“Early GPT-3 models like DaVinci 002 had a context window of approximately 4,000 tokens.”