[greg]

Engineer on the team here. Some of the interesting challenges of doing ChatGPT for circuit design were:

1) How to communicate the circuit connections to ChatGPT in the text prompt. We ended up just using the standard netlist format. Apparently it knows about netlists from its training data. It also knows about datasheets.

2) How to get the needed accuracy and precision for questions like "calculate the resistance needed between component A and B". For that, we adopted the popular langchain framework to integrate specific "tools" into the AI for specific recognizable problems.

[arcticbull]

Out of curiosity, I noticed in the video that it recommended for a voltage regulator either an LM7805 or an LM1117 - either a super inefficient 5V regulator that has no business in new designs, or a 1.2V low-dropout 800mA regulator that is probably not appropriate to the other components it recommended (i.e. an LM7805 -> LM1117 would be 36% efficient at best).

It then also recommends a LiPo battery with a nominal 3.7V which obviously wouldn't be compatible with its LM7805 recommendation.

It also recommends an ESP8266 which is a 3.3V part so neither of the recommended regulators, nor the battery alone, would be appropriate.

Also, it recommends a 168MHz STM ARM micro - but it already has an ESP8266 with GPIO.

I'm not sure what it's suggesting I build but it doesn't really sound like a working environmental radiation logger ;)

How should I think about this?

[rock_hard]

In this case it seems to be giving examples of common ICs that are voltage regulators - these answers to high level questions are pretty vague by default.

We've had good results by asking follow up questions like "what power supply IC could I use to power a 3.3v microcontroller from a 3.7V nominal LiPo battery over the whole battery voltage range?".

I asked that question and I got: LM3940 and AMS1117-3.3 (LDO, high drop out but probably ok for low current - one isn't in stock but it doesn't have stock info yet), and TPS62203 and ADP2108 (bucks, both seem appropriate). These answers should get better as Copilot gets access to more and more tools.

[assimpleaspossi]

I find comments similar to yours about ChatGPT all over the internet. I finally took the time to ask that some questions over the last few days. All the answers it gave were mundane and expected from any click bait web site you can find. I Googled the same questions and found similar answers.

Some personal questions it refused to answer or give suggestions because "only you can decide" was the response.

I'm very suspicious of all this. I'm thinking the only real benefit--some may say--is the chat aspect. You can carry on a conversation rather than having to determine a proper Google search over and over again.

Which makes me think there is some hype involved.

[evancox100]

I mean the more fundamental problem with the power supply is it recommending a "USB-to-Serial converter for powering the device." (See second example image.)

I'm overall bullish on AI systems but I'm not sure this space has enough detailed context to train. Much if this knowledgeis in-house/proprietary. We'll see how fast these are able to improve over time.

[stavros]

Why is the battery alone not appropriate for the ESP8266? It runs just fine off a lithium battery.

[joshuanapoli]

How do you deal with designs whose netlist or part list exceeds the context size limit?

[rock_hard]

It’s indeed a challenge

For many applications such as talking to pdfs you can use a vectors db like pinecone

But that doesn’t work well for schematics because for of the interesting use cases you need access to the whole thing.

Here is what’s been working for us

1. Extreme compression and summarization of the data we dynamically put in the context 2. Leverage the extreme amount of world knowledge the cutting edge models have

[ugh123]

Do you do any fine tuning of the model for circuit analysis and other hardware design literature?

[ZephyrBlu]

I'm curious if you fine-tuned a GPT model with specific hardware-related examples. Based on the OpenAI docs it seems like a pain to get enough good prompt:response pairs for fine-tuning (A few hundred minimum).