Predicts the next token
Strong at language, code, reasoning and tool use, but it does not automatically hold a verifiable physical state.
Typical problem: sounding plausible is not the same as the world working that way.
Bilingual atlas · source-first research · updated 2026-07-08
World Models
Large language models learn to continue text. World models aim at something different: simulating the outside world internally and predicting what happens after an action.
Core Definition
A world model is not just a better chatbot. It is a model that can run the world forward inside itself.
A useful world model compresses observations into an internal state, then predicts future states from actions, time or conditions. It can be an explicit simulator, or a predictive module inside a video, robotics or autonomous-driving system.
The test is not only whether the image looks real. A serious world model should preserve state, accept interventions, support action planning and avoid collapsing during longer rollouts.
Do Not Merge These Ideas
The key difference between LLMs, video models and world models
Generates a visual sequence
Can produce coherent motion and may learn partial physics, but many systems still optimize for looking right.
Typical problem: beautiful output is not the same as controllable, interactive agent training.Predicts the world after action
The focus is state, causality, action and long-horizon consistency, so AI can learn, test and plan inside simulation.
Core value: try the future internally before paying the cost of real-world failure.Editorial Method
Source-first, cautious by default.
World model is an overloaded term. This site separates official claims, external evidence and editorial judgment so readers can see how strong each statement is.
Primary sources first
Papers, technical reports and official posts are used as the verification base. Commentary is useful for discovery, not proof.
Evidence labels
Cards mark whether a claim is official positioning, a paper/report claim, a product surface or this site's synthesis.
Capability boundaries
A polished demo is not treated as proof of long-horizon physics, controllability or downstream agent utility.
Review rhythm
Fast-moving claims are marked with a review date and should be rechecked before becoming evergreen explainers.
Recently reviewed
2026-07-08
Positioning moved to a bilingual, source-first research and product intelligence site; team cards now expose source and evidence status.
Trajectory
From acting in imagination to real-time interactive worlds
-
World Models
Ha and Schmidhuber pushed the idea of imagining future rollouts in compressed latent space into broader AI discussion.
-
Dreamer / model-based RL
DeepMind's Dreamer line showed that agents can train behavior policies inside learned latent dynamics.
-
Sora and Genie
Video generation entered the world-simulator narrative, while Genie turned video environments toward interactive agent training grounds.
-
Cosmos, V-JEPA 2, Genie 3, Marble, GAIA-2
The field split into physical AI platforms, representation learning, real-time worlds, 3D spatial generation and driving-specific world models.
-
Platformization and applications
Public materials show world models moving from research concepts into robotics, autonomous driving, 3D content and safety-testing workflows.
Team Map
Who is building world models, and what are they betting on?
This map summarizes public materials without treating marketing language as proof. Each card separates focus, maturity and caution.
Essential Reading
Read these first to understand the main line
The homepage keeps only sources that help build judgment: papers, technical reports and official posts. More references live in the source list.
First Explainers
Shareable pages for readers who want one clear answer.
The homepage is the map. Articles are the units that can travel: one question, one cautious answer, bilingual by default.
Explainer
What is a world model?
A non-specialist introduction to prediction, action and internal simulation.
BoundaryIs Sora a world model?
Why video generation can point toward world simulation without proving full interactivity.
ComparisonGenie, Cosmos and V-JEPA: three routes
Interactive worlds, physical AI infrastructure and representation learning are not the same bet.
Current Judgment
World models are unfinished, but they have already changed the direction of AI competition.
The frontier has shifted from "who can generate realistic video" toward "who can generate controllable, interactive and verifiable worlds." That matters for robotics, autonomous driving, industrial simulation and 3D content, because the expensive part is real-world trial and error.
The caution remains: many systems show partial ability, not long-horizon physical understanding. A valuable world model must be usable by downstream agents, not just impressive as a demo clip.
What to watch next
- Can interaction extend from minutes to longer sessions while preserving spatial and physical consistency?
- Can robots and autonomous-driving systems use these models reliably for training, validation and planning?
- Will evaluation move beyond image quality toward action consequences, causality and controllability?
- Will the cost gap between open models, APIs and commercial products shrink quickly?
FAQ
How should a non-specialist understand this?
Will world models replace large language models?
Not in the short term. The likely form is a combination: LLMs handle language, planning and tool orchestration, while world models handle physical state prediction and simulated environments.
Why do autonomous driving and robotics need them so much?
Because real-world trial and error is costly and risky. A reliable world model could generate rare scenarios, predict action outcomes and stress-test systems before deployment.
Is every video model automatically a world model?
No. A video model may learn object motion and scene continuity, but world models also require state persistence, interaction, intervention and usefulness for action planning.
How will this site be updated?
The first version is a manually curated static site. The next step is to split teams, papers and explainers into structured content, add bilingual article pages, then test subscriptions or sponsorship only after there is repeat interest.
Follow / Feedback
Help shape the first useful update loop.
There is no account system yet. For now, use GitHub Issues to request an update digest, suggest a source, or point out a weak claim that should be rechecked.