# World model

> Source: https://aiwiki.ai/wiki/world_model
> Updated: 2026-09-16
> Fact-checked: 2026-09-16
> Categories: Artificial Intelligence, Machine Learning
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "World model." aiwiki.ai, 16 Sept 2026. https://aiwiki.ai/wiki/world_model
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

A world model is an [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) system that learns an internal representation of how an environment works, enabling it to predict future states, simulate the consequences of actions, and support planning without needing to interact with the real environment at every step. In the simplest terms, a world model answers the question: "If I take this action in this situation, what will happen next?" The concept is inspired by the cognitive science idea that humans carry mental models of the world in their heads, constantly running internal simulations to anticipate outcomes before acting. The term was popularized for modern AI by a 2018 paper from [David Ha](https://aiwiki.ai/wiki/david_ha) and Jurgen Schmidhuber, and by the mid-2020s world models had become the organizing thesis behind systems such as DeepMind's Dreamer and Genie, Meta's [V-JEPA](https://aiwiki.ai/wiki/v_jepa), OpenAI's Sora, NVIDIA's Cosmos, and Fei-Fei Li's [World Labs](https://aiwiki.ai/wiki/world_labs) [1][3].

World models have become one of the most actively discussed topics in AI research as of 2025-2026. Their appeal is straightforward: an agent that understands the dynamics of its environment can plan ahead, reason about cause and effect, and generalize to new situations far more efficiently than one that relies purely on trial and error. The concept spans multiple subfields, from [model-based reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) (where a learned dynamics model reduces the need for real-world interaction) to video prediction systems (where models forecast future video frames) to interactive world generators (where users can explore AI-generated environments in real time). The term has also become entangled with the debate over whether large video generation models like [Sora](https://aiwiki.ai/wiki/sora) constitute genuine world models or are simply sophisticated pattern matchers [1].

## What is a world model in AI?

A world model is a learned, predictive model of an environment's dynamics: given a current state (and optionally an action), it predicts the next state. Unlike a fixed physics engine, a world model is learned from data, usually in a compressed latent space rather than over raw pixels. Three properties distinguish a world model from a plain generative video model: (1) it represents state, often as a compact latent vector; (2) it predicts forward in time; and (3) in its strongest form it is action-conditioned, so an agent can ask counterfactual "what if I do X" questions and use the answers to plan. Systems that only generate a single plausible video from a prompt, without responding to actions, sit at the weak end of this definition.

## History and foundational work

### Early ideas

The notion that intelligent agents should build internal models of their environments is not new. In [control theory](https://aiwiki.ai/wiki/control_theory), [model predictive control](https://aiwiki.ai/wiki/model_predictive_control) (MPC) has used explicit dynamical models for planning since the 1960s. In AI, the idea of a "mental model" for planning dates at least to Kenneth Craik's 1943 book *The Nature of Explanation*, which argued that organisms carry "small-scale models" of the external world in their heads and use them to try out alternatives before acting. Ha and Schmidhuber framed their own work with a related line from the systems theorist Jay Wright Forrester: "The image of the world around us, which we carry in our head, is just a model. Nobody in his head imagines all the world, government or country. He has only selected concepts, and relationships between them, and uses those to represent the real system" [3].

In [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning), the distinction between model-free and model-based methods has been central for decades. Model-free methods (like [Q-learning](https://aiwiki.ai/wiki/q-learning) and policy gradient algorithms) learn to act directly from experience without building an explicit model of environment dynamics. Model-based methods learn a transition model (predicting the next state given the current state and action) and use it to plan or generate synthetic training data. Model-based approaches tend to be more sample-efficient because the agent can "imagine" many trajectories without actually taking them, but they are only as good as the accuracy of the learned model [2].

### Ha and Schmidhuber (2018): "World Models"

The paper that brought the term "world model" into common usage in the [deep learning](https://aiwiki.ai/wiki/deep_learning) community was "World Models" by David Ha and Jurgen Schmidhuber, published in 2018. The paper proposed a three-component architecture for RL agents:

| Component | Architecture | Function |
|-----------|-------------|----------|
| Vision model (V) | [Variational autoencoder](https://aiwiki.ai/wiki/autoencoder) (VAE) | Compresses raw observations (images) into a compact latent representation (32 dimensions for CarRacing, 64 for VizDoom) |
| Memory model (M) | Recurrent neural network (MDN-RNN) | Predicts future latent states; captures temporal dynamics |
| Controller (C) | Small linear model | Maps the current latent state and memory to actions (just 867 parameters for CarRacing, 1,088 for VizDoom) |

The key insight was that the controller could be trained entirely inside the "dream" of the world model: the VAE and RNN together formed a [generative model](https://aiwiki.ai/wiki/generative_model) of the environment, and the controller learned a policy by interacting with this internal simulation rather than the real environment. As the authors put it, "We can even train our agent entirely inside of its own dream environment generated by its world model, and transfer this policy back into the actual environment" [3]. Ha and Schmidhuber demonstrated this on the VizDoom: Take Cover task and a car racing task. On CarRacing-v0, the agent reached 906 +/- 21, the first published method to solve the task (a score above 900 averaged over 100 trials); trained purely inside its dream and transferred back, the VizDoom agent scored 1092 +/- 556, well above the threshold of 750 considered to solve the environment [3].

The paper also explored a provocative idea: what happens when the agent trains purely in its learned model without ever interacting with the real environment? They showed this was possible but highlighted a limitation. Since the agent could exploit inaccuracies in the learned model (finding "cheats" that work in the dream but not in reality), pure dream training required careful regularization.

Ha and Schmidhuber's work proved that accurate latent dynamics models were sufficient for control, but also exposed the cost of modularity: the three components were trained separately, so the system could not fine-tune representations end-to-end. This limitation motivated the next generation of world model research [3].

### Dreamer series (2019-2023)

Danijar Hafner and collaborators at Google and the University of Toronto developed the Dreamer series of world model agents, which addressed the limitations of the Ha-Schmidhuber approach by training all components jointly and using more sophisticated policy optimization.

| Version | Year | Key advance | Notable result |
|---------|------|-------------|----------------|
| Dreamer (PlaNet) | 2019 | Recurrent State-Space Model (RSSM) for latent dynamics; end-to-end differentiable | Learned control from pixels in DeepMind Control Suite tasks |
| DreamerV2 | 2020 | Discrete latent representations; KL balancing | First world model agent to achieve human-level performance on the Atari 100K benchmark |
| DreamerV3 | 2023 | Normalization and transformation techniques for stable learning across domains | First algorithm to collect diamonds in Minecraft from scratch without human data; outperformed MuZero with far less compute |

Dreamer works by learning a world model that predicts future latent states from current states and actions. An actor-critic policy is then trained by "imagining" trajectories inside this model, with rewards predicted by the model itself. Because all components are differentiable, gradients flow from the imagined rewards back through the dynamics model and into the policy, enabling efficient end-to-end optimization [4].

DreamerV3 was particularly notable for its generality. A single, fixed set of hyperparameters worked across more than 150 tasks spanning continuous control, Atari games, procedurally generated environments, and the open world of Minecraft. Collecting diamonds in Minecraft is a long-horizon challenge that requires finding wood, crafting a pickaxe, mining stone, upgrading tools, locating iron, smelting it, and finally mining diamond ore, all from pixel observations with sparse rewards. No prior algorithm had accomplished this without human demonstrations or hand-crafted reward shaping. In the published runs, 24 of 40 random seeds collected at least one diamond within 100 million environment steps, with the first diamond appearing after about 29 million steps. DreamerV3's success was published in *Nature* in 2025, where the authors wrote that Dreamer is "the first algorithm to collect diamonds in Minecraft from scratch without human data" [4].

## What is JEPA and why does Yann LeCun favor it?

[Yann LeCun](https://aiwiki.ai/wiki/yann_lecun), Meta's former VP and Chief AI Scientist, has been one of the most vocal proponents of world models as the path to machine intelligence. LeCun has argued repeatedly that [large language models](https://aiwiki.ai/wiki/large_language_model) (LLMs), despite their impressive text generation abilities, will never achieve genuine understanding of the physical world because they operate only on discrete tokens and lack the ability to predict continuous, high-dimensional sensory states [5].

LeCun's proposed alternative is the Joint Embedding Predictive Architecture (JEPA). The core idea is that instead of predicting raw pixels or tokens (which is computationally expensive and forces the model to predict every irrelevant detail), a JEPA-based system predicts in a learned abstract representation space. Two encoder networks process inputs (for example, two different views of a scene, or a current frame and a future frame), and a predictor network learns to predict the representation of one from the other.

The approach has been implemented in a sequence of published models from Meta:

| Model | Year | Domain | Description |
|-------|------|--------|-------------|
| I-JEPA | 2023 | Images | Predicts abstract representations of masked image regions from surrounding context; no pixel-level reconstruction |
| V-JEPA | 2024 | Video | Predicts abstract representations of masked video segments; learns temporal dynamics without generating pixels |
| V-JEPA 2 | 2025 | Video + robot action | 1.2B-parameter world model pre-trained on 1M+ hours of video; an action-conditioned variant enables zero-shot robot planning |

I-JEPA (Image Joint Embedding Predictive Architecture) learns by masking large portions of an image and predicting the representation of the masked region from the visible context. Because it operates in representation space rather than pixel space, it can focus on semantic and structural information rather than low-level textures. V-JEPA extends this to video, learning to predict missing temporal segments in representation space, which forces the model to learn about motion, object permanence, and physical dynamics [6].

### V-JEPA 2 (June 2025)

Released on June 11, 2025, V-JEPA 2 is a 1.2-billion-parameter video world model pre-trained on more than 1 million hours of internet video plus 1 million images, with no action labels in the first stage. Meta described it as "a world model trained on video that enables state-of-the-art understanding and prediction, as well as zero-shot planning and robot control in new environments" [15]. A second, action-conditioned model, V-JEPA 2-AC, was post-trained on fewer than 62 hours of unlabeled robot video from the open DROID dataset, then deployed zero-shot on Franka robot arms in two different labs to pick and place objects by planning toward image goals. On benchmarks, V-JEPA 2 reported 77.3% top-1 accuracy on Something-Something v2 (motion understanding) and 39.7 recall-at-5 on Epic-Kitchens-100 (human action anticipation) [15]. V-JEPA 2 is the clearest demonstration to date of the JEPA thesis: a model that learns physical dynamics in latent space from passive video, then transfers to a robot with minimal action data.

LeCun views JEPA as a stepping stone toward what he calls Autonomous Machine Intelligence (AMI), a system architecture in which a world model sits at the center, surrounded by modules for perception, memory, cost estimation, and action. After leaving Meta, LeCun co-founded [AMI Labs](https://aiwiki.ai/wiki/ami_labs) (Advanced Machine Intelligence), a Paris-headquartered startup, to pursue this vision; in March 2026 it announced a $1.03 billion seed round at a $3.5 billion pre-money valuation, one of the largest bets yet on the thesis that the path to general intelligence runs through world models rather than next-token prediction [35].

## Types of world models

The term "world model" is used to describe several related but distinct approaches:

### Model-based RL world models

These are dynamics models learned within a reinforcement learning framework. The model takes a state and action as input and predicts the next state (and often the reward). The agent uses this model to plan by simulating future trajectories internally. Examples include the Dreamer series, [MuZero](https://aiwiki.ai/wiki/muzero) (which learns a latent dynamics model for board games and Atari), and various model-based approaches used in robotics.

The strengths of this approach are sample efficiency (fewer real-world interactions needed) and the ability to plan ahead. The weakness is that errors in the model compound over long horizons: a small prediction error at each step can accumulate into a wildly inaccurate trajectory after many steps, a problem known as model compounding error [2].

### Video prediction models

Video prediction models take a sequence of video frames (and sometimes conditioning signals like text or actions) and generate future frames. These models learn to predict how visual scenes evolve over time, capturing information about object motion, occlusion, and scene dynamics.

Models in this category include:

- **SVG (Stochastic Video Generation)** by Denton and Fergus (2018), which used variational methods to model uncertainty in future frames.
- **FitVid** and other deterministic video prediction models used for robotic planning.
- **[Sora](https://aiwiki.ai/wiki/sora)** by [OpenAI](https://aiwiki.ai/wiki/openai) (2024), a large-scale diffusion-based video generation model that produces high-fidelity video from text prompts.

Whether video prediction models are truly world models is a subject of ongoing debate (discussed in a later section).

### Learned simulators and interactive world models

These systems go beyond passive frame prediction by allowing interactive exploration. A user or agent can take actions within the generated world, and the model produces the next state in response, functioning like an AI-generated video game or simulator.

[Google DeepMind](https://aiwiki.ai/wiki/google_deepmind)'s Genie and Genie 2 are the most prominent examples. NVIDIA's Cosmos platform represents a commercial approach, providing world foundation models specifically designed for physical AI applications like robotics and autonomous driving. Odyssey's Odyssey-2 family and Starchild-1 (described in the Odyssey section below) are commercial real-time examples that stream generated video in response to typed text, with Odyssey-2 producing a new frame every 50 milliseconds and Starchild-1 adding synchronized audio [25][28].

### JEPA-style representation predictors

As described above, JEPA-based models predict in an abstract representation space rather than in pixel space. This avoids the computational burden and noise of pixel-level prediction while (in theory) focusing the model on the aspects of the world that matter for decision-making. With the release of V-JEPA 2 in 2025, this approach moved from pure research demos toward zero-shot robot control, though it remains earlier in deployment than commercial video-generation systems. I-JEPA, V-JEPA, and V-JEPA 2 are the main published examples.

## Video generation as world simulation

### Sora and the "world simulator" claim

When OpenAI introduced Sora in February 2024, its technical report described the model as a "world simulator," arguing that video generation models trained at sufficient scale would implicitly learn to simulate the physical world. Sora can generate photorealistic videos from text prompts, depicting complex scenes with moving objects, changing lighting, and plausible (if not always physically accurate) interactions [7].

The claim provoked significant debate. Supporters argued that a model capable of generating coherent video must have learned something about how the world works: objects fall when dropped, cars move along roads, water flows downhill. If the model can consistently generate physically plausible outcomes, it has, in some functional sense, learned physics.

Critics offered several counterarguments:

- **Memorization vs. understanding.** A 2025 study using the Physics-IQ benchmark (developed by INSAIT and Google DeepMind, comprising 396 real-world test videos) found that Sora produced the most visually realistic output, achieving the best multimodal-LLM realism score of 55.6%. But across all systems tested, the best physical-understanding score was only 24.1% of the maximum (achieved by VideoPoet, not Sora), indicating that even the most realistic-looking models fail to capture underlying physics. The study reported essentially no correlation between visual realism and physical understanding [8].
- **No action conditioning.** A true world model should respond to actions: "If I kick the ball, where does it go?" Sora generates video from text prompts but cannot be conditioned on a sequence of actions, so it cannot be used for interactive planning.
- **Out-of-distribution failure.** Research showed that video generation models rely heavily on memorizing patterns from training data rather than learning general physical principles. When presented with scenarios outside their training distribution (such as objects with unusual physical properties), the models failed to generalize, and scaling up data and model size did not improve this [8].
- **Correlation vs. causation.** Video models learn correlations in pixel patterns ("objects that look like this tend to move like that") rather than causal models ("gravity accelerates objects at 9.8 m/s squared"). This distinction matters for tasks that require counterfactual reasoning or precise prediction.

The Physics-IQ authors summarized the gap bluntly: "visual realism does not imply physical understanding" [8]. The debate is not merely academic. If video generation models are genuine world models, then scaling up video generation (as OpenAI, Google, and others are doing) is a path toward AI systems that understand the physical world. If they are not, the field needs fundamentally different architectures.

### The spectrum between frame prediction and world understanding

A useful way to think about the debate is as a spectrum rather than a binary:

| Level | Capability | Example |
|-------|------------|---------|
| Frame interpolation | Predicting the next frame given previous frames; no understanding of physics | Simple video codecs |
| Statistical video generation | Generating plausible video from text or context; learns correlations in visual patterns | Sora, Runway Gen-3 |
| Stylized physics | Understanding that dropped things fall and rolling things move, without precise equations | Current best world models |
| Approximate physical simulation | Predicting outcomes of interactions with reasonable accuracy; responds to action conditioning | Research frontier (Genie 3, advanced model-based RL) |
| Precise physical simulation | Accurate physics with correct equations of motion | Traditional physics engines (not learned) |

Current video generation models operate at the "statistical video generation" level, occasionally reaching "stylized physics." Current model-based RL world models and interactive systems like Genie operate closer to "stylized physics" or "approximate physical simulation" for restricted domains.

## Genie and Genie 2 (Google DeepMind)

Google DeepMind's Genie project represents one of the most ambitious efforts to build interactive world models.

### Genie (February 2024)

The original Genie, published in February 2024, is an 11-billion-parameter model trained on a filtered set of 30,000 hours of unlabeled internet gameplay video (curated from a much larger pool of public 2D platformer footage). It learned a latent action space (a set of abstract "controls") entirely from watching videos, without any labeled action data. Users could provide a single image (a photo, a sketch, or an AI-generated scene), and Genie would generate an interactive 2D environment that could be explored using the learned controls. DeepMind described Genie as "the first generative interactive environment trained in an unsupervised manner from unlabelled Internet videos" [9].

Genie was notable for several reasons. It demonstrated that interactive world models could be learned from passive video without action labels. It showed that a single model could generate diverse 2D platformer-style worlds. And it introduced the idea of "world generation" as distinct from "video generation": the output was not a pre-determined video but an environment that responded to user input.

### Genie 2 (December 2024)

Genie 2, announced in December 2024, extended the concept to 3D environments. From a single image and optional text description, Genie 2 generates an interactive 3D world that users can explore using a keyboard or mouse. The generated environments include object interactions (opening doors, bursting balloons), animated characters and NPCs, lighting and reflections, and basic physics simulation [10].

Technically, Genie 2 uses an autoregressive latent diffusion model that generates the world frame by frame, simulating the consequences of each user action. It maintains memory of parts of the scene that are not currently visible and renders them accurately when they come back into view. The model was trained on video data and does not use a traditional rendering engine [10].

DeepMind positioned Genie 2 as useful for training and evaluating [AI agents](https://aiwiki.ai/wiki/ai_agents): rather than building handcrafted simulation environments, researchers could generate an unlimited curriculum of novel worlds for agents to explore.

### Genie 3 (August 2025)

Genie 3, released in August 2025, was described by DeepMind as its "first world model to allow interaction in real-time." The post states that "given a text prompt, Genie 3 can generate dynamic worlds that you can navigate in real time at 24 frames per second, retaining consistency for a few minutes at a resolution of 720p" [11]. The system maintains visual memory of the environment "extending as far back as one minute ago," so scenes stay consistent as a user looks away and back. Genie 3 also introduced "promptable world events," letting a user change the running world on the fly, for example altering weather or introducing new objects and characters. DeepMind summarized it as "a general purpose world model that can generate an unprecedented diversity of interactive environments," positioning it as a tool for training and evaluating embodied agents [11].

## World Labs and spatial intelligence

World Labs, co-founded by [Fei-Fei Li](https://aiwiki.ai/wiki/fei_fei_li), pursues a distinct interpretation of world models centered on "spatial intelligence": building models that perceive, generate, reason about, and interact with the 3D world. The company emerged from stealth in September 2024 with $230 million in funding from backers including Andreessen Horowitz, NVIDIA's venture arm NVentures, and Radical Ventures, at a reported valuation of around $1 billion [16]. World Labs describes its products as "large world models."

In November 2025, World Labs launched its first commercial product, Marble, which generates persistent, downloadable 3D environments from a single image or text prompt [17]. Unlike systems such as Genie that synthesize frames on the fly as a user explores, Marble produces a fixed 3D scene up front, which reduces the "morphing" and drift seen in autoregressive video world models and lets users export the result as Gaussian splats, meshes, or video. World Labs thus represents a 3D-native, geometry-first approach to world modeling, contrasting with the video-prediction approaches of Sora and Genie and the latent-prediction approach of JEPA.

## NVIDIA Cosmos

NVIDIA launched the Cosmos platform at CES 2025 as a suite of world foundation models (WFMs) designed for physical AI development. Unlike research-oriented projects like Genie, Cosmos is aimed at commercial applications, particularly autonomous driving and [robotics](https://aiwiki.ai/wiki/ai_robotics).

Cosmos models generate physics-based videos from combinations of text, image, video, robot sensor data, and motion data. They are trained to handle physically based interactions, object permanence, and realistic rendering of industrial environments (warehouses, factories) and driving environments (roads, weather conditions, lighting variations) [12].

For autonomous vehicles, Cosmos integrates with NVIDIA's Omniverse simulation platform. Developers can use Cosmos Transfer to amplify variations of sensor data, turning thousands of real-world miles of driving data into billions of virtually driven miles. This data flywheel approach addresses one of the biggest bottlenecks in autonomous driving development: the need for vast amounts of diverse training data [12].

NVIDIA released Cosmos as an open platform, and early adopters include [1X Technologies](https://aiwiki.ai/wiki/1x_technologies), [Agility Robotics](https://aiwiki.ai/wiki/agility_robotics), [Figure AI](https://aiwiki.ai/wiki/figure_ai), Foretellix (for autonomous vehicle testing), Skild AI, and Uber. The company also released Cosmos tokenizers (for converting continuous data into discrete tokens suitable for transformer-based models) and guardrails tools [12].

A major release of Cosmos in March 2025 expanded the model suite and physical AI data tools, coinciding with NVIDIA's broader push into what CEO [Jensen Huang](https://aiwiki.ai/wiki/jensen_huang) calls "physical AI," the application of AI to systems that interact with the physical world [12]. NVIDIA followed with [Cosmos 3](https://aiwiki.ai/wiki/nvidia_cosmos_3), launched at GTC Taipei in a press release dated May 31, 2026, which it describes as a single open "omnimodel" built on a mixture-of-transformers architecture that combines vision reasoning, world generation and action prediction, with generation across text, image, video, ambient sound and action [36].

## Odyssey

[Odyssey](https://aiwiki.ai/wiki/odyssey_ai) is an AI lab founded in 2023 by Oliver Cameron and Jeff Hawke, two founders from the self-driving industry, that builds what it calls "foundation world models" [20][21]. Its models are causal and autoregressive: each new frame is generated from prior frames and the user's streaming inputs, so the output can be steered while it plays rather than fixed in advance. Odyssey draws this line explicitly against bidirectional video generators. Its Odyssey-2 Max post argues that models such as Sora, Veo, Kling and Runway "generate past, present, and future jointly from a prompt fixed in advance," which "rules out real-time interaction, since future frames would have to condition on actions the user has not yet taken," whereas "a world model must instead be causal, predicting each state from prior states and actions" [30]. In June 2026 the company raised a $310 million Series B at a $1.45 billion valuation led by Natural Capital, with Amazon, GV, AMD Ventures, EQT and IQT participating; Amazon Web Services became its preferred cloud provider and Odyssey said it was working with Amazon's Annapurna Labs to run its models on [AWS Trainium](https://aiwiki.ai/wiki/aws_trainium) chips [20][21]. NVIDIA's venture arm [NVentures](https://aiwiki.ai/wiki/nvidia_nventures) and Samsung Next had invested in February 2026 [22].

| Release | Date | What Odyssey says it is |
|---------|------|-------------------------|
| Explorer | December 18, 2024 | A purpose-built world model for film and gaming, announced alongside Pixar co-founder Ed Catmull joining the board [23] |
| Odyssey-1 | May 28, 2025 | A research preview of a "real-time, playable world model" with navigation input [24] |
| [Odyssey-2](https://aiwiki.ai/wiki/odyssey_2) | October 27, 2025 | Real-time video steered by typed text, producing a new frame every 50 milliseconds (20 frames per second) [25] |
| Odyssey-2 Max | April 21, 2026 | The third and largest Odyssey-2 model, with about 3x the parameters and 10x the training compute of Odyssey-2 Pro, trained on several hundred NVIDIA B200 GPUs [30] |
| [PROWL-1](https://aiwiki.ai/wiki/prowl_1) | May 12, 2026 | An adversarial reinforcement learning curriculum for improving a world model [26][27] |
| [Starchild-1](https://aiwiki.ai/wiki/starchild_1) | May 17, 2026 | A real-time world model that generates synchronized audio and video [28][29] |
| [Agora-1](https://aiwiki.ai/wiki/agora_1) | May 18, 2026 | A multi-agent world model in which up to four players share one generated world [31] |
| [CaliBench](https://aiwiki.ai/wiki/calibench) | August 10, 2026 | A benchmark for whether video world models reproduce the correct distribution of physical outcomes [32][33] |
| [Odyssey-3](https://aiwiki.ai/wiki/odyssey_3) | September 15, 2026 | A foundation world model used as a frozen backbone for control policies across robots, cars, drones and video games [34] |

Odyssey-2 Max is the model for which the company has published benchmark numbers. Odyssey reports a VBench 2 physics sub-score of 58.52 for Odyssey-2 Max, up from 49.67 for Odyssey-2 Pro, and says this was the highest among the publicly available world models it evaluated, which included NVIDIA's Cosmos-Predict2.5-14B at 44.92; these are Odyssey's own measurements, and the comparison deliberately excludes bidirectional video models on the grounds that they are not interactive [30]. Odyssey describes the architecture as an autoregressive diffusion transformer trained with [flow matching](https://aiwiki.ai/wiki/flow_matching) in a continuous latent space, distilled to few-step denoising so that rollout can run in real time, and trained in three stages: general visual dynamics, interaction and task conditioning, then long-horizon stability [30].

### Odyssey-3: a frozen world model as a backbone for control

Most systems in this article use a world model either to generate video or to plan by rolling futures forward inside the model. Odyssey-3, announced on September 15, 2026, uses one differently: as a pretrained feature extractor whose internal representations are read out by small, separately trained action decoders, with the world model itself left frozen in the experiments where Odyssey specifies this [34]. Odyssey describes Odyssey-3 as an autoregressive [diffusion transformer](https://aiwiki.ai/wiki/diffusion_transformer) trained on "a vast collection of visual observations of the world," and says the same pretrained model was used as the base for policies that control robot arms, humanoid robots, a car, a simulated drone and video game characters, in each case with between a couple of hours and a few tens of hours of paired observation-and-action data from the target system [34]. For driving, the post states that "the pretrained world model remains frozen throughout training, extracting visual representations that feed into a relatively small driving policy, which predicts waypoints ahead of the car"; the drone and gaming policies follow the same recipe [34].

The reported results are the company's own and come without success rates, trial counts or variance. With 20 hours of simulated driving data, Odyssey says the model drove a car in closed loop on roads in India, and that policies trained entirely in simulation "traveled about 77% as far between safety-driver interventions" as policies trained on real footage [34]. With tens of hours of robot demonstrations it reports control of several robot arms, including recovery behaviors (reorienting a gripper after a missed grasp, retrieving an object dropped in an unusual position) that were absent from the demonstrations [34]. Humanoid policies were built by [Flexion](https://aiwiki.ai/wiki/flexion_robotics), which Odyssey announced as a research partner, on top of Odyssey-3 with tens of hours of teleoperation data; Odyssey says these "generalize better to environmental changes than the VLA baselines tested," without naming the [vision-language-action](https://aiwiki.ai/wiki/vision_language_action_model) baselines [34]. A game policy trained on roughly two hours of GTA V footage produced movement in Red Dead Redemption 2 without further training [34]. At announcement Odyssey said it would release the model publicly "in the coming weeks" [34]. The approach is closest in spirit to V-JEPA 2-AC (a video-pretrained model retargeted to a robot with little action data), but Odyssey-3 is a generative model rather than a JEPA-style representation predictor, and the claim is breadth across embodiments rather than a score on any single benchmark.

### Starchild-1: audio and video together

The video prediction and interactive systems described elsewhere in this article are visual only. Starchild-1, previewed on May 17, 2026, is Odyssey's attempt at a multimodal world model: it autoregressively generates synchronized audio and video while responding to streamed text, speech and action inputs, and its technical report says it runs at up to 24 frames per second on modern hardware [28][29]. Odyssey calls it "the world's first multimodal world model"; that is the company's own claim [28]. The technical challenge the report emphasises is that audio and video evolve at different temporal frequencies and information densities, so during a long rollout errors in one modality can destabilize the other. Odyssey's answer was a causal distillation pipeline that converts a bidirectional joint audio-video teacher model into a causal, autoregressive student, plus an asynchronous [KV cache](https://aiwiki.ai/wiki/kv_cache) design that handles the two modalities on different schedules [28][29]. The contrast Odyssey draws is with offline audio-video generators such as Veo, which produce a clip of fixed length whose future is set once generation begins [28].

### Agora-1: shared state, separate rendering

Agora-1, released as a research preview on May 18, 2026, is a multi-agent world model in which up to four players are matched into one generated deathmatch built on the game GoldenEye, and every participant sees a consistent view of the same simulated world [31]. Its architectural point bears on the game-environment use case discussed below. Conventional interactive world models fold simulation dynamics and rendering into a single network; Agora-1 splits them. One model is trained on the internal state of the game (positions, health and so on) and learns how that state changes in response to player actions; a second, DiT-based model renders the shared state from each player's viewpoint, conditioned on the state rather than on prompts or images [31]. Odyssey describes this as "loosely analogous to the structure of a modern game engine," except that both parts are learned, and notes that because the state is explicit it can be edited directly to generate new levels [31]. The post contrasts this with Multiverse and Solaris, earlier multi-agent systems that concatenate players into one split-screen representation or one sequence; Odyssey says the sequence approach scales poorly with player count and that both struggle to stay consistent when players lose sight of each other, while a third system, MultiGen, keeps an explicit shared state as Agora-1 does but models dynamics and rendering differently [31].

### PROWL-1: an adversarial curriculum

PROWL-1 (Prioritized Regret-Driven Optimization for World Model Learning), described in a May 2026 paper by Odyssey researchers with Ilija Bogunovic of University College London, attacks a data problem rather than an architecture problem [26][27]. Passively collected gameplay under-samples the rare, interaction-heavy transitions where world models fail, so PROWL trains a [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) agent to seek out trajectories on which a diffusion-based world model makes large errors, while a KL constraint keeps the agent close to realistic behaviour so it cannot exploit the model with unnatural actions. The world model is then fine-tuned on the discovered trajectories, and a Prioritized Adversarial Trajectory buffer re-ranks them by prediction error, action fidelity and learning progress so that training concentrates on unresolved failures [27]. The experiments used a world model pretrained with [Diffusion Forcing](https://aiwiki.ai/wiki/diffusion_forcing) on human Minecraft data in the MineRL framework; the paper reports improved robustness on held-out out-of-distribution trajectories compared with training on passive data alone, and also that weak behavioural constraints produce reward hacking [27]. Odyssey's framing is that world models, like language models, need scalable feedback loops, and its Odyssey-3 post describes PROWL as the mechanism by which agents trained inside a generated world can in turn improve the world model [26][34].

## Are video generation models world models?

This question has become one of the most contested in AI research. The arguments break down roughly as follows:

**Arguments that video generation models are (or can become) world models:**

1. Scale may be sufficient. As video models get larger and are trained on more data, they may implicitly learn enough physics to be functionally useful as world models.
2. Emergence. Just as LLMs exhibit emergent abilities at sufficient scale, video models may develop emergent physical understanding.
3. Practical utility. Even imperfect physics understanding is useful. A model that "knows" dropped things fall and rolling things continue is more useful for planning than no model at all.

**Arguments that video generation models are not world models:**

1. No causal understanding. Video models learn correlations, not causal mechanisms. They cannot answer "what if" questions involving novel interventions.
2. Physics-IQ results. Empirical testing shows a large gap between visual realism (best score 55.6%) and physical understanding (best score 24.1%), and this gap does not close with scale [8].
3. No action conditioning. Most video generation models cannot be interacted with; they produce a single pre-determined trajectory rather than responding to agent actions.
4. Generalization failure. Models fail on out-of-distribution physical scenarios, suggesting memorization rather than understanding [8].

A paper published in 2024 by researchers from several institutions, titled "Sora and V-JEPA Have Not Learned The Complete Real World Model," argued that neither generative video models nor JEPA-style models had yet achieved genuine world understanding, and that both approaches had fundamental limitations that needed to be addressed [13].

The truth likely lies between the extremes. Video generation models have learned some aspects of world dynamics (enough to generate plausible videos), but fall short of the kind of accurate, generalizable, action-conditioned physical reasoning that would qualify as a true world model. The field is moving toward hybrid approaches that combine the visual generation capabilities of diffusion models with the interactive, action-conditioned structure of RL-based world models.

### Calibration: measuring the distribution of outcomes

Most of the evidence in this debate, including Physics-IQ, scores individual generations: is each video sharp, temporally consistent and physically plausible? Odyssey's CaliBench work, published in August 2026 and accepted at Transactions on Machine Learning Research, argues that this misses a second failure mode [32][33]. Many physical processes are genuinely stochastic, and a simulator that is meant to stand in for the world has to reproduce the right distribution of outcomes, not just one plausible outcome. In the authors' words, a model "can pass every per-sample check, every video looks great, and still be badly miscalibrated, collapsing onto the single most likely outcome every time" [32]. The benchmark therefore uses nine scenes whose outcome distribution is known in closed form (a Galton board's binomial spread over bins, coin-flip-like left-or-right outcomes for a Y-fork, a T-junction and a double pendulum, uniform dice, cards and lottery tumblers, and the skewed red-black-green distribution of a European roulette wheel), generates 32 videos per model for each scene from that scene's fixed starting frame, reads the outcome of each with a vision-language model (Gemini 3.1 Pro, which agreed with a human annotator on 152 of 162 sampled videos, about 93.8 percent), and measures the total variation distance from the true distribution [32][33]. Two quantities are reported separately: scorability, the fraction of generations that yield a readable outcome at all, and [calibration](https://aiwiki.ai/wiki/calibration) among those that do [33].

Across six image-to-video models (WAN-2.7, [Seedance 2.0](https://aiwiki.ai/wiki/seedance_2), HappyHorse-1.0, [Veo 3.1](https://aiwiki.ai/wiki/veo_3_1), Runway Gen-4.5 and Cosmos3-Super), the paper reports that the dominant failure is mode collapse: models pile probability onto a few outcomes, most scene-model pairs are significantly miscalibrated, and in the extreme case Veo 3.1 produced the same die face on every random seed [33]. No model dominated all nine scenes. The authors also found that lowering [classifier-free guidance](https://aiwiki.ai/wiki/classifier_free_guidance) increases outcome diversity but reduces scorability, a trade-off between rendering valid frames and reproducing the right distribution [32]. The argument for why this matters is a safety one: a world model used to stress-test a driving stack or a robot policy in simulation will "hide precisely the rare, dangerous tail events you built the simulation to surface" if it collapses onto the likeliest future [32]. The protocol, conditioning frames and a summary metric (mean normalised total variation, mnTV) are released on GitHub [33].

## What are world models used for?

### Planning in reinforcement learning

The original and most established application of world models is in RL. An agent with an accurate world model can "think ahead" by simulating future trajectories in its model before acting, reducing the amount of real-world interaction needed. This is the approach used by Dreamer, MuZero, and many robotics systems.

### Autonomous driving

World models for autonomous driving predict how traffic scenes will evolve: where other vehicles will go, how pedestrians will move, and what will happen if the ego vehicle takes a particular action. NVIDIA's Cosmos, [Wayve](https://aiwiki.ai/wiki/wayve)'s GAIA-1, and various academic projects pursue this direction. The appeal is that a world model can generate unlimited training scenarios, including rare dangerous situations that are hard to encounter (or safely create) in real-world driving [12]. Wayve has since extended the GAIA line through [GAIA-2](https://aiwiki.ai/wiki/gaia_2_wayve) and [GAIA-3](https://aiwiki.ai/wiki/gaia_3_wayve) (December 2025), a 15-billion-parameter latent diffusion world model that the company frames as a foundation for evaluating its driving software rather than only synthesising footage, and [GAIA-4](https://aiwiki.ai/wiki/gaia_4_wayve) (August 2026), which brings generated radar into the simulation alongside cameras and puts Wayve's driving model inside a closed loop, so that its decisions change the sensor inputs it receives next [37][38]. Odyssey-3's driving experiment shows a third use of a world model in this domain, as a frozen backbone whose representations feed a small waypoint-predicting policy (see the Odyssey section above) [34].

### Robotics

In [robotics](https://aiwiki.ai/wiki/ai_robotics), world models help robots predict the outcomes of manipulation actions ("If I push this object, where will it end up?") and plan multi-step tasks. The combination of world models with language models (as in SayCan-style systems) allows robots to plan at multiple levels of abstraction: the language model decomposes a task into steps, and the world model simulates whether each step is likely to succeed. V-JEPA 2-AC's zero-shot pick-and-place on real Franka arms, using fewer than 62 hours of robot video, illustrates how a video-trained world model can be retargeted to physical control with minimal action data [15]. Odyssey-3 (September 2026) pushes the same idea through a generative model: a frozen world model serves as the representation backbone for policies on robot arms and, through Flexion, humanoids, with tens of hours of demonstrations per embodiment, according to the company [34].

One robot-focused example is [Hydra-0](https://aiwiki.ai/wiki/hydra_0), described in an August 18, 2026 preprint. It conditions a video world model on `action flow`, a set of sparse image-space point trajectories with per-point visibility. The representation can describe visible motion from robot links, grippers, human hands, or manipulated objects without passing embodiment-specific joint commands to the video model. In forward prediction, however, the proposed deployment pipeline still uses a controller and Isaac Lab physics simulation to roll out a candidate command, then projects the robot's surface motion through calibrated camera geometry. Hydra-0 predicts the scene's visual response to that supplied motion. This forward path does not by itself infer or execute commands on arbitrary hardware.[18][19]

Hydra-0's policy study is narrower than closed-loop evaluation. Each RoboLab generation starts from the real first observation and replays the full achieved trajectory; the policy is never queried on generated frames. The authors reported a Pearson correlation of 0.96 between replayed and reference aggregate success rates across 300 episodes, but this result measures open-loop replay. A separate inverse-control proof of concept takes desired object flow from one held-out human flexible-pipe-bending demonstration, then uses a learned readout to produce 14-degree-of-freedom commands for one bimanual YAM robot. The single-task demonstration does not establish general control across embodiments. As of September 3, 2026, the project page listed code as coming soon and provided no public Hydra-0 checkpoints.[18][19]

### Game environments and simulation

Genie and Genie 2 demonstrate the potential for world models to generate interactive environments for gaming, training, and evaluation. Instead of hand-crafting game levels or simulation scenarios, developers could use world models to generate limitless variations, potentially reducing the cost of content creation and testing. Odyssey's Agora-1 extends this from one player to several by keeping an explicit shared game state that a separate learned renderer draws from each participant's viewpoint, and Odyssey positions such worlds as environments for multi-agent reinforcement learning as well as for play [31].

### Prediction and forecasting

World models can be applied to any domain where predicting future states from current conditions is valuable: weather forecasting, economic modeling, protein dynamics, and more. [Google DeepMind](https://aiwiki.ai/wiki/google_deepmind)'s GraphCast weather model and similar systems share the underlying principle of learning dynamics from data to predict future states.

## Competing approaches (2025-2026)

As of early 2026, several major research groups are pursuing distinct approaches to world models:

| Group | Approach | Philosophy |
|-------|----------|------------|
| Yann LeCun / AMI Labs | JEPA (Joint Embedding Predictive Architecture); V-JEPA 2 | Predict in abstract representation space, not pixel space; LLMs are insufficient for physical intelligence |
| Google DeepMind (Genie team) | Interactive world generation from video data | Learn to generate explorable 3D environments; useful for training and evaluating agents |
| OpenAI (Sora team) | Large-scale video generation as implicit world modeling | Sufficient scale in video generation may yield emergent world understanding |
| NVIDIA (Cosmos) | Commercial world foundation models for physical AI | Practical tools for autonomous driving and robotics; data amplification for training |
| Danijar Hafner et al. (Dreamer) | Model-based RL with learned latent dynamics | Compact, efficient models for planning and policy optimization in RL |
| Fei-Fei Li / World Labs | Spatial intelligence and 3D world understanding (Marble) | 3D-native, persistent scene generation as a foundation for world modeling |
| Odyssey (Cameron and Hawke) | Causal autoregressive diffusion transformers; Odyssey-2 Max, Starchild-1, Odyssey-3 | Real-time, action-conditioned generation as the defining property of a world model; a pretrained world model as a frozen backbone for control across embodiments [30][34] |

These approaches are not mutually exclusive, and the eventual winning strategy may combine elements of several. The JEPA approach and the Dreamer approach share the idea of operating in learned representation spaces. The Genie approach and the Sora approach share the idea of learning from large-scale video data. Cosmos bridges research and commercial application [14].

## Current state (2025-2026)

World models are at an inflection point. The concept has moved from a niche topic in model-based RL to a central theme in AI research, driven by several converging trends:

- **Convergence of video generation and world modeling.** The enormous investment in video generation (Sora, [Runway](https://aiwiki.ai/wiki/runway_ml), Pika, and others) has produced models with impressive visual generation capabilities. Whether these models can be upgraded to true world models, or whether fundamentally different architectures are needed, is a key open question.
- **Commercial investment.** NVIDIA's Cosmos, Google DeepMind's Genie, World Labs' $230 million-plus raise, AMI Labs' $1.03 billion seed round (March 2026) and Odyssey's $310 million Series B (June 2026) represent billions of dollars of investment in world model research and infrastructure [16][21][35].
- **Beyond silent video and single players.** Starchild-1 (May 2026) generates synchronized audio and video in real time, and Agora-1 (May 2026) lets several participants share one generated world, extending world models along two axes the earlier systems in this article did not cover [28][31].
- **Evaluation is catching up.** Benchmarks such as CaliBench (August 2026) test whether a world model reproduces the right distribution of physical outcomes rather than only whether each generated video looks plausible; its authors found mode collapse across all six frontier video models they tested [33].
- **Robotics demand.** The boom in [humanoid robots](https://aiwiki.ai/wiki/humanoid_robot) and AI robotics is creating urgent demand for world models that can help robots plan and learn in simulation before deployment. V-JEPA 2's zero-shot transfer to robot arms in 2025 is an early concrete result in this direction [15].
- **Genie 3 as a milestone.** Google DeepMind's Genie 3 (August 2025) achieved real-time interactive world generation at 720p and 24 fps with minute-scale consistency, marking the first system from DeepMind to allow interaction in real time and fast enough for practical agent training and evaluation [11].
- **Persistent limitations.** Despite progress, no system has yet demonstrated a learned world model that can accurately predict physical dynamics across a wide range of scenarios. The gap between visual plausibility and physical accuracy remains large (best Physics-IQ physics score 24.1%), and out-of-distribution generalization continues to be a weakness [8].

The field is likely to evolve rapidly over the next few years. If JEPA-style architectures or Dreamer-style models can be scaled to handle the complexity of the real world, they could enable a new generation of AI agents that genuinely understand their environments. If video generation models prove to be a dead end for world understanding (as some critics predict), the field may pivot toward more structured approaches that incorporate explicit physical reasoning. In either case, world models remain central to the broader goal of building AI systems that can operate effectively in the physical world.

## See also

- [Reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning)
- [Foundation model](https://aiwiki.ai/wiki/foundation_model)
- [Sora](https://aiwiki.ai/wiki/sora)
- [Google DeepMind](https://aiwiki.ai/wiki/google_deepmind)
- [AI robotics](https://aiwiki.ai/wiki/ai_robotics)
- [Variational autoencoder](https://aiwiki.ai/wiki/autoencoder)
- [Odyssey (AI lab)](https://aiwiki.ai/wiki/odyssey_ai)
- [CaliBench](https://aiwiki.ai/wiki/calibench)

## References

[1] Themesis. (2026). "World Models: Five Competing Approaches." https://themesis.com/2026/01/07/world-models-five-competing-approaches/

[2] Sutton, R. S. & Barto, A. G. (2018). *Reinforcement Learning: An Introduction* (2nd ed.). MIT Press.

[3] Ha, D. & Schmidhuber, J. (2018). "World Models." https://worldmodels.github.io/

[4] Hafner, D., et al. (2025). "Mastering diverse control tasks through world models." *Nature*. https://www.nature.com/articles/s41586-025-08744-2

[5] Entropy Town. (2025). "Why Fei-Fei Li, Yann LeCun and DeepMind Are All Betting on World Models." https://entropytown.com/articles/2025-11-13-world-model-lecun-feifei-li/

[6] Meta AI. (2024). "V-JEPA: The next step toward advanced machine intelligence." https://ai.meta.com/blog/v-jepa-yann-lecun-ai-model-video-joint-embedding-predictive-architecture/

[7] OpenAI. (2024). "Sora: Creating video from text." Referenced in: https://openai.com/sora

[8] Motamed, S., et al. (2025). "Do generative video models understand physical principles?" https://arxiv.org/abs/2501.09038

[9] Google DeepMind. (2024). "Genie: Generative Interactive Environments." https://arxiv.org/abs/2402.15391

[10] Google DeepMind. (2024). "Genie 2: A large-scale foundation world model." https://deepmind.google/blog/genie-2-a-large-scale-foundation-world-model/

[11] Google DeepMind. (2025). "Genie 3: A new frontier for world models." https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/

[12] NVIDIA. (2025). "NVIDIA Launches Cosmos World Foundation Model Platform to Accelerate Physical AI Development." https://nvidianews.nvidia.com/news/nvidia-launches-cosmos-world-foundation-model-platform-to-accelerate-physical-ai-development

[13] Liu, et al. (2024). "Sora and V-JEPA Have Not Learned The Complete Real World Model." https://www.arxiv.org/pdf/2407.10311

[14] Introl. (2026). "World Models Race 2026: How LeCun, DeepMind, and Others Compete." https://introl.com/blog/world-models-race-agi-2026

[15] Meta AI. (2025). "Introducing the V-JEPA 2 world model and new benchmarks for physical reasoning." https://ai.meta.com/blog/v-jepa-2-world-model-benchmarks/

[16] TechCrunch. (2024). "Fei-Fei Li's World Labs comes out of stealth with $230M." https://techcrunch.com/2024/09/13/with-230m-in-funding-world-labs-is-building-large-world-models/

[17] TechCrunch. (2025). "Fei-Fei Li's World Labs speeds up the world model race with Marble, its first commercial product." https://techcrunch.com/2025/11/12/fei-fei-lis-world-labs-speeds-up-the-world-model-race-with-marble-its-first-commercial-product/

[18] Li, Hongyu, et al. (2026). "Hydra-0: Action Flow for Generalist World Modeling and Control." arXiv:2608.18077v1. https://arxiv.org/abs/2608.18077

[19] NVIDIA Isaac. (2026). "Hydra-0: Action Flow for Generalist World Modeling and Control." Project page. https://nvidia-isaac.github.io/video_to_data/hydra-0/

[20] Cameron, O. (2026). "Our $310 Million Fundraise to Accelerate World Simulation." Odyssey, June 17, 2026. https://odyssey.systems/our-series-b

[21] Bort, J. (2026). "World model maker Odyssey nabs $1.45B valuation backed by Amazon and other big names." TechCrunch, June 17, 2026. https://techcrunch.com/2026/06/17/world-model-maker-odyssey-nabs-1-45b-valuation-backed-by-amazon-and-other-big-names/

[22] Cameron, O. (2026). "Odyssey Announces Investment from NVentures and Samsung Next." Odyssey, February 12, 2026. https://odyssey.systems/investment-from-nvidia-and-samsung

[23] Cameron, O. (2024). "World Models for Film, Gaming, and Beyond." Odyssey, December 18, 2024. https://odyssey.systems/introducing-explorer

[24] Cameron, O. (2025). "Introducing Odyssey-1: A Playable World Model." Odyssey, May 28, 2025. https://odyssey.systems/introducing-odyssey-1

[25] Cameron, O. (2025). "Introducing Odyssey-2: A General-Purpose World Model." Odyssey, October 27, 2025. https://odyssey.systems/introducing-odyssey-2

[26] Hawke, J. (2026). "Introducing PROWL-1: Learning Through Discovery." Odyssey, May 12, 2026. https://odyssey.systems/introducing-prowl-1

[27] Guzel, A. H., Seidenschwarz, J., Graham, B., Sadeghi, J., Hawke, J. & Bogunovic, I. (2026). "PROWL: Prioritized Regret-Driven Optimization for World Model Learning." arXiv:2605.18803. https://arxiv.org/abs/2605.18803

[28] Cameron, O. (2026). "Starchild-1: The First Real-Time Multimodal World Model." Odyssey, May 17, 2026. https://odyssey.systems/introducing-starchild-1

[29] Team Odyssey. (2026). "Starchild-1: A real-time multimodal world model." Technical report. https://starchild.odyssey.ml/starchild-1.pdf

[30] Cameron, O. (2026). "Introducing Odyssey-2 Max: Scaled World Simulation." Odyssey, April 21, 2026. https://odyssey.systems/introducing-odyssey-2-max

[31] Cameron, O. (2026). "Agora-1: The Multi-Agent World Model." Odyssey, May 18, 2026. https://odyssey.systems/introducing-agora-1

[32] Sadeghi, J. (2026). "Introducing CaliBench: Are World Models Physically Calibrated?" Odyssey, August 10, 2026. https://odyssey.systems/introducing-calibench

[33] Sadeghi, J., Seidenschwarz, J., Allardice, J., Srinivasan, S., Graham, B. & Hawke, J. (2026). "CaliBench: Are the Stochastic Dynamics of Video World Models Physically Calibrated?" arXiv:2608.16829, accepted at Transactions on Machine Learning Research. https://arxiv.org/abs/2608.16829 (code: https://github.com/odysseyml/CaliBench)

[34] Cameron, O. & Hawke, J. (2026). "Introducing Odyssey-3: A General-Purpose Physical Intelligence." Odyssey, September 15, 2026. https://odyssey.systems/introducing-odyssey-3

[35] TechCrunch. (2026). "Yann LeCun's AMI Labs raises $1.03 billion to build world models." March 9, 2026. https://techcrunch.com/2026/03/09/yann-lecuns-ami-labs-raises-1-03-billion-to-build-world-models/

[36] NVIDIA. (2026). "NVIDIA Launches Cosmos 3, the Open Frontier Foundation Model for Physical AI." NVIDIA Newsroom, May 31, 2026. https://nvidianews.nvidia.com/news/nvidia-launches-cosmos-3-the-open-frontier-foundation-model-for-physical-ai

[37] Wayve. (2025). "GAIA-3: Scaling World Models to Power Safety and Evaluation." December 2, 2025. https://wayve.ai/thinking/gaia-3/

[38] Wayve. (2026). "GAIA-4: Multimodal World Models Powering Closed-Loop Simulation for Safe and Scalable Autonomy." August 3, 2026. https://wayve.ai/thinking/gaia-4/