According to Andrej Karpathy, the priorities in AI model training have shifted significantly over time. During the pretraining era, success depended on large, diverse, and high-quality internet text datasets, which enabled models to learn general language patterns and facts (source: Andrej Karpathy, Twitter). In the supervised finetuning era, the focus switched to conversational data, often generated by contract workers who create question-answer pairs to improve model performance in structured, real-world interactions (source: Andrej Karpathy, Twitter). This shift highlights new AI business opportunities in the creation and curation of high-quality human-labeled conversational datasets, which are now critical for advancing large language models and maintaining competitive differentiation in the generative AI market.