A talk on world models

AI World Models

Can a machine imagine the future?

predict · simulate · choose

Opening question

Are you reacting to the present — or acting on a predicted future?

Catching a ball, crossing a street, pouring water, playing chess — intelligent action happens before the outcome arrives.

Can a machine do something similar?

Roadmap

Where we're going

01What is a world model?definition · action-conditioning · the litmus test
02Memory & representationpartial observability · latent state · learning from experience
03From prediction to planninguncertainty · rollout · imagine → evaluate → act
04Where it fits & what countspolicy vs simulator vs agent · are LLMs & video models world models?
05Why it's still hardfailure modes · long-horizon drift · evaluation
06Testing it: Voyagea counterfactual social testbed
01
01
Physics simulator versus a learned world model

A working definition

A world model is an internal predictive model of how an environment changes — especially in response to an agent's actions.

In plain words: a model of “how the world will change if I do this.”
Not a crystal ball that reveals the one true future — a sandbox you can act inside to imagine many.

The minimum idea

Current state + ActionPossible next state

(zt, at) → ẑt+1

Click an action and watch how the same state + a different action produces a different result.

Prediction

What happens next?

NOWNEXT

Forecasts one timeline, projected from the past.

World-model reasoning

What happens if I do this?

NOW
A → ✓B → ?C → ✕

The litmus test: if the future doesn't change when the action changes, it isn't a world model.

02
02

Partial observability

Out of sight isn't gone.

What the camera sees isn't the whole state of the world. A world model must keep an internal picture through occlusion and missing information.

Like a baby learning that a hidden toy still exists, the agent must remember what it cannot currently see.
Observation compressed into a latent world representation
Observation ot
Latent state · the gist zt

Learning from experience

Observe → Act → See result → Learn

(ot, at, ot+1)

One piece of experience = what I saw · what I did · what happened next.

Baby version: try an action, get surprised, update the model.

1Observeball at the centre of the table
2Act + predictpush it toward the edge
3See surprisethe ball falls off
4Update modellearn edges and falling
03
03

Uncertainty

The future is usually not one answer, but a set of possibilities.

P(zt+1 | zt, at)

In plain words: how likely each possible next moment is — given where things are and what you do.

A pedestrian at a corner might turn left, turn right, or stop. A good model predicts — and owns its uncertainty.

Rolling the model forward

Imagine further by running the model on its own predictions.

ztnow · ball at the centre
t+1rolling toward the goal
t+2likely in the goal
t+3maybe overshooting?
t+4honestly, guessing

Like replaying a chess game in your head: a few moves ahead is sharp — twenty moves ahead, you're mostly guessing. Each step leans on the last, so small errors compound.

Observe, imagine, evaluate, act — the planning loop

Live interaction

Same state.
Which action wins?

Goal: land the blue ball in the green zone without it falling off. Predict first, then reveal.

Let the room predict or vote, then click an action — or “Simulate all three futures”.

Receding horizon

Plan several steps.
Execute one.
Observe again.
Replan.

The agent need not commit to the whole distant future at once. It keeps updating its own plan.

End-to-end example

Let's watch one robot think it through.

1Observecameras and sensors
2Representbuild a latent state
3Imaginesimulate several actions
4Evaluatesuccess, risk, cost
5Actexecute one step, then feedback
The world model is not the whole agent — it is the part the agent uses to predict consequences.
04
04

Environment / Simulator

decides what really happens

observation ↑↓ action

Agent

Perception
Memory
World Model
Planner
Policy
Goal / Value
World model predicts consequences.Policy chooses actions.Simulator determines what actually happens.

Concept boundaries

Generating a plausible future is not the same as having a reliable world model.

SystemMainly predictsAction-conditionedNaturally supports planning
LLMthe next token / languagenot necessarilynot necessarily
Video modelfuture visual framesnot necessarilynot necessarily
Physics simulatorstate changesusuallyyes
Control world modelconsequences of actionscore abilitymain goal
World model applications across robotics, driving, games and science
Robotics · Self-driving · Simulation · Science — real systems today
Demo · NVIDIA Cosmos

State of the art · 2025–26

This isn't sci-fi any more.

Google DeepMind

Genie 3

Type a sentence → a 3D world you can walk around in, in real time.

OpenAI

Sora 2

Generates video that mostly obeys real physics — "video as a world simulator."

Meta · Yann LeCun

V-JEPA 2

Learns the world from raw video, then drives a real robot arm to pick unseen objects.

NVIDIA

Cosmos

"World foundation models" that simulate the physical world to train robots & cars.

Wayve

GAIA

Imagines rare, dangerous driving scenes so self-driving cars can practice safely.

Scan any code, or click a card on the presenter laptop to open the demo.

See it move

Watch a world model dream up a place.

Genie 3 builds a world from a prompt and keeps it consistent as you move through it — exactly the "remember what you can't currently see" idea, at scale.

Demo · Google DeepMind — Genie 3

Drop a clip at assets/demo.mp4 to play it right here —
or scan the code to watch the real Genie 3 demo.

Your turn

Now you play a world model.

Decart · free

Oasis

Play AI-Minecraft — every frame is dreamed up live, with no game engine.

Microsoft Muse · free

AI Quake II

Walk through a shooter that no engine ever rendered.

Odyssey

Explorer

Step into a streamed, interactive 3D world you can move around in.

World Labs

Marble

Turn a sentence or a photo into a 3D world you can keep and explore.

Best on a laptop (Chrome) · click any card to open one live on stage.

05
05
Failure modes of learned world models
Broken physicsObject inconsistencyCounting failureCausal failure

Model drift

Each step is only a little wrong — far ahead, it can leave reality entirely.

Short-term prediction can be good while a long rollout drifts off — which is why your GPS keeps recalculating. Better to glance at reality often than trust one long guess.

Evaluation

What makes a world model actually good?

01

One-step accuracy

Is the next-step prediction correct?

02

Multi-step consistency

Is a long rollout stable?

03

Counterfactual accuracy

When the action changes, does the prediction change sensibly?

04

Planning improvement

With the model, do real actions get better?

06
06

Our project

Voyage isn't a world model — but it can test whether an agent has a useful one.

Voyage Environment
decides real outcomes
AI Agent
chooses actions
contains?
Internal World Model
predicts consequences

Counterfactual test

NPC trust: 72

Pick an action — compare the agent's prediction with the simulator's actual result.
A crystal ball versus a sandbox where a robot explores possible futures

A world model does not give AI a crystal ball.

It gives AI a sandbox.

Represent the present.Simulate possible futures.Choose an action.

Thank you

Questions?

Represent the present.Simulate possible futures.Choose an action.

AI World Models — does AI get a crystal ball, or a sandbox?

1 / 34