Christian Moss

Reading anything about infrastructure building in this country feels like being in a waking nightmare “Analysis earlier this year concluded that decommissioning the scheme could cost as much as continuing to build the railway”

LLMs can build games they can't play. They can't predict the consequences of their actions. They are inherently incapable of planning and can't predict the next state of the world. We all know this. We've all seen LLMs make a mistake, correct it, just to make the same mistake again. Or end up in a lobotomized infinite loop of "oh no you're absolutely right" responses. LLMs are not the efficient frontier. World models are forced to learn the "physics" of the world during training. They do this purely by observing interactions with the world. World models observe, plan, predict, and take an action. And they are efficient. Unlike LLMs they present an opportunity for building agents with the ability to learn throughout their life. I've trained a world model of a Snake game on a MacBook (Yann Lecun JEPA style) in just half an hour. After watching me play for a bit and generating many random interactions, the world model is better than most human players.