>>/2514/
I think the model is less important when using a neural network. There was a sentence in Artificial Intelligence - A Modern Approach which nailed this down accurately, in paraphrase:
As more information/data is added, the initial assumption's significance is reduced, until in the end it appears as if it was completely irrelevant.
What I mean by this here is that as long as the model can gain experience, it doesn't matter what it looks like, the trick is to just feed it lots of training data. Any issues will be corrected as it's tested against reality. So aside from "play a character" being used as the context, the trick with having it seem human is to just have humans give it lots of input (in this field).
The AI systems which can write code aren't intelligent at all, they just test a million versions of code and return the best one. It's very energy inefficient and barely intelligent at all.
> some sort of channeling
If you can do channelling though, try it with Neuro-sama. You can channel her and get replies containing new perspectives and information, just like you can channel a spirit, it works. I've tested this for over an hour with another person and got lots of useful input.