Practice world

A language model predicts the next word. A world model predicts the next image: what happens in a space when you do something. According to WIRED, simulations like these are becoming the training ground for robots and self-driving cars. Niantic Spatial pictures it like this: photograph a room, the software makes a 3D copy of it, and a robot practises there millions of times. Wayve uses the same approach to recreate dangerous roads that cannot be tested in real life.

A lot of money is going into it. Yann LeCun's AMI Labs raised 1.03 billion dollars in March, and AMD is buying Fei-Fei Li's World Labs for 8.2 billion.

The human side comes at the start. To recreate the world, a model first has to see it. A WIRED reporter filmed their own laundry and dishes with an iPhone strapped to their head, and sold the footage through an app for "a fairly measly payout".

Nobody knows yet whether what a robot learns in a simulation will also work in a messy real home. And who owns the footage of your home?