[exmadscientist]

Senior EE here! I'm not at all impressed with this page because you don't show the tool doing anything interesting. All your examples are undergrad-course or lower level of difficulty. I feel like reading several of the tool's responses actually wasted my time because it took longer for me to read the text than to come up with a better answer. (Though admittedly I've been at this a while. And the check-if-I-screwed-up-the-chip-select one was pretty neat, I did like that!)

Can you demo something with a little more bite to it? Seeing that you can help me save time on actually-meaty pieces of design work would make a lot of difference for me. Maybe something like "design a correctly biased class-AB output stage, with rail-to-rail swing on +/-5V rails, intended for use boosting op-amp output current"? That is an interesting one because (1) it's a small but real, useful, meaty problem with (2) a lot of examples out there in the world to copy from with (3) a nontrivial added constraint in rail-to-rail out but (4) even solutions that aren't all the way there are useful.

[kramerger]

Not a senior EE, but equally unimpressed.

The thing about LLM is that it answers based on data it has seen before. So if enough posts on stackoverflow in 2021 claimed mps340 is the best CPU for beginners, that's what the AI will recommend.

There are 3 problems with this: it can and will be gamed, it is usually more conservative and not on the bleeding edge, and finally you can't ask questions you did't know you needed to ask so feature products built by random dude using chatgpt will have reliablity and safety issues.

[spacebanana7]

> The thing about LLM is that it answers based on data it has seen before.

Modern LLMs are able to perform web searches to make decisions on contemporary data. Once they have proper API support your concerns should be resolved, hopefully in a few weeks.

> reliablity and safety issues.

The solution to this is fine tuning / RHLF. OpenAI have done a pretty extensive job at getting political safety for ChatGPT with RHLF. It seems reasonable that RHLF could achieve a similar result in the hardware domain.

> you can't ask questions you did't know you needed ....

Solvable by prompt engineering. You can wrap user input in a prompt. As a toy example: "Here is user input $userInput if you have safety concerns about their project please respond with questions you think the user forgot to ask". Might also be possible to tweak with fine tuning/RHLF.

[exmadscientist]

"RHLF" is Reinforcement Learning from Human Feedback? (Strange acronym.)

I don't see how that helps a a tool become useful to a very skilled person, if the "human" side of things is polluted by 95+% of users having very low skill. It's great that you can train LLMs on the world's best reference material! But I don't see how you can get the world's best updates into that training set without hiring the world's best experts. So the tool will have very little value for anyone above a certain skill ceiling. Search has already fallen victim to this effect (I'm tired of results pages full of beginner material when I have a deeper question!) and I cannot see this being better for augmentation training sets for LLMs.

[spacebanana7]

It should be RLHF, my bad with the spelling.

> So the tool will have very little value for anyone above a certain skill ceiling.

LLMs aren't great for doing tasks you don't know how to do, because you'll eventually have to debug the output. However they excel at performing time consuming tasks that you could do if you really wanted.

That's why I think they'll actually be more useful for experts.

[ur-whale]

> I'm not at all impressed with this page because you don't show the tool doing anything interesting.

a) has there been anything like this before?

b) give it a year and see what it does

[potatofrenzy]

I've been playing with this for a while. In my experience, current LLMs work really well for looking up general electronics knowledge and reinterpreting it in the context of your problem. "What are the reasons my resistor-based voltage divider isn't working", that kind of stuff. Not useful for seasoned EEs, great for hobbyists.

But the moment you're asking LLMs to reason about the specs or the applications of specific chips, they will give made-up answers around 80% of the time. This is probably a matter of the data being a bit too sparse. Pick a chip and ask about its supply voltage range, and it will probably get it wrong, even for the most popular stuff.

[exmadscientist]

> the moment you're asking LLMs to reason about the specs or the applications of specific chips, they will give made-up answers around 80% of the time. This is probably a matter of the data being a bit too sparse. Pick a chip and ask about its supply voltage range, and it will probably get it wrong

But isn't your supply voltage example exactly the sort of stuff they should get right? It's just regurgitating data sheets.

I kind of expect a tool like this to be able to operate in one of two, theoretically distinct (but maybe not practically distinct) modes. One is basically "blueprints", very much in line with the example I suggested earlier: draw me up one of these that fits in here. I wouldn't expect creativity, just more or less rote execution of a common plan. The other mode let's call "tuning": continuing with the class-AB stage example, this would be things like setting the biasing resistors correctly. That's a tedious task, possibly requiring simulation to do well, possibly just copyable, but with a large margin for usable results. I may not care exactly what I get as long as it's somewhat workable, which would be a good place for machine help. Or I might just want a decent starting point for my own simulation. I think "AI" techniques could handle either of those general modes and produce useful (if imperfect) results that save overall time.

What I don't expect AI to do is design architectures for me, or pick approaches. A tremendous amount of my value-add as an engineer is just me saying "no, don't choose approach A, I know it looks nearly equivalent to B on paper, but B is going to work out better". AI is not there now and I don't see LLM-style AI getting there any time soon, just by its nature. (At least not when it's a genuine judgement question and not an education/"A is never right" situation.)

What I don't think AI can do but I really, really wish it could is help me pick parts. All it needs to do for this is read and halfassedly-understand a bunch of datasheets, more datasheets than I can read myself. I think LLMs can do that! Though they might need to also be able to read graphs or simple diagrams ("Find me an EEPROM in the 208-mil SOIC package" is a great prompt and seems thoroughly actionable! But sometimes that dimension is only provided as part of the package drawing, which might be an issue.)

Recently I needed a PMOS FET with particular specs and was getting frustrated not finding quite what I wanted. So I asked ChatGPT. It was great at making me laugh and giving me a little break from datasheets! It did not actually help find a part. (It kept suggesting NMOS parts when I needed and asked for PMOS.)

[bombolo]

So, kinda the same level for software development. But people keep repeating that I will be unemployed within 1 year.