What Is a World Model? The Field, the Money, and the Test | WhatAICanDo Skip to content

What Is a World Model? The Field, the Money, and the Test

Devin
Updated date:
7 min read

Core argument: A world model is a system that has learned how the world behaves well enough to predict it: show it a state and an action, and it tells you the next state. With that one ability come planning, simulation, and practice without consequences — the things today’s language models, which only predict text, cannot do. In 2026 the field drew an $8.2 billion acquisition, a record $1.03 billion seed round, and four competing technical routes. This page is the series hub: the definitions, the money, the test, and links to every chapter.

Welcome to the world models series — plain-language explainers on the technology trying to give AI a working picture of physical reality. Start here, or jump straight to any chapter below.

1. The short answer

A world model is a system that learned the reflex: given the current state of an environment — a snapshot of how everything is at one moment — and an action, predict the state that follows. Train that reflex on enough recordings of the world — video, robot logs, driving miles — and it becomes something more: a simulator the system can roll forward inside its own head.

Why does that matter? Because prediction enables three things text prediction never will:

That third one connects directly to the most visible flaw of today’s AI. Language models hallucinate partly because text is their only channel to reality — a point chapter 1 covers in detail. A model with a picture of the world has a second channel. That is the promise, and the industry is betting billions on it.

2. Why 2026 is the year it got expensive

Three money events frame the field.

In March, Yann LeCun — Turing Award winner and formerly Meta’s chief AI scientist — raised $1.03 billion for Advanced Machine Intelligence (AMI Labs), reported as the largest seed round in European history. His bet: world models, against the grain of an industry built on large language models (LLMs). In late May, NVIDIA released Cosmos 3, an open world foundation model for physical AI that it positions as the standard the whole ecosystem builds on. And in late September, AMD agreed to acquire Fei-Fei Li’s World Labs for roughly $8.2 billion in stock — a chipmaker buying a modeling team because physical AI needs simulation environments.

Around those anchors sits a wider ring: General Intuition raising at a $2.3 billion valuation in June, and Decart, Odyssey, and Emulate — the last founded by the researchers behind Genie 3 — each commanding billion-dollar valuations on the simulation story. Chapter 4 maps all of them, with the numbers and sources.

The pattern underneath the money: models reproduce quickly once published, but proprietary data (gameplay archives, driving miles, robot logs) and compute do not. Watch who owns those two.

3. Four ways to build a world model

Researchers disagree about what a world model should actually predict, and the disagreement defines four technical routes. Each has a working system behind it today.

Imagination training — compress what you see into small internal codes, predict how they change, and train agents entirely inside the model. Dreamer 4 (September 2025) is the clearest example: an agent that learned to reach the Minecraft diamonds benchmark without playing the real game first.

Latent prediction — predict meaning instead of pixels. Meta’s V-JEPA 2, trained on over a million hours of video, handled pick-and-place of new objects in environments it had never seen with success rates between 65 and 80 percent — and the action-conditioned version got there with 62 hours of robot data.

Video simulators — generate the frames themselves, interactively. Google DeepMind’s Genie 3 produces navigable environments in real time at 720p and 24 frames per second, with spatial memory: walk away, walk back, and the world is still there.

3D scene models — rebuild the stage rather than the view. World Labs’ Marble turns text or images into persistent, exportable 3D worlds; its successor Atlas folds generation, reconstruction, and simulation into a single model.

Chapter 3 explains each route’s mechanics, trade-offs, and where they are converging.

4. The honest test

“World model” is doing heavy marketing duty right now, so the series applies one yardstick everywhere. Three questions:

  1. Can you act in it — does an input change what happens next?
  2. Does it stay consistent when you look back?
  3. Can it answer a what-if question about a situation you describe?

A system that fails two of the three is a video generator with good publicists. That test is why the Sora question (chapter 8) has no settled answer, and why open weights — publishing the model files for anyone to download — alone do not make a world model.

5. The series so far

Coming next: new chapters weekly.

6. Questions people actually ask

Is a world model the same as a game engine? A game engine’s rules are written by engineers; a world model’s rules are learned from data. Learning lets it capture places no designer built — and makes it capable of being confidently wrong. An engine’s rules can be read and checked line by line; checking a learned model’s rules is an open problem.

Do world models replace large language models? No. Language models carry the semantic layer — what things are called and what humans have written. World models carry the dynamics layer — what happens when things move. Systems that act in the real world need both.

Do LLMs have world models? Partly, and the field is split. LLMs clearly build internal representations that track some structure of the world — it is why they can reason about microwaves and melted ice. Whether that amounts to a predictive, actionable world model or only fragments of one is an open research fight. Chapter 2 covers the distinction.

Is Sora a world model? By the series’ three-question test: it generates frames, but you cannot act in it — so it fails question 1. OpenAI frames video generation as a path toward world simulators; LeCun calls such systems artistic video generators. Chapter 8 gives the question a full hearing.

Sources

  1. D. Hafner et al., “Training Agents Inside of Scalable World Models” (Dreamer 4), arXiv:2509.24527, September 2025.
  2. Meta AI, “Introducing the V-JEPA 2 world model”, June 11, 2025; arXiv:2506.09985.
  3. Google DeepMind, “Genie 3: A new frontier for world models”, August 2025.
  4. World Labs, “Marble: A Multimodal World Model”, November 2025; “Atlas: A World Model for Spatial Intelligence”, September 2026.
  5. TechCrunch, “AMD will acquire Fei-Fei Li’s World Labs for $8.2 billion”, September 28, 2026.
  6. TechCrunch, “Yann LeCun’s AMI Labs raises $1.03 billion to build world models”, March 9, 2026.
  7. TechCrunch, “General Intuition’s $2.3B bet that video games can train AI agents for the real world”, June 25, 2026.
  8. NVIDIA Newsroom, “NVIDIA launches Cosmos 3, the open frontier foundation model for physical AI”, May 31, 2026.
  9. D. Ha and J. Schmidhuber, “World Models”, 2018.
  10. Y. LeCun, “A Path Towards Autonomous Machine Intelligence”, 2022.

This hub updates as the series grows. Last updated: October 21, 2026. Newest chapter: do world models hallucinate.

Previous
Future of AI in the Next 10 Years: Five Forces, Not Dates
Next
When Will AGI Happen? Every Major AGI Prediction, on the Record