I think this will probably come eventually. The thing is right now both the cloud and the self-hosted side of AI is evolving very rapidly, a lot of what is true today wasn’t true 6 months or a year ago, and may not still be true a year in the future.
In my opinion, making specific authoritative recommendations at this point is maybe a little premature. It sounds like you have some experience with the self-hosted LLMs so I’m sure you’ve experienced some of this first hand.
I think the TL;DR of the current unofficial recommendations would be:
- Where privacy is the priority, use a permissively licensed model (e.g. Mistral, Llama 3) that you host and run locally, with an open source privacy respecting frontend (E.g. Ollama).
- Where running a local model is not an option, consider using Duckduckgo Chat or Brave Leo, and understand that they can’t offer perfect privacy and your queries will be shared with the model providers in cases where Brave/DDG aren’t hosting the model themselves.
There is a work-in-progress draft proposal you can check out here. I want to stress that it is a proposal/draft not finished, and not official or necessarily agreed upon.