“Amjad Masad asserts that the current era of LLM development involves significant manual work, such as data labeling and creating contrived reinforcement learning environments, a departure from the earlier scaling-focused era of models like GPT-2 and GPT-3.”