Ask HN: What extent can LLMs be used to simulate the operation of silicon chips?

[K33P4D]

If they are complex enough, can they simulate data handling and manipulation, for e.g. amplifying a soundtrack while the conventional norm would be to run the audio track through a set of hardware devices for amplification.

I have also read instances where LLMs are used as eyes and render the environment on a virtual canvas as they "see it", they are simulating the operations of a complex biological organ and sense.

[wmf]

Sure, LLMs can eventually simulate anything in a lossy, ridiculously inefficient way. It's not a good idea though.

I'm somewhat reminded of the paper "Could a Neuroscientist Understand a Microprocessor?" which calls out the weakness of black-box analysis and simulation techniques. https://journals.plos.org/ploscompbiol/article?id=10.1371/jo...

[GianFabien → wmf]

LLMs can only re-mix what already exists on the internet. If copious amounts of information doesn't already exist on the internet about specific subject matter, then at best you'll get some great sounding nonsense.

[wmf → GianFabien]

Yeah, I'm assuming "simulate a chip" means the LLM would be trained on a ton of documentation related to the chip. Otherwise you'd have no chance.

[ATMLOTTOBEER]

To an arbitrary extent. It would probably be really inefficient though. See - universal approximation theorem.

https://xuwd11.github.io/am207/wiki/functorch.html#:~:text=A....

[ekimehtor]

I suppose you could use the LLM to generate a function that would perform the operation you want. I know image classifiers have been used for audio, the audio is converted into a waveform Image and then processed. It should also be possible for AI to generate schematics including logic gates or even functional ICs such as amplifiers or possibly the entire catalogue of 7400 series chips.