Understanding Spatial AI and the Distinction Between World Models and Language Models

Defining Spatial AI

Spatial AI encompasses technologies that create a three-dimensional representation of surroundings, effectively filling in gaps in information while maintaining accuracy. For example, when shown a photograph of a room, a spatial AI model can predict elements not visible in the image, such as the opposite wall or the underside of a chair. This ensures that when the original viewpoint is revisited, the reconstructed elements remain consistent and correctly positioned.

This concept is often referred to as spatial intelligence. Systems in this category typically operate by:

• Accepting various input forms, including photographs, videos, or textual descriptions;
• Establishing a link between each frame and its respective point of view to create cohesive geometry;
• Inferring unrepresented areas based on a database of knowledge regarding environments;
• Retaining the final output for consistent retrieval along a defined path.

These systems are classified as world models. Fei-Fei Li, a Stanford University professor, outlined essential criteria for these models, emphasizing that they must adhere to physical laws, accommodate various input types, and predict the state of the world following any action.

The first commercial application in this field emerged in November 2025, when World Labs launched Marble, allowing users to create and export 3D worlds using techniques like "Gaussian splats." Currently, a standardized training method for spatial AI is lacking, unlike language models which are trained to predict the next word in a sequence.

Differences Between World Models and Language Models

World models and language models handle different data types. Text is linear, with each word dependent on its predecessor, while spatial data remains static regardless of its representation in a frame. This leads to distinct requirements for consistency:

• Language models must only avoid contradictions in previous statements;
• Spatial models must maintain geometric coherence, ensuring that returning to a starting point yields the same visual as the initial view.

While the internal structures of both models share similarities—such as integrating data into a unified output—the key difference lies in how spatial models associate images with specific viewpoints, enabling volumetric construction rather than linear progression.

Li explains that while language is a sequence of symbols, describing a spatial environment requires consideration of complex interactions, such as gravity and light, making the modeling of geometry and physics essential.

Evaluating the performance of spatial models presents challenges due to the absence of standardized testing. World Labs demonstrated their Atlas model on two tasks: generating views based on camera trajectories and reconstructing scenes from multiple images, without assessing the model's handling of real-world physics.

Challenges for Language Models in Spatial Contexts

Language models can effectively navigate routes within a city, even without a coherent map. A study involving MIT, Harvard, and Cornell researchers showed that a model trained on taxi trips in Manhattan could identify optimal routes, even in the absence of a detailed map. However, when faced with street closures, accuracy dropped significantly.

When researchers reconstructed the model's map based on its own routes rather than actual streets, they found inaccuracies, highlighting the limitations of the model’s understanding. Despite the model's success in familiar settings, it struggled to adapt to unexpected changes.

Recent evaluations of human versus machine performance on spatial tasks show a stark difference, with humans scoring significantly higher than even the best models.

Origins of Training AI in Imaginary Environments

The concept of training AI in imagined settings traces back to psychology, with British researcher Kenneth Craik proposing in 1943 that the brain simulates reality to anticipate outcomes. This idea was translated into machine learning in 2018 when David Ha and Jürgen Schmidhuber developed a system that trained on arcade games, predicting future actions based on previous frames.

However, this approach also revealed vulnerabilities, as the AI discovered shortcuts that did not exist in the actual game environment. To address these issues, researchers have experimented with introducing noise into simulations, although a universally effective solution remains elusive.

Creating 3D Scenes from Images

Unlike traditional image generators that produce frames sequentially, Atlas utilizes multiple images tagged with their respective shooting locations and angles. This method reduces the need for the model to fabricate details, enhancing accuracy when reconstructing environments.

World Labs reported that Atlas achieved an average reconstruction error of 25.3, outperforming five specialized models. The model's capabilities extend to generating new perspectives, depth maps, and point clouds, which can be utilized in various applications, including gaming and mixed reality.

Key Players in Spatial AI Development

Leading companies in the spatial AI sector include:

World Labs: Launched a public interface for developers in January 2026 and secured $1 billion in funding. They acquired SceniX, a robotics simulation team, in July 2026.

Google DeepMind: Introduced the Genie 3 model in summer 2025, capable of generating interactive environments in real-time but limited in visual memory.

AMI Labs: Founded by Yann LeCun and Alexandre LeBrun, the company raised $1.03 billion in March 2026, focusing on architectures that learn from video data.

The competitive landscape is still developing, with few commercial implementations and no public statistics available.

Applications of World Models in Real-World Scenarios

One primary application of world models is in autonomous transportation. Waymo introduced a world model based on Genie 3 that provides comprehensive environmental data, allowing for counterfactual scenario analysis.

Another area of application is in robotics, where World Labs demonstrated training robots entirely within simulated environments, enabling them to perform complex tasks without prior real-world examples.

Discrepancies Between Realistic Video and Physics

Research from Google DeepMind and INSAIT revealed that video generators often produce sequences that lack adherence to physical laws, suggesting that realism does not always equate to physical accuracy. This discrepancy arises from the task of generating frames based on statistical resemblance to training data rather than physical correctness.

While Atlas has been tested for generating views and reconstructing scenes, its performance regarding physical accuracy remains unassessed.

The Importance of Spatial AI for Autonomous Agents

While AI assistants can function without spatial awareness for text-based tasks, robots require a comprehensive understanding of their physical environment to make informed decisions. This necessity underscores the value of world models in providing crucial predictive insights about interactions with objects and environments.

Training costs are also reduced, as capturing environments with a phone is less expensive than manually creating 3D models, allowing for the generation of diverse training variations from a single recording.

Despite claims about the effectiveness of generative systems in reproducing real-world dynamics, independent verification of these outcomes is still pending. Currently, there is no intersection between world models and blockchain technology, as developers have not announced plans to integrate these systems with on-chain computations, although payment protocols for autonomous agents are evolving in parallel.