An example of AI coming on-device with small models - this just needs 8GB CPU RAM. Info:
Firefox users often struggle recalling previously visited websites, especially when relying solely on Firefox’s direct-match history search.
When exact keywords are forgotten, it can be hard to find the right input to recall important sites from browsing history. Human memory is not generally based on URLs and page titles, but rather on contextual meaning.
Firefox Solution
To solve this problem, we’re testing client-side natural language querying of history in the address bar .
Once enabled, Firefox will download a small language model (Xenova/all-MiniLM) and create a vector database based on your existing history. This new resource will provide the relational context to enable Firefox history suggestions that are semantically related to your search. All matching is done on-device. No new data collection or model training is involved, and your browsing history doesn’t leave your device. The existing history data on your device is made more rich and referential . This means your private browsing history stays private, which makes it difficult for us to understand whether your experience is improving.
For example, if you’re a basketball fan, you might have visited “NBA.com”. Existing Firefox history search would not match the term “basketball” to NBA.com because the term does not appear in the page title or URL. The Semantic History search improvement would match, as “NBA” and “basketball” are closely related in semantic terms.
Firefox Beta can be installed over the release version and uses the same profile, and has just gotten v142, including this. It is not as new and rapidly-upgraded as “Nightly”.
Testing Campaign, Thru Aug 1:
More detail:
And more local AI is in the works for Firefox…
