Commentator Lisan al Gaib argues reinforcement learning requires massive batch sizes to stabilize training against high variance and weak signals

The analyzed training setup used a 7,040 global batch size.