4 ms·
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 electronic
by ddevnyc 2mo ago
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?
- embedding-shape 2mo ago> 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.
- ddevnyc 2mo ago> 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.
- chrisjj 2mo ago> everything we can "know" can be written down Do you know how to ride a bike?
- ddevnyc 2mo agoahhhh, that's such a great example
- embedding-shape 2mo agoI 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.
- echelon 2mo agoThey can be used to create synthetic training data for robotics. It's not the best approach, though.
- ddevnyc 2mo agoi bet they're good for it. however robotics aren't electronics. that's a completely different space.
- chpatrick 2mo agoI've had great results on electronics with Claude and Gemini.
- jononor 2mo agoSuck 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?
- ddevnyc 2mo agospice 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.
- adrian_b 2mo agoIt depends on how you hold it. If you have accurate device models, and you know how to use simulations, you can predict quite well the behavior of the analog circuits that you design. Simulations that work like the real thing are an absolute necessity in the industry. It is frequent to design the analog parts of some integrated circuit, when you might have to wait months to get prototype samples that you can measure in the laboratory. If the prototypes do not work exactly as expected, you will probably have the chance to make once a correction of the schematics and/or layout, which will be incorporated in the final product. But when not even the second try does not work as designed, that is likely to be a failure, because any other redesign would add months and many millions of $ until the product can reach the market. So "spice" simulations that can be trusted are a necessity. When designing integrated circuits you cannot use empirical methods that work when you make manually a single-use board, where you can replace components or adjust resistors, capacitors or inductors until you get what you want. For integrated circuits, you must know in advance how the circuit will work, because you cannot tweak it. It is true however that it is easy to configure simulations that will give non-sense results, due to lack of convergence or other numerical errors, so to avoid this, one typically needs both experience with laboratory measurements and experience in using the simulator. I doubt that an LLM would successfully run SPICE simulations, because there are few examples of written instructions of how to do this in the right way (which varies depending on the domain of applications, and each SPICE-like simulator may have its own quirks).