Hi all
I see all those cool projects about people adding AI to give their home a personality.
I was wondering if there’s a way to install a full local LLM in order to help with coding.
One that would be able to save my automation directly (or create helpers).
One that would know what entities I have and therefore be able to create usefull automations.
Personally I don’t want to talk to my house, I want to talk to my installation.
Is this something feasible? is someone working on this?
I found this link; but it’s too complex for me, plus I wouldn’t have the necessary hardware: Personal Copilot: Train Your Own Coding Assistant
The trouble with using AI to write automations is that there are as many (maybe more) examples of incorrect yaml online as there are of correct yaml - just have a look at the posts in this forum. Since it doesn’t actually know anything - LLMs are just statistical language models - AI is hard put to tell the the difference.
So the right way to write a Blueprint, (well let’s call it best practice and the method suggested in the docs), is to write an automation or script that does a simplified version of that you want first, then add the blueprint magic sauce. Getting the logic working is not always easy, and becomes obfuscated in all that variable substitution stuff, making everything frustrating.
Get what you want working, then make the blueprint out of that by adding the !inputs and variables and such. Try to avoid device triggers and selectors. You can make them work, but people tend to struggle with those when dealing with them outside of the GUI editor context.
There are generally 2 reasons to make a blueprint.
because you want to do the same thing multiple times, like my script blueprint to talk to my broadlink and program buttons to turn on and off my AV equipment. I have re-used that script blueprint about a dozen times in my config.
I understand and I agree.
But at the same time, using an LLM has advanced me: I still need an evening for 1 automation, but now it’s ten times more intelligent (f.e. my light switch on in the evening if I open the door; with LLM it now takes the status of the light in account).
The topic was also an opener: I don’t know much about LLM’s but I understand you need to train them. I assume you can tell it to use this dataset as high quality and another dataset as lower quality?
I also assume that if you would embed this in your application, the LLM could write and execute it’s own test; f.e. asking the user ‘did your income light switch on?’.
But for this to happen, someone smarter than me should train it?