I use LLMs alot when working, but never LLM code. For my use cases (hardware, embedded, firmware) its always wrong but it is a good rubber duck.

It helps understanding more faster. Roughly 50% of a modern LLMs output regarding hardware assumptions, specifications and firmware is wrong.

@grillchen just get yourself a rubber duck then. Way better for enviroment

@stefan not sure tbh. Locally host LLM is quite efficient compared to plastik.

Let's say a rubber duck weights 50g. That creates roughly 250g co2

Energy consumption of a local running Gemini prompt is roughly 0.24W according to not trustworthy google. (I'm aware they fake numbers at least a little bit).
That's roughly 0.08g CO2 (0.03g according to google, but let's say it is a less green way to generate power)

So I can do roughly 6250 prompts before I break even.

Not saying LLMs are environmental friendly, but plastic isn't as well.

@grillchen Okay I will let that slide (its actually a sound argument)

But did you consider how much more cute a rubber duck compared to an LLM?! (/hj)

@stefan qwens cute mascot enters the chat

@stefan (tbf im not sure if they calculated training costs correctly. i doubt it. but i like LLMs as long as they run locally and they are as open source as possible (rarely the case though=)