> Eventually, this arms race ends with a computer vision model that looks at the screen, classifies visual elements as ads, and draws a rectangle over anything that looks like an ad.

Surely a better way to think of this is that a person's AI agent will simply find, via whatever means necessary, the information that the person is interested in. It's hard to imagine the person will still be operating on the old "pull" model where they deliberately visit a specific website or open a specific app.