Computer scientist Louis Castricato quit his studies at Brown University to start a new company, Overworld, which aims to develop AI that can understand and navigate the physical world. This new field of research is known as ‘world models,’ which teach AI systems how to react in physical environments.

What are World Models?

World models are AI systems that learn the statistical structure of space and time, allowing them to predict the consequences of their actions. This is different from language models, which learn the statistical structure of text. World models have the potential to enable AI systems to work like robots, adapting to their environment and interacting with objects in a physical space.

Several prominent scientists, including ‘Godmother of AI’ Fei-Fei Li, are working on world models. Li describes the concept of a world model as ‘one of the most important and most overloaded terms in AI today.’ She has divided world models into three categories: renderers, simulators, and planners. Renderers prioritize visual fidelity, simulators create virtual training grounds, and planners predict what an AI agent or robot should do in an unstructured world.

Applications of World Models

World models have various applications, including robotics, video games, and weather prediction. Castricato’s company, Overworld, is building video game worlds where virtual characters can interact with objects in a physical environment. Other companies, such as Causal Labs and Extropic, are also working on world models for weather prediction and specialized computer chips.

Venture capitalists, such as Steve Jang, are investing in world model-focused companies. Jang believes that the future of AI will involve multiple types of models with different philosophies and architectures.

Original reporting: KTBS 3 (Shreveport) — read the source article.