“In early pre-training efforts at OpenAI, the dataset batches were not independent and identically distributed (IID), violating a core assumption of deep learning.”