“In 2017, Alec Radford's GPT-1 was trained on a narrow distribution of text, possibly fan fiction, totaling around a billion words, which limited its generalization capabilities.”