The claim runs against intuition: a page rendered as an image can be represented with roughly ten times fewer units than the equivalent text and still be recovered accurately. Squeeze it to twenty times and accuracy drops to a little over half. The authors present it as a route to long contexts and, along the way, as a forgetting mechanism: old context gets progressively lower resolution until it blurs, much as memory does.
What I took from it
That context is a budget, not an entitlement. Before accepting that "all the documentation goes into the model", I separate what has to stay readable word by word from what only has to be retrievable on demand. Those are different needs with very different costs, and treating them as one is the most common way to inflate a document-and-AI project.
The forgetting mechanism left me a question I now ask in any design: which information is allowed to decay over time, and who decides when. A system that never forgets is not more faithful, it is more expensive and more confusing, because it answers with repealed policy as confidently as with current policy.
And a practical consequence that reorders priorities: the same procedure turns scanned documents into training material at industrial scale. The dead archive every company keeps stops being an obstacle and becomes raw material.
What I left out
The enthusiasm. This is early-stage work and the authors say so plainly; the numbers come from their own benchmarks with their own models. Above all, transcribing a page accurately is not the same as reasoning about what it says, and conflating the two is exactly the mistake that gets paid for later, in a demo that collapses at the first hard question.