The Machines That Dream in Physics

The hottest idea in AI isn't a large language model. It's a machine that can close its eyes, imagine what happens next—and be right. In the spring of 2026, Yann LeCun—Turing Award winner, one of the three "godfathers" of deep learning—stood on a stage in Paris, picked up a pen, and balanced it on its tip. Then he asked the room a question a four-year-old could answer. "What happens if I let go?" Every toddler knows. The pen falls. They can't write the equations for it, but they know , deep in their bones, the way you know a dropped mug is bad news before it hits the tile. A large language model, LeCun pointed out, doesn't know it the same way. Ask GPT-5 and it will produce a gorgeous paragraph about the pen's descent—torque, gravity, the clatter on the floor. But it isn't reasoning about the world. It's doing autocomplete. It's guessing the next word from a trillion words it has read, blind to the actual physics happening six inches from LeCun's hand. That gap—between describing the world and understanding it—is the most important fault line in artificial intelligence right now. And an increasingly loud faction of the field believes the industry has spent the last five years sprinting confidently in the wrong direction. For years we've been mesmerized by fluency. LLMs got bigger, hungrier, and startlingly good at the syntax of thought: code, sonnets, legal briefs, therapy-speak. But fluency turned out to be a magic trick, not a mind. Chatbots still fumble anything requiring a gut feel for space, time, cause, and consequence—the stuff that makes a plumber good at their job and a self-driving car safe on a wet road. So the smart money, and a startling amount of actual money, has pivoted to a different bet. The new obsession is the world model : a machine built not to predict the next word , but the next state of reality . Feed it what's happening now, propose an action, and it tells you what the world will look like a moment later. It's the difference between a librarian who has read every manual on building a house and a carpenter who can feel how the wood will splinter. This idea has cracked Big Tech wide open. In late 2025, LeCun walked out of Meta—the company he'd steered for over a decade—to found Advanced Machine Intelligence (AMI) Labs , raising north of $1 billion in seed funding on the bet that the whole industry's flagship technology is a dead end. Around the same time, Fei-Fei Li—the scientist whose ImageNet dataset arguably kicked off the modern AI boom—launched World Labs , which rocketed to a $5 billion valuation chasing something she calls "spatial intelligence." Google DeepMind, a startup called Verses running on 1990s neuroscience, and a dozen others have piled in. The prize, if it works, is enormous: robots that can walk into a kitchen they've never seen and make you a sandwich, factories simulated so faithfully you can test a year of disasters before breakfast, a genuine path to the long-promised (and long-deferred) dream of artificial general intelligence. The catch: world models have a nasty habit of hallucinating themselves into incoherent soup. Whether the field can fix that—compounding errors, a gremlin called exposure bias , and physics that quietly break—is the whole ballgame. Here's how the machines that dream in physics actually work, who's building them, and why some very serious people think this is the road to a mind. The Wall Between Words and Worlds To see why world models are a genuinely different animal—not just a bigger chatbot—you have to understand what an LLM actually is . An LLM lives in a flat, one-dimensional universe made of text. Through the now-famous transformer architecture and its "attention" mechanism, models like OpenAI's GPT-5 or Anthropic's Claude do one thing with superhuman skill: predict the most probable next chunk of text. Their "knowledge" of gravity, of what's hidden behind a chair, of why a glass tips over—all of it is secondhand, inferred from how often certain words appear near other words. Their grasp of thermodynamics is really a grasp of thermodynamics vocabulary . Scale that up and you get something remarkable but lopsided. More parameters tighten the web of word associations—better math, better prose, better mimicry of reasoning. What they don't buy you is a felt sense of the physical world. There's no body, no feedback, no consequences. Which is why LLMs remain brittle exactly where the real world is unforgiving: spatial reasoning, long-horizon planning, and the sort of common-sense physics that never once appears in a training sentence because no human bothers to write it down. A world model is built on the opposite premise. It borrows an idea from cognitive science: that brains—yours, a crow's, an octopus's—are fundamentally prediction engines , constantly running a simulation of what's about to happen and flinching when reality disagrees. A world model tries to internalize the rules of an environment—gravity, friction, momentum, the fact that objects don't wink out of existence when you look away—and then run that simulation forward. Here's the same idea in a table, because the differences are structural, not cosmetic: Large Language Models Generative Video (e.g. Sora) Predictive World Models (JEPA, RTFM) What it predicts The next word, by probability The next pixel/frame The next state of the world, in the abstract What it "sees" 1D stream of text tokens 2D wall of dense pixels Multimodal, spatial, continuous—images, video, physical states How it "reasons" Pattern-matching in context Visual averaging of likely frames "Imagination": simulated rollouts in a compressed mental space Superpower Emergent knowledge, endless training data Gorgeous, cinematic output Physics-aware, generalizes to new tasks, cheap to run Fatal flaw No body, no grounding, confident hallucination Errors snowball; blurs the future into mush Hard to train without collapsing; drift over long horizons The industry's dawning consensus isn't that one side wins. It's that they're puzzle pieces . An LLM is a brilliant reader of human intention—perfect for parsing "clean up the kitchen" into steps. But it needs a partner that actually knows a mug will shatter. The chatbot proposes; the world model checks it against physics. The Prophet of the Latent Space The whole point, in one photo: a robot that acts by imagining. AMI Labs is built on the V-JEPA family of latent world models—the same technology that lets an arm manipulate objects it has never seen, guided only by a goal image. Image: © Meta AI (editorial/press use). No one has been ruder about the LLM gold rush than Yann LeCun. For years he's called the strategy of scaling autoregressive language models "a dead end" on the road to real intelligence—an opinion that grew awkward inside Meta as the company doubled down on exactly that. In November 2025 he left. By the following spring he had over a billion dollars and a new company, AMI Labs, built around a single stubborn conviction. The conviction is this: predicting the world in pixels is both wasteful and dumb. Picture a video model trying to guess the next frame of someone walking down a hallway. To do it in pixel space, the model has to render everything —the weave of the carpet, the exact fall of the light, every strand of hair. Worse, the future is uncertain. The person could go left or right. A pixel-predictor, forced to hedge, splits the difference and paints a smeared, ghostly average of both. Run that forward and the world dissolves into blur. But a robot doesn't need the carpet fibers. It needs to know: person, moving, roughly there, about this fast. It needs structure and momentum, not texture. So why model the texture at all? JEPA: Predicting in Your Head, Not on the Canvas LeCun's answer is the Joint Embedding Predictive Architecture , or JEPA. The trick is to stop predicting pixels entirely and instead predict in latent space —an abstract, compressed shorthand the machine invents for itself. A neural network visualized layer by layer. JEPA makes its predictions in this abstract inner space—reasoning about structure, not repainting textures. Visualization: Serenechan3, CC BY-SA 4.0. It works in two moves. First, an encoder takes raw input—an image, a clip of video—and crushes it down into a compact vector of numbers, deliberately throwing away the incidental junk (lighting, texture) and keeping the meaningful structure (what's where, what's moving). Then a predictor takes that compressed snapshot, plus a nudge representing an action or an uncertainty, and guesses the next compressed snapshot. No rendering. No pixels. Just thought predicting thought. There's a famous way for this to go horribly wrong, and it's almost funny. It's called representation collapse . If the encoder is lazy, it discovers a loophole: map every image to the exact same vector. Now the predictor's job is trivial—the next state always equals the current state, so it scores a perfect zero error—and the model has learned absolutely nothing about the world. It's the machine-learning equivalent of a student who answers every exam question with "42" and technically never contradicts himself. Earlier versions of JEPA (Meta's I-JEPA for images, V-JEPA for video) fended off collapse with a bag of fiddly engineering tricks—stop-gradients, momentum-updated "target" networks—that worked in practice but nobody could really prove worked. LeJEPA: From Recipe to Theorem Then, in late 2025, LeCun and researcher Randall Balestriero published the paper that reads like a mission statement for AMI Labs: LeJEPA , subtitled Provable and Scalable Self-Supervised Learning Without the Heuristics. The headline innovation has an intimidating name— Sketched Isotropic Gaussian Regularization , or SIGReg—and a simple job. It forces the encoder's outputs to spread out evenly, like marbles scattered across a table rather than piled in one corner. That single constraint makes the lazy "map everything to one point" cheat mathematically impossible. The bag of heuristics goes in the bin. What makes the paper a genuine landmark isn't the tidiness, though—it's the guarantee . LeJEPA comes with a proof: under reasonable assumptions, the model provably recovers the world's true hidden variables from messy observations, with a single tuning knob and costs that scale linearly. It turns a lucky recipe into a theorem. For a field that mostly advances by trial, error, and vibes, that's a rare and load-bearing thing. V-JEPA 2: 62 Hours to a Robot That Improvises The proof of the pudding is a robot arm. AMI Labs builds on V-JEPA 2 , a 1.2-billion-parameter video world model trained on over a million hours of internet video. It learns in two stages, and the second one is the jaw-dropper. Stage one is pure observation: no actions, no labels, just watching a million hours of the world happen and learning to fill in masked-out chunks of video—effectively teaching itself the background physics of reality. Stage two bolts on control using a shockingly tiny amount of robot data: just 62 hours of a robot arm fumbling around. Because the model already understands motion and objects from stage one, those 62 hours are enough to let it control a real arm zero-shot —reaching, grasping, and pick-and-placing objects it has never seen, guided only by a goal image and a single ordinary camera. No task-specific training, no environment-specific data collection. You show it a photo of where things should end up, and it figures out how to get there by imagining its way forward. Here's the family tree at a glance: Architecture The big idea How it dodges collapse What it's for I-JEPA Predict the meaning of images, not their pixels EMA "target" networks Computer vision, spatial reasoning V-JEPA 2 Predict features across a million hours of video Masked feature prediction, frozen encoder Zero-shot robot planning, anticipating human action V-JEPA 2.1 Denser prediction across every token; multimodal Deep self-supervision through the network's guts Grasping, depth estimation, scene understanding LeJEPA Provably recover the world's latent variables SIGReg (spread the embeddings out) Stable, heuristic-free foundation models This is LeCun's whole vision in miniature: a machine that can imagine the consequences of an action, weigh the cost, and act—without ever rendering a single pixel. The Geometer: Fei-Fei Li and the Case for 3D Not a photograph, and not a hand-built game level: an entire explorable 3D world conjured by World Labs' Marble from a single image. This is what Fei-Fei Li's "spatial intelligence" looks like once it ships as a product. Image: © World Labs (editorial/press use). Where LeCun wants to compress the world into abstract thought, Fei-Fei Li wants to give machines something more concrete: an explicit, three-dimensional sense of space . "Language literally comes out of everybody's head," she said in a 2026 talk. "There's no language in nature. But the world is far more complex." Language, in her framing, is a thin human overlay—a one-dimensional, digital afterthought. The physical world underneath is irreducible: geometry, occlusion, gravity, the stubborn fact that things have insides and backs and take up room. To crack intelligence, she argues, you need machines that grasp that . She calls it spatial intelligence , and it's the founding bet of World Labs, the company she launched with a trio of graphics and vision heavyweights and a $5 billion valuation. Marble, and "3D as Code" World Labs' flagship, Marble , does something video generators can't. Hand it a single photo, a text prompt, a short clip, or a rough 3D sketch, and it produces a persistent, explorable 3D world —one you can walk through from any angle without it dissolving. Crucially, it doesn't hallucinate frames one at a time. It builds an actual structured environment you can export: as "Gaussian splats" (a slick modern way to represent 3D scenes), as clean meshes for artists, or as stripped-down collision geometry for physics engines. The philosophy is "3D as code." Text became the universal interface for software; World Labs is betting that 3D becomes the universal interface for space —a shared format humans and machines can both generate, edit, and hand back and forth. A companion tool called Chisel lets a creator block out a scene with crude geometric shapes—here's a wall, here's a table-sized box—and lets the model paint in the rich detail on top. Structure and style, cleanly separated. An architect could rough out a building's bones and then audition a dozen interior designs inside it. RTFM: The Renderer That Learned to Remember A LiDAR point cloud of a Texas highway—the kind of explicit 3D structure spatial intelligence is built to capture. Visualization: Rebecca / Equator from USGS 3DEP data, CC BY-SA 4.0. World Labs draws a sharp line between a renderer (makes things look good) and a simulator (makes things behave right). Marble leans toward the former. To push toward real-time interaction, the company built RTFM —the Real-Time Frame Model—a "learned renderer" that runs on a single H100 GPU and generates novel viewpoints fast enough to feel live. Traditional graphics pipelines hand-code geometry, lighting, and reflections with decades-old algorithms. RTFM throws that out and represents the whole world implicitly , inside the activations of a neural network. But its cleverest move solves the oldest bug in AI-generated worlds: things forgetting they exist. Most video models have no memory. Turn around in one and the room behind you is gone—regenerated from scratch, different every time. RTFM instead treats already-seen frames as a spatial memory . Through a technique called context juggling , it pulls up the relevant nearby views whenever it needs to render a patch of space, baking in a quiet assumption that the world is a stable 3D place. Turn your back in an RTFM world and the vase behind you stays exactly where it was. Object permanence—the thing human babies figure out at eight months—finally solved, in a spatial cache. The Rest of the Arena: DeepMind, and a Brain-Inspired Heretic This isn't a two-horse race. Two other approaches are strange and important enough to matter. DeepMind: Worlds That Train Their Own Players Google DeepMind is building the holodeck. Its Genie 3 is an 11-billion-parameter world model that spins up playable, photorealistic 3D environments from a single sentence or image , running at a smooth 20–24 frames per second. Knock over a virtual vase, wander off, come back—the vase is still on the floor where it fell, rendered in crisp 720p. That persistence, learned entirely from watching 200,000+ hours of video and simulation, is the hard-won prize. Then DeepMind drops an agent into those dreamed-up worlds. SIMA 2 , powered by its Gemini models, is a generalist that learned hundreds of language-following skills across commercial video games. Drop SIMA 2 into a Genie 3 world it has never seen, give it a goal in plain English, and it parses the physics of a place that didn't exist a second ago and gets the job done—narrating its own intentions, asking clarifying questions, improving by trial and error. Put the two together and you get something almost eerie: one AI generating endless worlds to train another AI. A self-contained flywheel for teaching machines to act, with the added trick of domain randomization —wildly varying the simulated worlds so that skills learned in the dream transfer to reality. (DeepMind, worth noting, has already proven that structured "world" models can find real truths: AlphaFold cracked protein structures, and its GNoME system discovered 380,000 stable new materials. These aren't parlor tricks—they're verifiable science.) Verses and AXIOM: Betting Against the Transformer And then there's the heretic. Verses AI ignores the entire transformer playbook in favor of an idea from theoretical neuroscience: Karl Friston's Free Energy Principle , the claim that every living thing acts to minimize its own "surprise." Its system, AXIOM , doesn't passively predict the next token or pixel. It runs on active inference —continuously working to minimize surprise either by updating its beliefs to match the world, or by acting on the world to make it match its beliefs. The idea behind AXIOM, drawn by its originator. A system—whether a cell or a brain—is separated from the world by a boundary; it senses, predicts, and acts to keep "surprise" low, either updating its beliefs or changing the world to fit them. AXIOM turns this biological loop into an algorithm. Diagram: Karl Friston (Kfriston), CC BY-SA 3.0. AXIOM carves what it sees into distinct objects and their trajectories, spinning up new object-trackers as things appear and merging redundant ones to stay lean. Because it learns through Bayesian updating rather than the brute-force backpropagation that powers everything else, it learns frame by frame, on the fly, without catastrophic forgetting —the bane of conventional networks. The results are the kind that make people uncomfortable. On the Gameworld-10k benchmark, AXIOM beat Google's own DreamerV3—scoring higher while using less than half the compute and a tiny fraction of the parameters. It's a loud piece of evidence that the reigning "just add more data and GPUs" dogma might not be the only road to a mind. How a Machine Dreams: Training vs. Imagining To really get world models, you have to separate two phases that feel similar but aren't: teaching the simulator, and then using it to think. Building the Simulator Unlike an LLM's simple next-word drill, a world model has to learn two things at once: a compact way to represent a high-dimensional scene, and how that representation changes when something acts on it. It does this by self-supervised learning on oceans of data, absorbing the "background physics" of the world the way you absorbed object permanence—by watching, not by being told. Which surfaces the field's most stubborn gremlin: teacher forcing . During training, it's convenient to always feed the model the real next frame to learn from. But at showtime, there is no real next frame—the model has to run on its own predictions, imperfections and all. That mismatch between the sheltered classroom and the messy real world is called exposure bias , and it's a slow-motion disaster. Fixes like Dynamic Latent Bootstrapping work by gradually weaning the model off ground truth during training—forcing it to eat its own cooking early, so it doesn't choke on it later. Imagining: Thinking Before Acting Here's the payoff, and it's the whole point. A world model lets an agent think before it moves . Instead of reacting, it plays out the future in its head—a formalization of an old control-theory idea called Model Predictive Control . Faced with a task, the agent runs latent rollouts : it simulates hundreds of possible action sequences purely in its compressed mental space, far faster than they could happen in reality. A cost function scores each imagined future against the goal, and the agent picks the winner. It's chess played inside the machine's own head, then executed once. How a world model "thinks before it acts." At each instant it predicts a whole trajectory into the future (the curves to the right of k), scores it against the goal, executes only the immediate next move, then slides the window forward and re-imagines everything. Do this many times a second and you get planning. Diagram: Martin Behrendt, CC BY-SA 3.0. The frontier is making this cheap and long-range: Hierarchical Planning (HWM). Plan far enough ahead and the tree of possibilities explodes; the model loses the plot. HWM fixes this the way humans do—with abstraction. A high-level planner sketches the route in coarse "macro-actions" ("go to the drawer, then the counter"), and those become waypoints for a low-level planner that fills in the millisecond-by-millisecond motor control. The results are stark: on real Franka robot-arm tasks, HWM nailed tricky non-greedy pick-and-place from a single goal image 70% of the time , where a flat, non-hierarchical planner scored a flat zero —while using up to 4× less compute. Differentiable World Models (DWM). Instead of just sampling futures and picking the best, DWM makes the entire imagined rollout differentiable—meaning the agent can use calculus to backpropagate through its own daydream and tune its plan directly, optimizing on the fly for the exact situation in front of it before it ever moves a motor. The Friction of Reality So: is this real, or is it the most expensively funded hype cycle since the metaverse? The honest answer is that world models have one great enemy, and its name is compounding error . Exposure bias means a microscopic mistake at step one gets amplified at step two, and step three, until by step fifty the simulated world has drifted into a physics-defying fever dream. Cute when it's a video demo. Lethal when it's a self-driving car's world model confidently hallucinating an empty road where a pedestrian is standing. Why drift is so hard to beat. This is the Lorenz attractor, the classic picture of "sensitive dependence": the trajectory is colored by time, fading from red to blue, and two starting points a hair apart end up in totally different places. A world model has the same problem—one tiny prediction error, fed back into itself frame after frame, snowballs until the simulated future bears no resemblance to reality. Image: Dschwen, CC BY-SA 3.0. The field knows this is the existential question, and it's throwing everything at it. Patching the Drift Technique What it does The problem it kills AR / Rolling Forcing Trains the model on its own predictions, not ground truth Closes the classroom-vs-reality gap; stops cascading hallucinations Backwards Aggregation (BAgger) Reverses the model's own drifted rollouts to teach it how to undo mistakes Gives the model recovery reflexes instead of hoping it never errs Video RAG (VRAG) Anchors generation to explicit coordinates and a spatial memory buffer Fixes forgetfulness; keeps space coherent over long spans Dynamic Latent Bootstrapping Feeds generated latents back in during training, memory-efficiently Walks the model down inference-realistic error paths early Bolting Physics to the Bone The most rigorous fix doesn't patch the symptoms—it hard-codes the laws. Standard world models let energy and mass appear and vanish because nothing forbids it. Hamiltonian World Models forbid it, by building the actual mathematics of physics into the network's structure. They split the machine's internal state into a "phase space" of position and momentum governed by the equations of classical mechanics, then force the simulation to evolve in a way that conserves energy by construction. The result is a world model that literally cannot dream up a perpetual-motion machine—physically valid, stable over long horizons, a simulator rather than a hallucinator. New Rulers for a New Game The old way to judge these models—does the video look cool?—is dead. In 2026 a cinematic drone shot means nothing if the buildings underneath it violate structural integrity. A new generation of benchmarks measures whether a model is a faithful simulator : WorldModelBench probes seven real-world domains (robotics, driving, industry) and catches subtle cheats like an object quietly changing size—a violation of the conservation of mass. WorldModelGym feeds a model different choices and measures its regret —how much worse off an agent would be trusting the model versus an oracle. World-in-World closes the loop entirely, testing whether the model actually helps a robot navigate, manipulate, and recognize things in the real world. The shift is the whole story: the field is finally grading these things as engines , not as media generators. The Meter That Never Stops: Why World Models Want to Come Home Here's a plot twist almost nobody saw coming. The last five years trained us to believe that serious AI lives in the cloud—that intelligence is something you rent, by the token, from a warehouse full of GPUs in Virginia. World models may quietly invert that assumption. The physics-dreaming machine might be the thing that finally drags AI out of the datacenter and onto the hardware sitting on your desk. To see why, you have to notice a structural difference in how the two kinds of AI consume compute. Episodic vs. Always-On An LLM is transactional . You ask a question, it thinks for a second, it answers, the meter stops. Even a chatty session is a series of discrete little sprints with long idle gaps in between—which is exactly why per-token cloud billing works. You pay for the bursts. A world model doing its actual job—control, planning, keeping a robot or a car or a pair of glasses safely oriented in reality—is continuous . Remember how these things "think": Model Predictive Control, running latent rollouts, simulating dozens or hundreds of possible futures every fraction of a second , then doing it again, and again, for as long as the agent is awake and embodied. The simulation never stops. It can't stop, any more than your sense of balance can take a coffee break. Now do the arithmetic on renting that. A metered cloud API is a taxi with the meter running—fine for a trip across town, ruinous if you leave it idling in your driveway 24 hours a day for a decade. A world model streamed from the cloud is a meter that never stops spinning , for every device, forever. The economics are absurd on their face. Local hardware flips the whole equation. A MacBook, a mini-PC, a phone, a Jetson board bolted to a robot—that silicon is a sunk cost . You already bought it. The marginal cost of one more latent rollout isn't a line item on an invoice; it's a few joules of electricity you're paying for anyway. Continuous simulation is prohibitively expensive as a service and nearly free as a capability you own outright . For a workload that runs every waking millisecond, owning beats renting so decisively it isn't really a contest. The unglamorous future of AI compute? Hardware you already own. A Mac Mini's Neural Engine and a phone's NPU are sunk costs, not metered APIs—which matters enormously for a workload that never stops running. Photo: Chiru-2, CC BY-SA 4.0. Physics Doesn't Wait for Wi-Fi Cost is only the first problem with renting a world model. The second is the speed of light. Physical control loops are brutal about timing. A robot arm or a balancing machine typically needs to close its control loop at 50 hertz or more —a fresh decision every 20 milliseconds or less—or it wobbles, overshoots, and fails. A round trip to a cloud datacenter, on a good day, is tens to low-hundreds of milliseconds, plus jitter, plus the ever-present chance the connection simply hiccups. That's not a latency budget you can shave; it's a non-starter. You cannot catch a falling object over a Wi-Fi connection. This is the opposite of an LLM's situation. A chatbot can happily tolerate a second of round-trip lag—nobody notices. But a world model is an inference-time planner : its whole value is thinking a beat ahead of reality, right next to the sensors feeding it and the motors it commands. It has to be physically close to the action. On-device, that same loop drops to single-digit milliseconds because there's no network hop at all—new edge inference stacks are already pushing camera-to-decision pipelines below 10ms on phone-class hardware. The world model doesn't want to be in the cloud. Physics won't let it. There's a bandwidth version of the same argument. World models don't sip sparse text; they gulp dense, continuous sensor streams—video, depth, IMU—at high frame rates. Piping that firehose to a datacenter around the clock is a bandwidth and battery catastrophe (radios are power-hungry). Processing it where it's captured is the only sane option. Your Living Room Is Not a Training Set You Want to Upload Then there's the data—and this is where "local" stops being an optimization and becomes a moral necessity. A useful personal world model is, by definition, a model of your specific world : the layout of your home, the faces of your family, the choreography of your daily routines, the inside of your workplace. It is the single most intimate dataset a machine could hold. And world models don't learn it once from a public corpus the way an LLM ingests the open web—they learn it by continuous self-supervised observation of your actual environment , refining themselves frame by frame on a stream of your private life. Uploading that to someone's servers, continuously, forever, is a privacy proposition somewhere between "hard no" and "regulatory bonfire." Household-robot researchers already flag it as a pivotal and under-addressed risk; the emerging design pattern for always-on egocentric systems is privacy-by-design , where raw data never leaves the device without explicit consent. But here's the elegant part: the local hardware story and the privacy story are the same story. If the world model runs on-device, the sensitive data it feeds on never has to leave in the first place. On-device self-supervised learning means your living room trains your assistant without your living room ever being uploaded anywhere. (Where models genuinely must share what they learn, techniques like federated learning let them pool insights without pooling footage .) The Tailwinds—and the Honest Caveat Three currents are pushing in the same direction. First, the architectures were designed for this. LeCun's entire latent-space thesis exists to make prediction cheap; AXIOM runs on a minuscule parameter count and learns frame-by-frame with no giant replay buffers; Meta's open EB-JEPA library trains world models in a few hours on a single GPU. These aren't trillion-parameter behemoths straining a datacenter—they're deliberately lean, precisely because they're meant to run all the time. Second, the hardware already shipped. The Neural Engine in a current Mac does tens of trillions of operations per second; a flagship phone's NPU pushes 60–70; robotics boards like NVIDIA's Jetson Thor reach into the thousands. The silicon for on-device world models isn't a roadmap promise—it's on desks and in pockets today. Third, the incentives of everyone but the cloud vendors align: compute wants to come to the data, not the other way around. But intellectual honesty demands the caveat. Not every world model is edge-ready today. The generative branch—the ones that render gorgeous, playable pixels, like DeepMind's 11-billion-parameter Genie 3 or World Labs' RTFM, which still needs a full H100 to hit interactive frame rates—remains firmly datacenter-bound for now. Those are renderers , and rendering is expensive. It's the predictive, latent branch—JEPA, AXIOM, and their descendants—that's genuinely headed for your pocket. The most likely near-term shape is a split : pretrain the heavy, universal physics once in the cloud, then ship a small, personalized model that plans and adapts locally. Which, tidily, maps onto the same division of labor the whole field is converging toward—a cloud-friendly, latency-tolerant language model handling what you mean , and a local, private, real-time world model handling what happens next . The trillion-dollar question hanging over the AI industry is whether intelligence is a metered service or a thing you own. For chatbots, the cloud won. For the machines that dream in physics, the answer may turn out to be the device that's already sitting on your desk. The Horizon: 12 Months, and Then 5 Years The near-term payoff of world models is embodiment—robots that generalize to new environments with almost no new data. Photo: Willy Jackson, CC BY-SA 4.0. The next year belongs to robots. Because architectures like V-JEPA 2.1 and hierarchical planning generalize to new environments without mountains of task-specific data, expect a fast ramp in generalist robots that can improvise in kitchens and warehouses they've never seen. Expect, too, the great reconciliation: LLMs and world models stop competing and start merging into hybrid "Vision-Language-Action" systems—the chatbot as the eloquent front-of-house that understands what you want, the world model as the grounded engine that knows how not to knock the glass off the table. And expect Marble, Chisel, and RTFM to start quietly rewiring game design, CAD, and VR, where "type a sentence, get an editable world" is a genuine superpower. The next five years is where it gets vertiginous. World models are positioned to become the operating system of spatial computing. The first big commercial wave will be high-fidelity digital twins —living simulations of factories, power grids, supply chains—where planners run millions of counterfactual disasters in latent space before committing a single real dollar. Push further and the models get small enough to live on your body: a lightweight JEPA running on a wearable, building a private model of your surroundings on-device and warning a visually impaired user about the curb ahead—no cloud, no lag. And at the far end sits the reason any of this is funded at all. If the exposure-bias problem truly falls—if LeJEPA's proofs and Hamiltonian physics scale to open, chaotic, real environments—then LeCun's wager pays off. Machines would hold a persistent, layered understanding of reality. They would learn by watching, foresee the consequences of their actions, and plan their way through a messy world. The intelligence we've been building so far can describe the world with breathtaking fluency. The next kind won't just talk about reality. It will simulate it, act inside it, and—if the people betting billions are right—come to inhabit it alongside us. Image credits Genie 3 generated world — © Google DeepMind, from the official Genie 3 model page . Used for editorial/press illustration. V-JEPA robot (AMI Labs tech lineage) — © Meta AI, from the official V-JEPA 2 research page . Used for editorial/press illustration. Marble-generated 3D world — © World Labs, from the official Marble world model blog . Used for editorial/press illustration. LiDAR point cloud — Rebecca / Equator, from USGS 3DEP data, CC BY-SA 4.0 , via Wikimedia Commons . Humanoid robot (Ameca) — Willy Jackson, CC BY-SA 4.0 , via Wikimedia Commons . Mac Mini M4 — Chiru-2, CC BY-SA 4.0 , via Wikimedia Commons . Neural network visualization — Serenechan3, CC BY-SA 4.0 , via Wikimedia Commons . Model Predictive Control diagram — Martin Behrendt, CC BY-SA 3.0 , via Wikimedia Commons (rasterized from SVG). Lorenz attractor — Dschwen, CC BY-SA 3.0 , via Wikimedia Commons (rasterized from SVG). Free Energy Principle / Markov blanket diagram — Karl Friston (Kfriston), CC BY-SA 3.0 , via Wikimedia Commons . Sources & further reading Paradigm shift & LLM critique — Yann LeCun's critique of AGI feasibility (36Kr); "Why Yann LeCun Left Meta" (Fast Company); "When AI Learns How the World Works" (Goldman Sachs). AMI Labs — "How Yann LeCun's Startup Challenges the Logic Behind Today's A.I. Race" (Observer); "AMI Labs Raises $1.03B to Build Beyond LLMs" (StartupHub.ai); Wikipedia . JEPA architecture — "What Is JEPA?" (Turing Post); "I-JEPA" (Meta AI); "V-JEPA 2 world model" (Meta AI); V-JEPA 2 paper ; V-JEPA 2.1 paper . LeJEPA — LeCun & Balestriero, "LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics" ; "When Does LeJEPA Learn a World Model?" . World Labs — Marble, RTFM, spatial intelligence — World Labs "About" ; "Marble: A Multimodal World Model" ; "3D as Code" ; "RTFM: A Real-Time Frame Model" ; "Why Fei-Fei Li Thinks Spatial Intelligence Is AI's Next Frontier" (VAST Data); $5B valuation report (SiliconANGLE). DeepMind Genie 3 & SIMA 2 — Genie 3 model page ; "SIMA 2: A Gemini-Powered AI Agent for 3D Virtual Worlds" . Verses & AXIOM (active inference) — "AXIOM: Learning to Play Games in Minutes with Expanding Object-Centric Models" ; "Active Inference Explained" ; VERSES independent validation . Planning & inference-time methods — "Hierarchical Planning with Latent World Models" ( project page ); "Model Predictive Control with Differentiable World Models" ; "Sword: Style-Robust World Models via Dynamic Latent Bootstrapping" . Drift mitigation & Hamiltonian physics — "BAgger: Backwards Aggregation for Mitigating Drift" (CVPR 2026); "VRAG: Learning World Models for Interactive Video Generation" ; "Physically Native World Models: A Hamiltonian Perspective" ; "PH-Dreamer" . Benchmarks — WorldModelBench ( NeurIPS 2025 poster ); "WorldModelGym" (Reka AI); "WBench" . Robotics, on-device & privacy — "World Model for Robot Learning: A Comprehensive Survey" ; "Pretrained to Imagine, Fine-Tuned to Act" (NVIDIA); "Embodied Foundation Models at the Edge" ; "Privacy Risks in Reinforcement Learning for Household Robots" ; "Lightweight EB-JEPA" .