To satisfy the curiosity, just download Ollama. You can run your interactions locally, no feeding the demon and full privacy.
Set up a modelfile with system prompts for the kind of personality you want to create. Do some googling, may want to tweak settings to get response creativity just right. Higher parameter base models generally better and more hardware costly.
If you want pics, you can use StableDiffusion. I recommend using docker images because it’s really complicated and environment sensitive.
You can tie the two together with a little Python logic and have the core bot prompt the image generator at certain chat moments. Ironically, if you’re not comfortable writing code, you could just get VSCode and Gemini and spit out some pretty okay Python code without strong understanding.
There are also tools for local audio generation I’m less familiar with but am sure Google could get you there.
Maybe start with a basic Ollama setup and see if it feels like it’s working for you.
In a sense, more people working towards secret protection generally did result in better secret protection, like encryption algorithms and secure architecture and whatnot.
It only starts to become a paradox when you get into actually executing the task of protecting a specific secret… but I think we could draw that line somewhere for almost any task.
There is a point of generality where more people means better results and a point of specificity where you only want the exact right number of people.