In an experiment, DeepMind researchers investigated the extent to which AI agents could generalize learned knowledge to unseen tasks.