According to @karpathy, the pretraining era prioritized large, diverse, high‑quality internet text, while the supervised finetuning era prioritizes high‑quality conversational datasets, often produced by contract workers generating Q&A answers. Source: Andrej Karpathy on X, Aug 27, 2025. This shift indicates the bottleneck and value capture are moving toward ownership and production of curated conversational data and scalable labeling capacity, which directly affects where competitive advantage concentrates in AI models. Source: Andrej Karpathy on X, Aug 27, 2025. For crypto markets, the data‑scarcity theme aligns with on‑chain narratives around decentralized data curation and monetization, making data‑focused AI‑crypto segments a relevant area to monitor for liquidity and catalyst flow. Source: Andrej Karpathy on X, Aug 27, 2025.