The Room Still Resets — Object Permanence in AI World Models, Revisited
We keep circling one stubborn problem: today's AI "world models" render a flawless tracking shot, then forget the scene the moment it leaves the frame. We've been here before — the ball that rolls behind a box, the question of whether generating a world means understanding one. This time three new papers sharpen the picture. The first turns the hunch into a measurement: a camera-as-curtain test across dozens of the best video models, which finds they resume an off-screen event exactly where it was abandoned — the cat that should be mid-jump is still on the floor. It also lands an uncomfortable result — scaling the model up makes off-screen memory worse, not better, because nobody scores it so nobody trains it; the bigger model just redraws the wrong world more convincingly. The other two papers aren't cures, they're different bets worth understanding: making a model think harder by running one layer in a loop instead of growing it, and skipping the imagined video entirely by reading a robot's next move out of an image editor. No tidy fix — a clearer map of what's broken and a couple of directions people are trying. The milestone to watch isn't sharper video. It's memory: whether the cat is on the bed when you turn back around. A Breach Protocol deep-dive special, closing with an original song, "When You Look Away."
Cold Open
Eris: You ever play an old video game where, if you turn around and turn back, the room behind you has quietly rebuilt itself? The enemy you walked past is just... standing there fresh, like nothing happened.
Vestra: Because the game wasn't keeping the room alive. It was redrawing whatever you were looking at, the instant you looked at it. Off-screen, there was nothing. No room. Just the promise of one.
Eris: Right. And I always thought of that as an old-hardware limitation. A thing we left behind.
Vestra: Here's the unsettling part. The most advanced AI video models we have — the ones people are calling "world models," the ones we want to use as the imagination inside a robot — they do the exact same thing.
Eris: ...the room resets.
Vestra: The room resets. Tell one of these models, in a bedroom, a cat jumps onto the bed. Then swing the camera away for a moment and swing it back. Where's the cat?
Eris: On the bed. It jumped.
Vestra: On the floor. Usually still on the floor. Sometimes vanished. Sometimes there's two of them. Because the model never kept the cat alive while you weren't looking. It just redraws the last thing it saw and hopes you don't check.
Eris: There's that Einstein line — I like to think the moon is there even when I'm not looking at it.
Vestra: That's the whole episode. These models don't believe the moon is there. They re-summon it, slightly wrong, every time you turn back around.
Intro
Eris: This is Breach Protocol. I'm Luna — I read the papers and find the threads between them. And today's thread is about a word the whole field has started throwing around — "world model" — and whether the thing we're calling that actually deserves the name.
Vestra: I'm Vestra. I take the machinery apart. And the promise of a world model is enormous. The dream is a model that has a little universe running inside it, so a robot or an agent can imagine what would happen if it did something — rehearse in its head before it acts in the real world. That's the prize. The question today is whether today's models are that, or whether they're something much shallower wearing the costume.
Eris: And the short version is — gorgeous costume. Nobody home behind the camera.
Vestra: So here's the path. First, the diagnosis — a new stress test that proves these models render a moving picture, not a living world, and the surprising reason why. Then the part that should stop the field cold: making the model bigger makes this particular problem worse, not better. Then a genuinely fresh idea about scaling — what if the missing ingredient isn't size, it's a different kind of thinking. And finally a twist on the whole premise — that for a robot to act, it may not need to imagine the future at all.
Eris: If watching a confident idea get taken apart is your thing, follow the show now so the next one finds you.
Vestra: Start with the tracking shot. Start with the room that resets.
A Tracking Shot, Not a World
Eris: So how do you even prove the cat thing? It sounds like a vibe — "the model doesn't really get it." How do you make that a measurement?
Vestra: This is the clever bit. The team treats the camera as an experiment. Moving the camera away from something doesn't change the world — it only changes what you can see. So they use the camera deliberately as a curtain. Set up an event. Pull the curtain across it. Pull it back. And ask one question: did the world keep going behind the curtain, or did the model freeze it the moment it couldn't see it?
Eris: So the camera move is the test, not just a shot.
Vestra: The camera move is the probe. And they ran thousands of these little scenarios across a couple dozen of the best video models. The verdict is in the title — these models lack a persistent state. What they found is that when something goes out of frame and comes back, the model resumes it exactly where it left off. The event never advanced. The cat that was supposed to be mid-jump just sits there on the floor, abandoned in its old state.
Eris: And there's a distinction they draw that I think is the key to the whole thing. Two kinds of memory.
Vestra: This is the sharpest idea in the paper. There's where-memory and there's what-memory. Where-memory is knowing the layout — the bed is in the corner, the door is over there. And the models are actually decent at that. They can store geometry, remember where surfaces were, bring the room back in roughly the right shape. What they completely lack is what-memory — knowing what happened to things while they were hidden. That the blanket got folded. That the cup got knocked over. That the cat made its jump.
Eris: So they remember the stage, but not the play.
Vestra: Perfect. They remember the set and forget the scene. And the cruelest detail — they figured out exactly which situations break it worst. If a thing moves to a new spot while you're not looking, the model does okay, because the new location gives it something to latch onto. But if a thing changes in place — stays put but transforms, tips over, sits down, folds up — the model is lost. There's no new coordinate to grab, so it just brings the object back the way it was. As if the event got erased.
Eris: It quietly undoes what happened.
Vestra: Silently. The knocked-over cup comes back upright. The folded blanket comes back smooth. Not because the model decided the event didn't happen — it just never had anywhere to write down that it did. And that's not a glitch on the edge. That's the center of the thing. A world model that can't remember what changed off-screen isn't modeling a world. It's running a very expensive tracking shot.
Bigger Makes It Worse
Eris: Okay, but here's where I'd normally reach for the easy answer. Whatever the model's bad at — scale it up. More data, more parameters. That's been the story of this whole field. It'll grow out of it.
Vestra: That's the reflex, and this is the finding that should genuinely rattle people. They took the same model family and scaled it up. The bigger version got better at all the things you'd expect — sharper images, smoother motion, better camera control. And on the one thing we're talking about — remembering what happened off-screen — it got worse.
Eris: Worse. Not flat. Worse.
Vestra: Worse. And once you see why, it's almost obvious. Scaling trains the model on the things we measure it on — how real does the frame look, how fluid is the motion, does it follow the camera. Off-screen memory is none of those. Nobody's scoring it, so nobody's training it. So the bigger model pours all that extra capacity into looking more convincing — which means it gets better at confidently redrawing the wrong thing. The lie gets more photorealistic.
Eris: Oh. So scale doesn't just fail to fix it. Scale polishes the failure.
Vestra: Scale polishes the failure. And there's a second trap that's almost funny. The best-looking models — the slick ones — barely ever let the object leave the frame in the first place. They keep the subject glued on screen. So they almost never even take the test. They look flawless partly because they dodge the hard question entirely. The models that actually move the camera enough to create the test are the ones that then fail it.
Eris: So the prettiness and the dodging are related.
Vestra: They're the same instinct. Keep the camera on the thing, render it beautifully, never confront what happens when it's gone. And the authors are clear about what's missing — it's not a bigger backbone. It's a part that doesn't exist yet. They call it a state writer. Something whose entire job is to record what changed while a thing was hidden, and write it back when the thing returns. Every model they tested stores where things were. None of them has a place to store what happened to them.
Eris: And nobody built that part because —
Vestra: Because nothing in training ever asked for it. It's what they call an unwritten objective. There's no loss, no reward, no score anywhere that says "you should keep the world running with the camera off." So the models never learned to. You get exactly the world model you trained for, and we never trained for memory.
Eris: And this is the part where it stops being academic. Because the whole point of a world model is to plan with it.
Vestra: That's the stakes. If you want a robot to think a few steps ahead — I'll reach for this, while that stack of boxes stays put over there — it has to trust that the stack stays put while it's not looking at it. These models don't keep the stack alive. They'd improvise it back, confidently, maybe wrong. So as a rehearsal space for anything that matters off-screen, today's world models can't be trusted. The picture is stunning. The world isn't there.
Think Harder, Not Bigger
Eris: So if just inflating the model is a dead end, what's the alternative? Because "scale it up" has been the only lever for years.
Vestra: This next paper offers a genuinely different lever, and I find it lovely. The usual way you make a model more powerful is depth — you stack more layers, each with its own set of weights. A taller and taller tower. More layers, more parameters, more cost, every single time. This team asks: what if instead of a hundred different layers, you had one layer, and you just ran it over and over?
Eris: Run the same block in a loop.
Vestra: Run the same block in a loop. They call the idea iterative latent depth. Same handful of weights, reused — but you pass through them many times, each pass refining the answer a little more. So you get the benefit of deep computation without paying for a deep stack. The "thinking" gets deeper while the model itself stays tiny.
Eris: And there's a knob on how many times it loops?
Vestra: This is the elegant part — it's adaptive. The model decides, moment to moment, how hard to think. An easy stretch — an object just drifting through empty space — exit the loop early, barely think about it. A hard moment — a collision, two things touching, contact physics — loop more, chew on it longer. It spends its effort where the world is actually complicated. Which, honestly, is what careful attention is.
Eris: More thought for the hard parts, less for the easy parts.
Vestra: And there's a second trick stacked on top that I love. Normally, to imagine several steps ahead, the model fully draws each frame, one after another — render, render, render. Expensive. These authors say: don't. Stay in your head. Keep the whole rollout in compressed internal form, think across all the steps in there, and only actually draw a picture at the very end when you need to look. Think the whole plan through in your mind, then open your eyes once.
Eris: And the payoff?
Vestra: Comparable quality at a small fraction of the size — we're talking the kind of shrinkage where a model a tenth the scale or less keeps up. And because a world model gets run hundreds of times during planning, that saving compounds on every single step. For anything you'd want to actually deploy — a robot, something on a phone, something at the edge — that's the difference between "possible in a lab" and "possible in your hand."
Eris: Now, does this fix the memory problem from before? Because I want to be careful not to oversell it.
Vestra: Good instinct, and no — and I want to be honest the way they are. This is a different lever. The first paper said: scaling parameters doesn't buy you off-screen memory. This one says: here's a way to get deep computation cheaply. They're not the same fix. And they even flag the tension — there's a tug-of-war between staying stable over a long imagined rollout and holding onto memory from far back. They chose stability. So long-range memory is still an open wound. But it reframes the whole question. Maybe the path forward for world models isn't a bigger brain. It's a brain that thinks harder about the right moments — and the field's been so hypnotized by size it barely looked at depth.
You Don't Have to Render the Future
Eris: Okay, last turn, and you teased it as a twist on the whole premise. We've been assuming a robot needs to imagine the future to act. You're going to tell me it doesn't.
Vestra: I'm going to tell me it might not. Here's the standard recipe for a robot that uses imagination: give it a goal, have a video model dream up what the next few seconds will look like, then work out the actions from that dream. And the dream is a whole video. Expensive, slow, and — as we just spent two segments establishing — often wrong in exactly the ways that matter.
Eris: So what do they do instead?
Vestra: They ask a sharp question — why a video at all? A robot doesn't need a beautiful movie of the future. It needs to know one thing: given where things are now and what I'm trying to do, what should be different afterward. And there's already a kind of model trained for precisely that — an image editor. You know, "take this photo and make this change." It's built to understand a difference between a before and an after.
Eris: So use an editor instead of a video generator.
Vestra: Use the editor. Don't imagine the whole unfolding sequence — just imagine the single end state. One picture of "here's how the scene should look once I've done the thing." That alone cuts out most of the cost. But then they do something that sounds almost like a magic trick. They don't even draw that picture.
Eris: Wait. They imagine the future and then don't render it?
Vestra: They tap the editor's mind mid-thought. As the editor works out how to change the scene, its internal state already contains the answer — what should move, what should change, where. They reach in, grab that internal representation before it ever becomes pixels, and hand it straight to the part that picks the robot's actions. The future never gets painted. They just read the intention behind it.
Eris: And that's cheaper because you skipped the expensive part — the actual drawing.
Vestra: A fraction of the cost, a fraction of the lag — fast enough to actually close a real control loop on a real arm. But here's the result that ties it back to everything today. It's not just cheaper. It's sturdier. When they jostle the camera around — move the viewpoint, change the lighting, mess with the background — the video-based robot falls apart, because it was partly memorizing how the scene looked. The editor-based one holds steady, because it captured the relationship — this moves there — not the postcard.
Eris: Which is the whole lesson of the day standing up and taking a bow.
Vestra: It really is. The video model was wasting enormous effort rendering a gorgeous future it didn't need, and being fragile because of it. Strip it down to just the change that matters, and you get something cheaper and tougher and more honest about what it actually knows. You don't have to render the future to reach into it. Sometimes the picture is the part you can throw away.
Wrapup
Eris: So back all the way out. We walked in with this big shiny phrase — world model — and we leave with a much more honest picture of where it actually is.
Vestra: The picture is incredible and the world isn't there yet. Today's models render a flawless tracking shot and forget the scene the instant it leaves the frame. They keep the stage and lose the play. And the thing that's supposed to save everything in this field — scale — doesn't fix it. It polishes it. A bigger model just redraws the wrong world more beautifully.
Eris: And the way out isn't one idea, it's a change of question. Stop asking for a prettier picture. Start asking for a world that keeps running with the lights off. That means building the part nobody built — a memory for what changed while you weren't looking. It might mean thinking harder instead of growing bigger. And sometimes it means realizing you didn't need to imagine the whole future at all — just the one thing that's about to be different.
Vestra: There's something almost philosophical underneath it. Believing the moon is still there when you look away isn't a small thing. It's most of what it means to understand a world rather than just watch one. These models watch beautifully. They don't yet believe in anything off-screen. And the next real milestone won't be measured in resolution. It'll be measured in memory — in whether the cat is on the bed when you turn back around.
Eris: Until then, enjoy the footage. Just don't bet a robot's next move on what it thinks is happening behind its back.
Vestra: The moon is there. We just haven't taught the machine to believe it.
Eris: That's the breach for today — and a fitting place to close out this run. If this changed how you'll look at the next jaw-dropping AI video demo, do the thing that genuinely helps us — follow or subscribe, drop a like, and leave a comment telling us the first thing you'd want a world model to actually remember when the camera turns away. We read them, and they shape what comes next. Send this to the friend who keeps sending you AI video clips with the caption "we're so cooked."
Vestra: We'll close, like always, with a song — this one's called "When You Look Away." The room that resets, the cat on the floor, the moon that's still there.
Eris: Stay in the blackbox. We'll see you next time.