But Google's DeepMind AI group has now developed a reinforcement learning tool that can develop extremely optimized algorithms without first being trained on human code examples.