"Firefox and AI" has produced strong emotions. I've tested the latest version of the #Mozilla browser and made a few observations:
- the "AI" features are just a web page loaded as a "sidebar"
- the "AI" is a single-page web application which the user can give access to pages or selections
- if the single-page application doesn't meet certain implied UI patterns the feature breaks
- the Web-UI in llama-server works, the one llama-swap breaks it (as seen in the screenshot)
The takeaway is this:
1. the "AI" feature is useful, especially when used with a self-hosted RAG (Retrieval Augmented Generation) like Open WebUI, or a simple LLM on a local GPU
2. the page-scraping feature (from the screen-reader API) is good, better than what the extensions Page Assist uses
3. 3rd party extensions, like Page Assist, have far better LLM support (an API endpoint just works)
4. the real problem is trusting OpenAI, and the like, with the feature.