What is a world model in AI? Genie 4, Genie 3 and the 2018 paper that started the idea

A world model is software that has learned how a place behaves. Show it a scene and an action, and it predicts what happens next. The idea is old, and it is suddenly everywhere because Google DeepMind introduced Genie 4 on October 27, 2026. This guide explains what a world model is, what Genie 4 does and does not show, and how to read the next demo.

What is a world model in AI?

The clearest early definition comes from a 2018 paper, "World Models," by David Ha and Jurgen Schmidhuber. The authors built generative neural network models of reinforcement learning environments. Their model learned "a compressed spatial and temporal representation of the environment," and features from it fed a very compact, simple policy that could solve the task. The agent could even be trained entirely inside the model's generated "dream" and then transferred back to the real environment. The abstract on arXiv does not give a formal definition beyond that, and we use it here as the working one.

Put plainly, a world model does two jobs. It compresses what an environment looks like, and it predicts what changes when something acts on it. Once you can predict, you can plan, rehearse or play without touching the real thing.

Why people care now

Two things changed since 2018. Video generation became good enough to predict whole scenes, and large language models learned to reason about goals. DeepMind's Genie line combines them. Genie 3, announced August 5, 2025, generated navigable worlds at 720p and 24 frames per second. Its authors, Jack Parker-Holder and Shlomi Fruchter, wrote that "Genie 3 environments remain largely consistent for several minutes." They were equally direct about the ceiling: "The model can currently support a few minutes of continuous interaction, rather than extended hours."

That candor is the right lens for every world model demo. The headline question is not whether it looks real. It is how long it holds together.

What Genie 4 adds

DeepMind's October 27, 2026 post, by Raia Hadsell, Diego Rivas and Hang Qi, describes Genie 4 as a new world model for real-time interactive simulation that is natively multimodal, generating audio alongside visuals. It draws on the Gemini family: Gemini Omni anchors how things look and behave, and other Gemini models handle reasoning to keep scenes coherent.

Users can steer a world in real time by describing changes in natural language, such as the weather, the camera or new people and objects. The model simulates aspects of aerodynamics and collisions, and generates ambient sound and adaptive character dialogue as the scene runs. A Gemini 3.8 Flash model acts as a guiding agent that watches frames, breaks goals into steps, checks progress and adjusts the plan. A robotics demo features an Apptronik Apollo humanoid playing checkers, and mentions Gemini 4 Argon. Our report on Gemini 4 Argon's robotics results covers how simulated data fed those numbers.

What is not shown

The post is a research preview, not a product. Access starts with trusted testers. DeepMind plans access for Google AI Ultra subscribers through Project Genie, plus early-access partners using the Gemini API and Google AI Studio, and a waitlist is open. The page does not give a resolution, frame rate, session length, price or release date beyond "coming soon."

The limitations section is where the honest content sits. DeepMind says visual details and object permanence can drift over longer sessions or after multiple edits, and that complex physical interactions, real-time audio accuracy, legible text, spatial accuracy and long-term consistency are still active research areas. Testers in creative work, journalism, education, gaming and engineering are using it for previsualization, turning archival photos into simulations, rapid game prototyping and scenario generation. Those are examples DeepMind reports, not independent evaluations.

How a world model differs from a video generator

A video generator makes a clip from a prompt and stops. A world model takes an action at each step and predicts the consequence, so you can keep steering. That is the property the 2018 paper used: a compact policy acted on features from the model, and the agent could be trained inside the model's own dream before being moved back to the real environment. Genie 4's real-time steering by natural language, and the Gemini 3.8 Flash guiding agent that watches frames and adjusts plans, are the modern versions of the same loop. The question that matters for AI progress is whether skills learned in the dream transfer to reality. DeepMind's post reports tester uses but no transfer results.

Safety and provenance

All generated video and audio carries SynthID watermarking, and DeepMind says Genie 4 has safety guardrails and internal safety evaluations and went first to trusted testers. A watermark helps a viewer tell a generated clip from a recording. It does not tell you whether a simulated physics result is right, which is the bigger issue if people start using such worlds to train robots or test plans.

A reading guide for the next demo

Use four questions. How long does the world stay consistent, in minutes and in edits? Does an agent learn something inside it that transfers to reality, which is the 2018 paper's test? Is the physics checked against measurement, or does it only look plausible? And is anyone outside the lab running it? Genie 3 answered the first question with "several minutes." Genie 4 has not yet published an answer.

Our take

World models are one of the more interesting ideas in AI, and also one of the easiest to oversell. A model that can dream a place is useful for rehearsal. It is not evidence of understanding the physical world, and DeepMind's own limitations list says as much. Watch for transfer results and independent tests, because those, not the clips, will show whether this is a research tool or a path to general capability.