Companies like OpenAI and Google are running out of the data used to train artificial intelligence systems. Can new methods continue years of rapid progress?