Mozilla’s camp has a clear direction that they want to use on-device AI processing features due to user privacy concerns. Data is not sent back to servers. One clear example is the web page translation system that uses on-device translation software.
Recently, Firefox has added a new feature with a similar concept, called Link Preview. It previews the content of the destination link when you hover the cursor before clicking (the method is to point the cursor at the link and press Shift+Alt).
The backend process involves Firefox loading only the HTML content of specific web pages for parsing without the need to load the entire page. It extracts crucial information like meta tags and main content of the page (leveraging the existing Reader View feature). It then summarizes using the wllama model (llama.cpp Wasm version) and SmolLM2, a small model with 360 million parameters.
Mozilla states they selected these models based on performance and speed, displaying the first summary within 4 seconds and subsequent topics within 1 second each. This feature is still experimental through Firefox Labs and requires downloading the 369MB model for initial use.
TLDR: Mozilla focuses on on-device AI processing for user privacy, demonstrated through features like web page translation and Link Preview in Firefox, with backend processes summarizing content quickly using specific models for efficiency.
Leave a Comment