By integrating a 3D representation into models, we can better align the model's capabilities with the tasks we want it to perform, enhancing user affordances. Spatial intelligence is crucial for creating technology that understands and interacts with the 3D world, enabling a wide range of applications.

Spatial intelligence is about a machine's ability to perceive, reason, and act in 3D space and time, understanding how objects and events are positioned and how interactions affect those positions.

Visual spatial intelligence is as fundamental as language, and we have the right ingredients now, including advancements in compute and algorithms, to unlock its potential.

The future of spatial intelligence includes generating immersive 3D worlds for various applications, such as gaming and education, rather than just static images or videos.

Building a successful team for spatial intelligence requires multidisciplinary expertise, including system engineering, machine learning, and computer graphics.

The grounded thesis of language posits that human language is reflected in the operations of large language models, suggesting that the mechanisms for generating language in humans and machines are fundamentally similar.

Language models are currently the primary focus of AI development, which was not anticipated 15 years ago when we were exploring physical grounding in AI.

Language is an autoregressive model, and this autoregressive nature is fundamental to how we generate and understand language.

The distinction between linguistic and perceptual systems suggests that while they can communicate, they operate in fundamentally different spaces, each with its own rules and properties.