I haven't read the spec entirely, but couldn't "infer" mean "check with a compressed list of millions of urls already categorized offline and cached on disk" ?

... using an AI that's been trained on browsing history captured from billions of users. Pretty sure that's what "infer" means.

I'm responding to questions on whether or not this could work - "Isn't this just naive and may not actually produce useful results".

It could work in the degenerated case of just having a giant dictionary of the top hundred million websites and all their topics, so it stands to reason it could work with "ai" too. Each region's browsing behavior is so "head heavy" I would think you'd rarely need to interpolate to see things that weren't in the "training set" of offline categorized sites.

Yes, it is a classifier doing drum roll please topic modelling on the content you are looking at. But it runs the model locally so your privacy is totally preserved... right? Not that I consented to my CPU/GPU cycles being used against me in this way though.