What world-model research proposes for adaptive AI: action consequences, learning from failure and a practical way to test predictions in a simulator.
