The idea is to generate enormous amounts of reliable training data through the use of multiverse-style world simulators that can take a single real-world situation – or even just a text prompt, then ...