“The early version of GPT-3, DaVinci 002, had a context window of approximately 4,000 tokens.”