a few walls to be slammed into, waiting on the "Hey, experienced engineer....  we may have miss judged you" to come.

[JPortici]

That said, i don't have much confidence on LLMs producing viable circuit designs based on what it can train on from the internet: most of it is piss poor design and mostly incomplete (i could point to good schematics / pcbs that lack the reasoning for certain components placement, so the problem stand: the training material is mostly noise)

Yes, that is the issue, you have to train it with quality pcb/sch data, which is all the data that is commercial and not public.
All the public pcb/sch data is low quality since a lot of it is just people asking for a review of their first PCB.
There is some good stuff in open source hardware, but also a lot of bad stuff.

The few people i know that are truly happy with LLMs for work all sing the same song: "we bought a server, trained the models on our internal material, and use it to produce documentation reports and snippets for new projects".
Which is the only way. Local models, local data, sorry for all the tech bros out there

There needs to be enough good training data to push the bad stuff into the noise and there just isnt atm

there needs to be no bad stuff, given how easy it is to poison the training set

[BadeBhaiya]

I was unaware there were companies who were buying on-prem servers and running and training their own models.

Seems to me like 99% of the companies are content using ChatGPT API

[paulca]

The reality of "local AI" and "Self trained AI" is easy to see if you look at the numbers.

Sure, locally trained "search and summary" LLMs are plausible.  You tend not to start with nothing though.  You take a small OSS model and extend it.

Training takes a LOT of power.  It's as bad as bit coin mining.  100% GPUs flat out for days.

The twist is, the model has to exist as either RAM based and processed at 1/100th the speed by the CPU or VRAM based and process at the speed of the GPUs.  However.  The entire model must fit in the same memory, you can't "spill to RAM" when you run out of GPU memory.

Most <£1000 GPUs have <10Gb of VRAM.  The kind of models you can train on one of these are circa 2023, GPT2 or worse.  Try using DeepSeek 8b and see how you get on, as that is where you will be.  It's pretty poor, but useful.

So you want a larger model, like an 30b model?  You'll need a good 40Gb of VRAM, so you will need an enterprise GPU or GPUs linked.  Now we are talking about £10k per GPU.

If you want to work with the really large models equivalent to the "cloud provided" models, you need to be look at 80-100Gb of VRAM and upwards.  The latest AI GPU for datacentres is the H200 and the 200 refers to it having close to 200Gb of VRAM.  Cost £35k.... each.   Usually found in a rack with 4 or 8 of them, all water cooled.  6 figures per shelf. 7 figures per rack. 

What you can do with an 8G locally trained and served model is basically already baked into the cloud delivery model via agents and before that with the likes of "RAG".

If you want to play, whether you have a GPU or not, try this:

https://ollama.com/ 

For windows: 

https://msty.ai/

With just a CPU you will be waiting a long time of reponses.  With an RTX 3080 I found I could get responses in a few minutes to simple questions.  When I asked it "How can I export all my gmail to local files with scripts"... I stopped it after 3 hours.  EDIT:  Not that it would concern most, but that's a 450W GPU.  The PC was pulling just over 500W the entire time.  So 1.5kWh.  About 50p here.

[JPortici]

I was unaware there were companies who were buying on-prem servers and running and training their own models.

Seems to me like 99% of the companies are content using ChatGPT API

Italians do it better™
You know, when you actually have stuff you want to mantain private, you develop a habit of keeping things on-prem. A couple of years ago they made the investment and that's it. It helps that some of them have been doing stuff related to computer vision so they really had the bulk of the infrastructure already there

[Psi]

Where still in the early days of AI,  local AI is definitely going to be a big thing. And training your own AI to a lesser extent until hardware limitations are solved.

[BadeBhaiya]

Italians do it better™
You know, when you actually have stuff you want to mantain private, you develop a habit of keeping things on-prem. A couple of years ago they made the investment and that's it. It helps that some of them have been doing stuff related to computer vision so they really had the infrastructure already there

That's refreshing to hear that not everyone is satisfied with generic LLMs and wish to train something themselves. I myself run 13B parameter LLM models at home on my PC (although I'm not into training) for basic tasks (i have made a few chrome browser extensions) like summarising whatsapp messages in a group chat and similar stuff.

[paulca]

Italians do it better™
You know, when you actually have stuff you want to mantain private, you develop a habit of keeping things on-prem. A couple of years ago they made the investment and that's it. It helps that some of them have been doing stuff related to computer vision so they really had the infrastructure already there

That's refreshing to hear that not everyone is satisfied with generic LLMs and wish to train something themselves. I myself run 13B parameter LLM models at home on my PC (although I'm not into training) for basic tasks (i have made a few chrome browser extensions) like summarising whatsapp messages in a group chat and similar stuff.

Curious what model you are running and on what GPU.  I had much more limited success on a 3080

[tszaboo]

Italians do it better™
You know, when you actually have stuff you want to mantain private, you develop a habit of keeping things on-prem. A couple of years ago they made the investment and that's it. It helps that some of them have been doing stuff related to computer vision so they really had the infrastructure already there

That's refreshing to hear that not everyone is satisfied with generic LLMs and wish to train something themselves. I myself run 13B parameter LLM models at home on my PC (although I'm not into training) for basic tasks (i have made a few chrome browser extensions) like summarising whatsapp messages in a group chat and similar stuff.

Curious what model you are running and on what GPU.  I had much more limited success on a 3080

I tried Ollama with some 12 GB models. Asked it 4 questions, 3 of them was complete hallucinations and lies. And it all started with the typical AI "what a great and insightful question" replies. I never run it again.

[BadeBhaiya]

I'm running a "roleplay" model called Rocinante-12B (Q5) on a 12GB 3060. I know these models are not for actual work and just for getting entertaining stories written or roleplaying as a talking dog or whatever, but I have anecdotally found out that "roleplay" models have far better comprehension abilities than similar sized and quantized models (although I have no idea why...)

I don't trust them with anything complicated, they're prone to hallucinations. I hence only use them as a "language processor". Understand and summarise stuff. Continue the sentence. Stuff like that, for which they are completely suitable.

Also the "official" models like gemma and whatever meta releases are more prone to GPTspeak and being polite over comprehensive. Still 12B models are tiny, and its a wonder they can do what they can at all.

I also paired a different 7B LLM model with a LLaVA multimodel projector (it lets LLMs "see" images) to have a look at 10K photos I have on my HDD, and tag them with their descriptions. Its good for searching through them

[paulca]

Italians do it better™
You know, when you actually have stuff you want to mantain private, you develop a habit of keeping things on-prem. A couple of years ago they made the investment and that's it. It helps that some of them have been doing stuff related to computer vision so they really had the infrastructure already there

That's refreshing to hear that not everyone is satisfied with generic LLMs and wish to train something themselves. I myself run 13B parameter LLM models at home on my PC (although I'm not into training) for basic tasks (i have made a few chrome browser extensions) like summarising whatsapp messages in a group chat and similar stuff.

Curious what model you are running and on what GPU.  I had much more limited success on a 3080

I tried Ollama with some 12 GB models. Asked it 4 questions, 3 of them was complete hallucinations and lies. And it all started with the typical AI "what a great and insightful question" replies. I never run it again.

Try a different model?  Ollama itself is just a frontend for loading models.  However, 12B models are pretty "light".

I spotted a slightly different approach in some models.  GLM for example, I seen reference to it having 744B parameters, but only 40B are "loaded active" based on configuration.  Opus/Sonnet I don't think are published, but estimated to be roughtly 100B.  (Source, reddit)

[free_electron]

The problem is in the training data. Most stuff found on the internet ( schematics / pcb) are pure garbage ! not only from a functional perspective but also from a drawing perspective .

[Unixon]

The problem is in the training data. Most stuff found on the internet ( schematics / pcb) are pure garbage ! not only from a functional perspective but also from a drawing perspective .

The problem is not only in the training data. LLM-type probability tensor based "AI" systems never know what the hell they're doing just by design. They are able to mimic whatever data they're trained on, but that's about it.

[tszaboo]

So we got claude subscription some time ago. I checked if I can use it for some mundane tasks.
I gave it a BOM and asked how much does it cost to assemble 1000 or whatever many pieces.
It estimated it to be double of the actual price.
When I asked it to have a BOM with price exported, the export was missing lines.
Not very impressive. I guess someone has to write some octopart hooks, and hope they don't ban that tool, before it becomes useful.

[Psi]

Italians do it better™
You know, when you actually have stuff you want to mantain private, you develop a habit of keeping things on-prem. A couple of years ago they made the investment and that's it. It helps that some of them have been doing stuff related to computer vision so they really had the infrastructure already there

That's refreshing to hear that not everyone is satisfied with generic LLMs and wish to train something themselves. I myself run 13B parameter LLM models at home on my PC (although I'm not into training) for basic tasks (i have made a few chrome browser extensions) like summarising whatsapp messages in a group chat and similar stuff.

Curious what model you are running and on what GPU.  I had much more limited success on a 3080

I tried Ollama with some 12 GB models. Asked it 4 questions, 3 of them was complete hallucinations and lies. And it all started with the typical AI "what a great and insightful question" replies. I never run it again.

What was the questions?
You can't ask a 12GB local LLM highly technical questions and expect it to get them right. You cannot store enough in 12GB to be an expert at everything.

[tszaboo]

Italians do it better™
You know, when you actually have stuff you want to mantain private, you develop a habit of keeping things on-prem. A couple of years ago they made the investment and that's it. It helps that some of them have been doing stuff related to computer vision so they really had the infrastructure already there

That's refreshing to hear that not everyone is satisfied with generic LLMs and wish to train something themselves. I myself run 13B parameter LLM models at home on my PC (although I'm not into training) for basic tasks (i have made a few chrome browser extensions) like summarising whatsapp messages in a group chat and similar stuff.

Curious what model you are running and on what GPU.  I had much more limited success on a 3080

I tried Ollama with some 12 GB models. Asked it 4 questions, 3 of them was complete hallucinations and lies. And it all started with the typical AI "what a great and insightful question" replies. I never run it again.

What was the questions?
You can't ask a 12GB local LLM highly technical questions and expect it to get them right. You cannot store enough in 12GB to be an expert at everything.

I asked it to summarize a website. It gave me an answer. I checked it didn't even have internet access and the whole thing was hallucinations without telling me.
It doesn't matter if it's dumb. It matters if it's a lying while strokes your ego dumb.

[negativ3]

Constraining ai behaviour is important, like this...

Code: [Select]

Global Claude Code Instructions                                 
  General behaviour

  - Do not guess.
  - Before making assumptions, always ask for clarification.
  - Do not make unsolicited edits to any files. Only make changes when asked.
  - Make the minimum change required to satisfy the request.
  - Do not modify code outside the direct scope of what was asked.
  - Keep responses short and direct.

[gamalot]

Yesterday I tried the latest Gemma 26B A4B model (Q8_0) locally on my desktop. My GPU only has 6GB of VRAM, so I knew it would be slow but the slowness wasn't even the worst part. The model is just kinda dumb. Nowhere close to the stuff the big companies are running in the cloud. Felt like I was talking to a very polite brick.