I am particularly curious how multimodal models will work with types of knowledge that are inherently non-text. For example, SOTA LLMs really suck at electronics, especially analog electronics.
Is MiniMax H3 capable of logical / technical reasoning, or is it purely art oriented?
Suck at what aspect of (analog) electronics specifically? Not contradicting the claim, just want to understand it.
I have not tested yet, but I suspect that LLMs with a harness that can execute code can do SPICE simulations rather ok these days? I have seen MCPs for measurement equipment also, maybe they can even close the physical loop?
> I have seen MCPs for measurement equipment also, maybe they can even close the physical loop?
I'm working on this but for various reasons can't have my physical lab up at the moment. But yes there are many options for connected test equipment that could rather trivially be interacted with via LLM or pretty easy to write libraries.
spice is a very bad simulation. it's not like a unit test or a VM. it works very differently to the real thing, only approximating it in some highly controlled situations.
> work with types of knowledge that are inherently non-text
What are those things exactly? AFAIK, everything we can "know" can be written down, one way or another, even analog circuits.
Also, what SOTA LLMs are you referring to? GPTs been handling analog circuits fine for quite some time, I want to say for at least one year? I've been "pair programming" a bunch of working circuits with GPT models since probably GPT 5 or so.
> AFAIK, everything we can "know" can be written down, one way or another, even analog circuits.
writing down circuit diagrams is like cooking about music.
> what SOTA LLMs are you referring to?
I have done a survey among analog electronics designers just a couple weeks ago and they all said that their forays into LLMs were great for digital electronics, code, and firmware, but for analog they were pretty terrible, with a variety of LLMs, according to everyone.
> everything we can "know" can be written down
Do you know how to ride a bike?
I think that's different though, that's "doing" rather than just "knowing". You can ride a bike without knowing how it works, and obviously vice-versa too. I don't see circuit diagrams as "doing" though, but the soldering part of building circuits definitely is that way though, you can't just read about it and excel first time you pick up an iron, you have to practice and understand it with your body, like bicycling.
ahhhh, that's such a great example
I've had great results on electronics with Claude and Gemini.
what sorts of things have you designed with it?
They can be used to create synthetic training data for robotics. It's not the best approach, though.
i bet they're good for it. however robotics aren't electronics. that's a completely different space.