World Models Are Not Game Engines Yet

Google DeepMind's Genie 3 makes a sentence that sounded strange a year ago feel much more concrete: a generative model can produce an interactive 3D world in real time.

Given a text prompt, Genie 3 can generate a 720p environment that runs at 24 frames per second and remains navigable for a few minutes. The user can move through the generated scene, and the model updates the world in response to those actions. DeepMind can even trigger "promptable world events" such as changing weather or introducing objects and characters.

It is tempting to look at this and say that generative models are becoming game engines.

I think that is too early. But I do think we are seeing the beginning of a new simulation layer that may eventually sit beside, or inside, game engines.

A game engine stores a world. A world model predicts one.

The distinction starts with state.

A conventional game engine has an explicit representation of the world. An object has a transform. A character has health. A door is open or closed. An inventory contains a defined set of items. Physics operates on bodies with known properties. Gameplay code decides what transitions are legal.

The rendered frame is a consequence of that underlying state.

A generative world model approaches the problem from another direction. Genie 3 receives the previous trajectory and user actions, then generates what should plausibly happen next. DeepMind describes world models as systems that simulate how an environment evolves and how actions affect it. Genie 3 performs that prediction quickly enough to support real-time interaction.

Visually, those approaches can begin to converge. Internally, they are very different.

If I walk behind a building in Unreal or Unity and return ten minutes later, the engine does not need to remember what the building looked like from a sequence of pixels. The building still exists as structured state. A world model has to maintain enough information about its generated history to recreate a consistent continuation.

Genie 3 has improved this temporal consistency enough to keep environments coherent for a few minutes. That is an impressive step. DeepMind still lists interaction duration as a limitation, along with constrained action space and difficulty simulating interactions among multiple independent agents.

Those limitations map almost exactly to the things game engines are built to control.

The interesting part is not graphics generation

I do not think Genie 3 matters because it can generate pretty environments. We already have strong image and video generators.

The important change is action conditioning. A world model has to produce a sequence that changes in response to what the user does. If I turn left instead of right, the future has to branch; if I return to an earlier location, the environment has to remain believable.

This direction was visible in Genie 2. DeepMind's 2024 model could generate playable 3D environments from an image, respond to keyboard and mouse input, model interactions and maintain parts of a world that temporarily left the camera view. Its high-quality trajectories lasted up to roughly a minute.

GameNGen pushed on the same boundary from a narrower domain. The 2024 research system learned to simulate DOOM at more than 20 frames per second using a diffusion model conditioned on previous frames and player actions. Instead of running the original game renderer to produce every next frame, the model learned to predict the visual result of gameplay.

Genie 3 expands the idea from reproducing one known game toward generating many environments from prompts.

That is why I see world models as simulation research, not just generative media.

Where a world model still differs from a production engine

For actual game development, believable next frames are not enough.

I need authoring, persistent state, inspectable gameplay logic, save files, networking rules, collision, animation and scripts that behave predictably. Most importantly, I need to inspect the state when something goes wrong.

If a generated character fails to pick up an object, there is no way to recover the cause. The object may never have been represented as interactable, the model may have forgotten it existed, the action itself may have failed, or the visual generation may have produced a plausible frame that merely looks like failure.

In a conventional engine I can inspect the hierarchy, variables, collision volumes, animation graph and logs. The world is made from state I can query.

A generative world model can hide much of its simulation inside learned representations. That flexibility is what makes open-ended generation possible, but it makes production control harder.

For games, controllability is not a secondary concern. It is the product.

The first useful role may be before runtime

This is where I think world models could enter game production sooner than many people expect.

Not as replacements for Unreal, Unity or custom engines, but as machines for generating experience.

DeepMind's main framing for Genie 3 is agent training. A model can generate many environments, and an embodied agent can act inside them. DeepMind tested a recent SIMA agent inside Genie 3 worlds, using the generated environments as places where the agent could pursue goals through navigation actions.

SIMA itself is useful context. DeepMind's earlier system learned to follow natural-language instructions across several existing 3D games using only screen images plus keyboard and mouse controls, without access to source code or custom game APIs.

Put the two ideas together and an interesting pipeline appears: one model creates environments; another agent learns or is evaluated inside them.

That is different from shipping a game to players, but it can be valuable for prototyping, testing and training.

A designer could explore a rough world before building assets. AI characters could encounter generated situations that would be expensive to author manually. Gameplay ideas could be tested for mood, navigation or spatial composition before committing to a production scene.

For someone working with Blender or game engines, I would treat this like procedural generation at a higher semantic level: traditional systems generate from rules we write; a world model generates from learned priors. The trade is more flexibility for less explicit control.

I do not expect engines and world models to stay separate

My guess is that the long-term architecture becomes hybrid.

A game engine is very good at explicit state. A world model is becoming good at filling in what does not need to be explicitly authored.

That suggests a boundary similar to what is happening in rendering.

The engine could remain authoritative for gameplay state, geometry that matters for interaction, player position, networking, inventory, physics constraints and persistence. A learned world model could handle parts of environment generation, background behaviour, visual variation, animation, secondary interactions or content that does not need exact scripting.

The model should not decide whether the player owns a key if the game rules depend on it. It may be reasonable for the model to decide what clutter sits around the room or how a background pedestrian reacts to rain.

The engineering challenge becomes deciding which parts of the world must be true and which parts only need to be plausible.

My prediction

I do not think game engines disappear because world models can generate interactive video.

I think game engines gradually absorb world models as another simulation primitive.

Today an engine combines rendering, physics, animation, scripting, audio and networking. In the future, learned world simulation may become another layer in that stack. Developers could choose where deterministic simulation ends and generative simulation begins.

The boundary will move by project. Competitive games need tight authority and reproducibility; cinematic exploration may tolerate more generated state; agent-training environments may value variety over handcrafted content.

Genie 3 is still a research preview. Its worlds last minutes, its action space is constrained, and multi-agent interaction remains difficult. DeepMind presents it as a step toward world simulation and agent research, not a replacement for existing engines.

But the direction is becoming clearer.

A game engine computes a world from rules and explicit state.

A world model learns enough about worlds to predict one as you move through it.

Right now those are different technologies.

I would not assume they stay that way.

Sources

1. Google DeepMind, Genie 3: A new frontier for world models, 5 August 2025.
2. Google DeepMind, Genie 2: A large-scale foundation world model, 4 December 2024.
3. Valevski et al., Diffusion Models Are Real-Time Game Engines (GameNGen), August 2024.
4. Google DeepMind, A generalist AI agent for 3D virtual environments (SIMA), 13 March 2024.