“The focus on learning from scratch in models like AlphaZero caused the GPT paradigm of large-scale pre-training on human data to be overlooked by many in the research community, with the exception of OpenAI.”