Report
Three ways AI world models represent the world
The Turing Post’s guide groups world models by generated observations, compact latent states and explicit object-and-space structures.
TLDR
The Turing Post lays out three approaches to AI world models: predicting what comes next as an image, video frame or token sequence; compressing the world into an internal state and predicting how it changes; or explicitly representing objects, positions and interactions. Its guide cites Dreamer 4 and Nvidia’s Cosmos-Predict2.5 as examples of the first approach, JEPA and MuZero of the second, and Atlas of the third.
Combined views
2K
1 Source, first seen ago
