Imagine a standardized API where you provide either your own LLM running locally, your own LLM running in your server (for enthusiasts or companies), or a 3rd party LLM service over the Internet, for your optional AI assistant that you can easily disable.

Regardless of your DE, you could choose if you want an AI assistant and where you want the model to run.

Yeah. I'm really annoyed by this trend of having programs that could function offline require connecting to a server.

Not just hypothetically but practically too. A foss program called koboldai let's you run LLMs locally on your computer and a project that takes advantage of this is the koboldassistant project. You can essentially make your own Alexa,Cortana,Siri whatever that doesn't collect your data and belongs to you