Microsoft just put some
research out showing how small data set algos can perform as well as large foundation models, essentially by training off those models.
Which demonstrates the possibility over having your own open source AI, using a combination of personal hardware running open source algos trained off large privately maintained foundation models.
I think there will be an arms race between the large private models and the competition to provide the most accurate, most effective responses. And open source algos that train from those. The race will be between providing end users with answers with a bias in favour of those who own the private models, and the open source algos that train off them with a bias in favour of the user.
I haven't heard much about AI's leveraging qualitative data, specifically personally generated data. That's still the real dream. Ongoing, hi-fidelity observation and assessment of the organism, help in making decisions that most favour that organism. Which isn't something I would want occurring outside of my direct control. If this kind of AI integration was going to happen, this looks like a path towards that.