“The limitation of finite training data can likely be overcome by using synthetic data, where models generate new data for themselves, similar to how DeepMind's AlphaGo Zero trained by playing against itself.”