[Zirconocene]

I don't want to retype things again, so for those not on RL, here's some fun things I was turned on to with AI tools:

https://www.carpokes.com/viewtopic.php?t=4626

Cheers

[worf]

Well, for a dataset where N=2, seems pretty Ok.

Needs a better test. Feed it all the wiring diagrams and ask it what should be tested to make sure the alarm works and where those components are located...

[hernanca]

I am in agreement with the feedback in the carpokes discussion from PSU_Crash (an EE) and johnb.

"AI" in popular use today usually means Large Language Models (LLM's). This is the frontier, today. The underlying AI technology is Machine Learning (ML) via Neural Networks which started in the 1950's (if not sooner). ML is a tried and true pattern matching and pattern identifying technique based on statistical methods. It is what gives us facial recognition, voice recognition, handwriting recognition, tumor recognition....

LLM's are ML applied to text and are mind-boggingly large statistical representations of all the text of the mind-boggingly large number of documents they have been trained on. The interfaces to these LLM's keep getting augmented with various tricks (such as pre-processing a complex question into more focused sub-questions) to make them more accurate in their responses. However, it is still a pattern matching guessing game underneath.

I don't know what a "SPICE netlist" is for documenting a circuit diagram, but if it follows a syntax to represent a circuit diagram, then I agree that input could improve the "correctness" of the LLM output compared to feeding it a complex diagram. The LLM has to map that complex diagram (based on prior training) to a semantic representation and that mapping will likely not be as accurate as mapping from a syntax-based representation of a complex circuit diagram (again, based on prior training). The patterns in syntax-based text are, naturally, more reliably (and completely) identifiable during training than those that can be found from training on diagrams. (Note: code follows syntax, too... hmmm). Under the covers, the inputs (your query and any context you provide) are all mapped to statistically-driven semantic "representations", which then drive the output.

As suggested, look into prompt engineering. I am using some AI tools at work and I always tell them: "Don't make anything up". Another good technique to help guide the output is to tell it to act in a role, e.g., "Perform as an Electrical Engineer". You want to help guide the otherwise-blind LLM to the correct areas of its vast multi-dimensional semantic model, where the BS output it generates appears... more pleasing....