Forcing everything into a plug-in is architecturally more complex, and less performant... I'm imagining proxying from native code through JavaScript APIs, then back to native code for LLM operations and context storage. But might lead to creation of some new AI extension APIs.
So now we're debating compile-time feature flags vs run-time, and the overhead of running/maintaining multiple build configs. And picking good names for each... "Firefox Pro with AI" vs "Firefox Lite for Engineers". This isn't what Mozilla needs to be focusing on right now, imo.
Is it using AI if you don't click on the feature's related button? AFAIK when I choose to translate a page or a selection it only starts working when I do it.