Tech Briefing. Your daily dose of tech news — signal, not noise.
Learning Center
One technology, explained in plain English
DTBRIEF

World Models, Explained: How AI Imagines a Place You Can Walk Through

A world model does not play a video. It imagines the next frame, twenty-four times a second, based on what you just did. Here is how Genie 4 turns a sentence into a place you can walk through — and where it still falls apart.

A glowing 3D world unfolding from a single line of text, with a small character walking through a landscape that is still being drawn ahead of them, on a deep navy background with cyan and warm yellow accents
A world model draws the world as you move through it — the ground ahead of you does not exist until the model imagines it.
24 fps
Genie draws a brand-new frame 24 times a second — each one predicted from the frames and actions before it
200,000 hrs
Of gameplay video trained the first Genie to imagine how worlds behave, without anyone labeling a single action
~2 min
About how long a Genie 4 world stays consistent before details start to drift and change
720p
The resolution of Genie’s worlds — a playable sketch, not a finished game

1 What it is

A world model is an AI system that has learned how the world behaves — and can then imagine a world of its own, frame by frame, as you move through it. You type a sentence, a place appears, and when you press a key to walk forward, the model draws what you should see next. It is not playing back a video somebody recorded. It is inventing the world as you explore it.

Google’s Genie 4, which opened to Google AI Ultra subscribers on November 18, 2026, is the clearest example yet. It turns a text prompt or an image into a walkable, three-dimensional world rendered in real time at up to 24 frames per second, with synchronized sound that adapts as the scene unfolds — footsteps, weather, music, even spoken dialogue from characters. You can change the world mid-session by typing a new prompt while you move, and the world reshapes itself around you.

The idea is older than the hype. The first Genie, published by Google DeepMind in February 2024, could only make flat, two-dimensional platformer worlds at one frame per second, trained on 200,000 hours of public gameplay video. Genie 2, in December 2024, moved to three-dimensional worlds that lasted ten to twenty seconds. Genie 3, in August 2025, became the first to run in real time at 720p and 24 frames per second. Genie 4 adds sound and live editing — and each step took the same core trick further.

2 How it works: the four-step loop

Every world model runs the same loop, and Genie 4 runs it twenty-four times a second:

1. Imagine — you write a prompt, and the model generates the opening frames of the environment. 2. Act — you move with the keyboard, and the model reads your action. 3. Predict — the model draws the next frame, guessing what the world should look like after your move. 4. Remember — it checks its recent output, about a minute back, so the world stays consistent.

Underneath the loop is an autoregressive transformer — the same kind of model that writes text one word at a time, except here “what comes next” is an entire video frame. The model learned how worlds behave by watching enormous amounts of video: the first Genie inferred, entirely on its own, eight basic action patterns from adjacent frames of gameplay footage, with nobody labeling what “jump” or “move left” meant. Gemini models now run inside the simulation loop too, keeping the physics and the story consistent and adapting your prompts to the running scene.

Crucially, Genie does not build an explicit three-dimensional model of the scene the way a game engine does. There is no stored map, no pre-rendered geometry. The world exists only as the model’s memory of what it has shown so far, plus its prediction of what comes next — which is exactly why the world can drift.

3 World model vs video generator

A video generator makes a clip that plays from start to finish. You prompt it, it renders a few seconds of footage, and you watch. You cannot steer it, walk into it, or change your mind halfway through — the clip is already finished. It is a movie.

A world model makes the next frame based on what you just did, so the world reacts to each move. There is no finished clip, only a loop: action in, frame out, twenty-four times a second. It is a place, not a movie — and the place only exists while you are in it.

Video generatorRenders a finished clip from a prompt. Beautiful, but fixed: you watch it, you cannot steer it, and it ends after a few seconds.
World modelPredicts the next frame from your last action, 24 times a second. You can move, explore, and rewrite the scene mid-session — but the world can drift.
A video generator shows you a world. A world model lets you walk around inside one that is being imagined as you go.

4 Why everyone’s building it now

Three stories in today’s briefing are the same story. Google opened Genie 4 to AI Ultra subscribers in a rebuilt Project Genie on November 18, with a community gallery for discovering and remixing worlds. But the more important customer is not sitting at a keyboard. Waymo has already built a variant called the Waymo World Model from Genie 3, to simulate the rare, dangerous edge cases its robotaxis almost never meet on real roads — the crashes and near-misses you cannot safely stage in real life.

A robot arm practicing picking up objects inside a glowing simulated warehouse that looks like a video game world, on a deep navy background with cyan and warm yellow accents
The real customer for world models is not gamers — it is robots and self-driving cars that need infinite worlds to practice in.

That is the real market. Robots and self-driving cars learn by doing, and the real world is a slow, expensive, sometimes dangerous classroom. A world model is an infinite practice ground: every possible intersection, every weather condition, every strange object on the road, generated on demand. The same logic applies to the robot brains in yesterday’s lesson — vision-language-action models need worlds to practice in, and world models can supply them by the thousand.

And the timing is not an accident. The models got fast enough to run in real time, the sound arrived, and the editing became live, all in the same year that AI agents started doing real work. A world you can walk through, talk to, and rewrite in plain language is the natural interface for an agent that has to act in the world, not just describe it.

5 The honest part

First, the world forgets. A Genie 4 scene holds together for roughly two minutes before details start to drift — an object can come back changed after you look away, and readable text inside scenes is still unsolved. Google is open about the remaining gaps: camera latency, physical interactions, audio precision, and long-term scene permanence are all still listed as open challenges.

Second, it is a sketch, not a simulation. Genie’s worlds look right, but the physics is learned from video, not computed from laws — so a world can look convincing and still get the physics wrong. That is fine for a game or a brainstorm, and dangerous if you mistake it for a faithful copy of reality. Google still calls Project Genie a research prototype, and there is no public API or developer kit.

The honest summary: a world model is the cheapest way ever invented to imagine a place — and the most expensive mistake you can make is believing the place it imagined is real. The technology is real, the worlds are playable, and the two-minute memory is the number that tells you how far it still has to go.

Key terms

  • World model An AI system that predicts how an environment changes in response to actions, then draws that change frame by frame — a place you can move through, not a clip you watch.
  • Autoregressive Predicting what comes next from everything that came before. In a text model the next thing is a word; in a world model it is an entire video frame.
  • Latent action The hidden action patterns a model discovers by watching video — the first Genie found eight of them on its own, with no labels.
  • Promptable world event Changing a generated world with a text prompt while you are inside it, such as new weather or a new character.
  • Sim-to-real Training an AI in simulation and then using it in the real world — the reason robots and self-driving cars need world models.
  • Project Genie Google’s web app for exploring world models, where Genie 4 now runs for eligible Google AI Ultra subscribers.
From today’s briefing

This lesson comes from today’s Daily Tech Briefing (Episode 38, November 19, 2026): Anthropic's record IPO, Nvidia's slide & GTA 6's 99. Read the full briefing: Anthropic's $83 IPO — Daily Tech Briefing.

Sources

Google, “Genie 4 is now available” (Google Blog, November 18, 2026) — Project Genie rollout to AI Ultra subscribers, mid-session edits, native audio; Google DeepMind, Genie 4 announcement (October 27, 2026) — research preview, real-time interactive simulation; Wikipedia, “Genie (world model)” — Genie 1 (February 2024, 2D, 1 fps), Genie 2 (December 2024, 3D, 10–20s), Genie 3 (August 2025, 720p, 24 fps), Waymo World Model; AILoitte, “Genie 3 by DeepMind: Features, Access and Use Cases” (2026) — autoregressive frame prediction, one-minute memory, promptable world events.