← Invisible Machines · Español
This is a working model of the most common AI in your life: a feed that picks what you see. Tap the posts that pull your attention. Underneath, a recommender is learning from every tap — and quietly deciding what you get shown next. Watch the map on the right. Nothing here is saved or sent; it runs entirely on your device.
You only ever tapped things you found interesting. You never chose to narrow your world — yet under the engagement goal it narrowed anyway. Here's the mechanism, and it's real: the recommender can't see what you value, only what you click. The most clickable posts are the most intense ones, so those get shown more, so they get clicked more. The loop tightens around a smaller, louder slice of the world — pulling you (the red marker) away from the broad, balanced place you started (the green one) toward the loudest corner it can find.
Then you flipped the goal to wellbeing, and — same you, same moment — the feed opened back up. That's the whole point: how this power gets used is a design choice, not a law of nature. A feed optimized for your reflective interest instead of your reflexive attention is entirely buildable. It just has to be chosen.
That gap — between the power these systems hold over a billion days of attention, and the wisdom with which that power is aimed — is the thing worth closing.
Made to be understood, not monetized. No accounts, no tracking, nothing leaves your device — open the page source and check. Free to copy and reuse (CC0). This is a simplified model built to make a real dynamic visible; the arithmetic is illustrative, the mechanism is not.