4mo ago

Is deepseek weak on circuit design?

I've tried using deepseek (first time I've ever used an LLM, so maybe I'm being dumb) to help me a little with designing some circuit because my reference book was leaving out a LOT of crucial information.

The results have been ... subpar. The model seems to be making quite elementary mistakes, like leaving floating components with missing connections.

I'm honestly kinda disappointed. Maybe this is a weak area for it. I've probably had to tell deepseek more about designing the circuit in question than it has told me.

Edit: I realised I was just being dumb, since LLMs aren't designed for this task.

16 comments

I wouldn’t have thought an LLM to be of use for circuit design in the first place, so I wouldn’t have been disappointed.
- I mean, people use it for making code.
  
  Maybe I'm wrong but the amount of code available online you can use to feed a model is probably a few orders of magnitude larger than circuit designs.
  
  LLMs are often trained on up to some 80% code, depending on use although usually it's probably lower, as that has been shown to improve their logical/thinking skills.
  Basically, if the task can be done with only words and there is a lot of data of it present day LLMs can probably get really good at it if properly trained for it, but for things like circuits, where a lot of the data is likelly to be graphical or there might just not be much of it, LLMs aren't yet as good at.
I am in the circuit design field. What is the reference book ?
- Foundations of oscillator design. I wanted to find a way to improve the bandwidth of some of the FET oscillators in the book but I've found different resources.
  
  And DeepSeek generated a Spice model of a standard oscillator design with missing connections? That's a worry.
Though I'm not in electronics, I am in software development, and I have tried using various language models for basic circuit development for very small side hobbies, and I have found it to be mixed at best.
It works better when it's integrated into a search engine so that it has real humans explaining things so that they can just do the summary and not build something from scratch.
General rule that's helpful for keeping in mind with generative AI models is they can only be as knowledgeable as the subject matter they have been trained on. And even then, it's only "can" of potential, not a guarantee, as training on the material doesn't necessarily mean it will answer correctly with regards to that material.
Which makes intuitive sense if you compare to a human, but is easy to miss in all the black box hype surrounding AI. No matter how clever a human being is, if they don't know something, they don't know it and thinking about it can only do so much. Now imagine that, but also missing key capabilities that humans have, like the ability to ask questions and learn long-term information from them in real-time.
Side note: The one subject I can think of where thinking analysis alone may work functionally to uncover new knowledge is, like, mathematical proofs where it's abstract A, B, therefore C logic, and that's also something LLMs don't have the design or capability for.
Deepseek has some legit reasons to have hype, but primarily it's hype relative to other LLMs and their training. There are still a lot of hurdles in getting LLMs past common problems.
AIs for EDA are more about optimizing lengths of traces and looking for bugs, as far as I know from outside the industry. LLMs are generally not used for it and the AIs that are used are as far as I know proprietary. I also cannot confirm there's any current effort to train LLMs to help with EDA, though it's possible that someone else is doing that and I just don't know about it.
You might want to ask it to break it's work down into steps, then ask it to double check it's work for each step
Not perfect, but you'd be surprised how much asking the questions in the right way will improve performance
You might have better luck asking this in a more electronics-centric channel
- Does lemmy even have such a community? I've never seen anything like it.
  But my post isn't really about electronics (or maybe it can be). I'm not looking for advice on building circuits.
  I'm just making an observation about DeepSeek's capabilities. Maybe an AI focused channel would be better?
  
  I suggested electronics communities because electrical circuit design is a rather niche topic, and I wouldn't think that many people have tried using LLMs for circuit design. But maybe some people in electronics and electrical engineering communities may have tried it and will have some advice.
  What sort of output are you asking for? Is the AI giving you an image, or some sort of SPICE config, or something else?
  I'm not sure which Lemmy client you're using, but I'm using Boost and doing a quick search I see the following communities :
  https://discuss.tchncs.de/c/electronics
  https://discuss.tchncs.de/c/askelectronics
  https://lemmy.world/c/ece
Sometimes LLMs get basic things wrong. This is true.

16 comments