Core argument: An LLM predicts what text comes next. A world model predicts what state of the world comes next, after an action. The first learns from descriptions of the world; the second learns from the world itself. They are not competing versions of the same product. They are different layers, and the most interesting AI of the next few years will come from systems that need both.
This is the second chapter of an ongoing series on world models. The first chapter asked why language models hallucinate; you can read it here. This one explains what researchers mean when they say a model should “understand the world” — and why some of the most famous names in AI are betting enormous sums that the next breakthrough comes from outside language.
1. The short answer
Every large language model, the technology behind ChatGPT, Claude, and Gemini, runs on one trick: given some text, predict the piece of text that plausibly comes next. Answers, code, essays — all of it is assembled from that single mechanism, trained on an enormous amount of human writing.
A world model does something structurally different. Given the current state of an environment and an action, it predicts the state that follows. State can mean camera frames, sensor readings, or the position of objects on a table. Action can mean “move the arm left,” “turn the wheel,” or “walk through this door.” The model is trained not on writing about the world but on recordings of the world doing things: video, simulation, robot logs.
The distinction in one sentence: an LLM has read a billion descriptions of the world; a world model has, in effect, watched the world move. One knows what people said happens when a glass slides off a table. The other can predict that it falls.
2. The line cook, not the cookbook
Picture two people. The first has memorized every cookbook ever printed. Ask about béchamel and they recite the ratio, the temperature, the classical variations. The second is a line cook who has burned a thousand pans. Their knowledge is less complete on paper and far richer in every way that matters: they know what dough feels like when it is about to over-proof, and they catch the sauce before it splits.
Most of what separates them is a different kind of knowledge. The cookbook reader has the world as described. The cook has the world as experienced — including all the small consequences nobody bothers to write down, because writers assume you know them.
This is the bet, roughly. An LLM is the cookbook reader, magnificently well-read and physically innocent. A world model is an attempt at the cook: something that has watched cause and effect enough times to expect them.
The analogy breaks in one place, and the break matters more than the analogy. The line cook gets real physics as feedback, every second, with actual consequences. A world model trained inside a simulator gets feedback only as good as its simulator. If the simulated glass shatters the wrong way, the model learns the wrong physics perfectly, with total confidence. Everything depends on how honest the training environment is — which is why the quality of simulators and video data is now a serious research question, not a footnote.
3. What each system actually does
The LLM side we covered in the previous chapter, so one paragraph will do. A language model compresses an enormous amount of text and then predicts plausible continuations. Because its only channel to reality is more text — other people’s summaries, already lossy — it can produce confident statements with nothing behind them. That is the hallucination mechanism, and it is a property of learning from descriptions rather than a bug someone forgot to patch.
A world model learns what researchers write as a transition function: given state and action, predict the next state. Three uses fall out of that, each building on the last.
Prediction: tell me what happens next. This is the base case, and it is what makes the other two possible.
Planning: because the model can roll the world forward in its head, a system can try actions mentally before committing to any of them, and pick the one whose imagined outcome looks best.
Imagination training: once the model is good enough, you can generate experience inside it — endless practice situations for an agent, without paying for real-world trials. Reinforcement learning researchers have pursued this for years; the Dreamer line of research calls it imagination training, and its latest version trains agents entirely inside learned world models. It is the cheapest route to let a robot practice ten thousand failures without breaking ten thousand mugs.
A concrete anchor for what this looks like today: Meta’s V-JEPA 2, released in June 2025. It is a 1.2-billion-parameter model trained on more than a million hours of unlabeled video. One design choice matters a lot: it does not predict pixels. It predicts in an abstract representation space, on the theory that a model forced to summarize “what is where and what is moving” learns what matters, while a model predicting every texture wastes its capacity on the appearance of things. The action-conditioned version, fine-tuned on roughly 62 hours of robot-arm footage, was then deployed on Franka robot arms in two separate labs with no robot-specific training — and picked up and placed objects with success rates reported between about 65 and 80 percent. Zero-shot, in the sense that the specific robots and rooms were never part of its training.
That last sentence is the whole point. Nobody handed the robot a description of the room. The model had watched enough of the world to expect how objects behave, and planning against that expectation worked well enough to be useful.
4. The money is on both sides
If you follow this debate only through headlines, you come away thinking the field has split into two armies. The funding says something more interesting.
Yann LeCun, Turing Award winner and formerly Meta’s chief AI scientist, has argued for years that autoregressive generation — predicting the next token or the next pixel — is the wrong road to machines that actually understand. In March 2026 he put serious money behind that position: his new company, Advanced Machine Intelligence (AMI Labs), raised a 3.5 billion pre-money valuation, reported as the largest seed round in European history, with an explicit mission to build world models rather than larger language models.
Google DeepMind, meanwhile, refuses to pick. Gemini keeps improving on the language side. On the world side, DeepMind previewed Genie 3 in August 2025: a model that generates interactive 3D environments from a text and image prompt, in real time at 24 frames per second, worlds you can walk around in rather than watch. In early 2026 it began widening access through Project Genie. One lab, both layers, deliberately.
And OpenAI has marketed Sora, its video generator, with the phrase “world simulator” since its early technical report — a claim that sounds strong until you read the same report’s own caveats, which note that Sora gets basic interactions like glass shattering wrong. Independent evaluations have been blunter. A survey titled “Is Sora a World Simulator?” and an ICML 2025 paper that tested video generators against physical laws both reached versions of the same conclusion: the videos look right, and the physics inside them systematically does not hold up — objects appear, vanish, or fail to conserve momentum in ways no real table ever allowed.
So the honest reading of the landscape is not “world models versus LLMs, pick a winner.” It is that the same organizations are building both, because the two layers do different jobs. The real argument inside the field is narrower and more interesting: which layer is the bottleneck on the way to more capable systems, and how much genuine understanding a convincing video actually contains. Those are open questions, and Chapter 8 of this series will come back to the second one.
5. My take
Three positions I hold with varying confidence.
First, and highest confidence: the distinction is real, and the complementarity is the actual story. Language models carry the semantic layer — what things are called, what humans have written about them. World models carry the dynamics layer — what happens when things move. A robot that can read the manual and practice in imagination needs both; neither is a substitute for the other. The “which one wins” framing sells better and explains less.
Second: treat “world model” as a marketing word until proven otherwise. The honest test is simple. Can you act in it? Does it stay consistent when you look back? Can it answer a what-if question about a situation you describe? A model that fails two of the three is a video generator with a good publicist. That test will matter more as the term spreads through press releases, and it is the yardstick I will apply throughout this series.
Third, medium confidence: for people deciding what to actually use, this is not a choice you face today. LLMs are products you can call; world models mostly live inside robotics and driving stacks or inside research previews. The practical takeaway is not “switch tools.” It is to notice that the tools you use in five years will likely assume both layers exist — and that the people building them are hiring accordingly.
6. Questions people actually ask
Do world models replace LLMs? No. They take different inputs and produce different outputs. A model that predicts camera frames has no way to answer an email, and a model that predicts text has no way to catch a falling cup. Systems that need both — and most embodied systems do — will use both.
Is Sora a world model, then? It is a video generator whose marketing uses the phrase “world simulator.” OpenAI’s own technical report lists its physics failures, and independent evaluations find systematic violations of basic mechanics. When video models gain reliable action control and consistency over time, the label gets easier to defend. By the honest test — act in it, look back, what-if — Sora today fails at least two of three.
Can I build with world models today? Mostly only indirectly. The technology ships inside autonomous-driving development, robot training pipelines, and limited research previews like Genie. There is no general-purpose “world model API” for everyday developers yet. If that changes, it will be big news, and this series will cover it.
Sources
- Meta AI, “Introducing the V-JEPA 2 world model”, June 11, 2025.
- “V-JEPA 2 and V-JEPA 2-AC”, arXiv:2506.09985.
- CNBC, “Meta launches AI world model to advance robotics, self-driving cars”, June 11, 2025.
- eWeek, “Yann LeCun, Meta’s former AI chief, launches $1B world-model startup AMI”, March 10, 2026; see also Dealroom coverage of the round.
- Google DeepMind, “Genie 3: A new frontier for world models”, August 2025.
- OpenAI, “Video generation models as world simulators”, technical report, 2024.
- Zheng Zhu et al., “Is Sora a World Simulator? A Comprehensive Survey on General World Models and Beyond”, arXiv:2405.03520.
- “How Far Is Video Generation from World Model: A Physical Law Perspective”, ICML 2025, arXiv:2411.02385.
- Danijar Hafner et al., “Training Agents Inside of Scalable World Models” (Dreamer 4), arXiv:2509.24527, 2025.
Next in this series: how does a model actually learn to “imagine”? A plain-language tour of the four technical routes — Dreamer-style imagination training, JEPA, video simulators, and 3D scene models.