As neural nets become more advanced, they will increasingly become more data efficient at learning new tasks, requiring fewer examples to train additional behaviors. Pre-training a massive neural net like GPT will become increasingly useful at training any arbitrary new task.