Yo from another concerned user
The discussion thread you tried to get traction with is very much my style. I've done a lot of work surrounding personality, behaviour and socially intelligent AI training. I took the solo route though. I never reached out for live participation like this. I really like the idea.
I train models on a corpus of 100% my conversational prompts to AI models. I distilled my conversations with frontier models and flipped the roles conversationally. The model prompts as the user, I become the assistant and then I train on response only. I use a 10K dataset modest by most standards but the change in personality and tone when you remove the diversity of scraped data is actually noticeable immediately (if your dataset is diverse enough).
I myself am a bartender so a lot of my interactions with AI models come across as friendly banter between coworkers while waiting for a brain atlas to finish building or talking about philosophical nonsense while a set of SAE weights is training. I also dodge the hubris claims around training an AI model to "be me" by stating the fact that it doesn't matter where you are, if you're feeling lonely or excited and want to share something you can always talk to a bartender and they will always respond to match your energy. So I offer a model that's approachable and has a unique ability to listen.
I've been doing this work with purpose and direction since probably last November? You don't see much interest surrounding the culture of how the future of AI interaction is being formed today. It's buried beneath standard benchmarks and the theory that more data is greater than sharper data when it comes to personality or individuality with the argument always stating that the more a model has to pull from the better they can respond. I'm curious what would you prefer? A model with access to an increasingly larger portion of our populations personalities through online scraping with the ability to change cadence verbosity and overall performance or a sharper single personality that tries to match your cold start prompt as best as possible with the possibility that you and that model just don't have a very great conversation?