英国王立協会特集号に見る、世界モデルの最前線とAIの未来

A world model is an internal representation that helps an organism or AI predict what happens next and choose an action. Sakana AI's overview of a Royal Society journal issue on world models emphasizes that this includes representing the acting system itself. For a robot, knowing its own state is described as the basis for adjusting movements to the situation.

Sakana uses this perspective to distinguish fluent language behavior from understanding how actions affect the world: large models can do much by learning patterns in text, which does not guarantee they grasp cause and effect. Sakana says it will pursue the theme through world-model and Physical AI research at its RSI Lab. The overview is a summary of the issue's themes, not a report of Sakana's own experimental results.